Jason Brownlee Thesis PDF

Clonal Selection as an Inspiration for
Adaptive and Distributed Information

Processing
Jason Brownlee
2008
A dissertation presented for fulfillment of the requirements for the award of

Doctor of Philosophy
Abstract
The development of adaptive and intelligent computational methods is an important fron-

tier in the field Artificial Intelligence given the fragility of top-down software solutions to
complex problems involving incomplete information. This dissertation describes a system-
atic investigation of the Clonal Selection Theory of acquired immunity as a motivating
information processing metaphor of a series of adaptive and distributed Computational
Intelligence algorithms. The broader structure and function of the mammalian immune
system is used to frame the cellular theory and classically inspired approaches, providing
the additional distributed perspectives of a ‘host of tissues’ called the Tissue Paradigm
and a ‘population of hosts’ called the Hosts Paradigm. This investigation was motivated
by three open problems in the broader field of Artificial Immune Systems, specifically the
perceived impasse in the development, identity, and application of the field, the promise of
distributed information processing, and the need for a framework to motivate such work.
The state of Clonal Selection Algorithms is investigated in the context of immunolog-
ical theory, and considered in the context of broader related machine learning fields and
adaptive systems theory. A systematic approach is adopted in considering the adaptive
qualities of clonal selection beyond a cellular perspective, involving the identification of
the lymphatic system and lymphocyte migration as a motivating metaphor for intra-host
distributed systems, and host immunisation and evolutionary immunology as a motivat-
ing metaphor for intra-population distributed system design. Relevant immunophysiology
and theory was reviewed, abstracted to computational models and algorithms, and sys-
tematically assessed on model pattern recognition problems to demonstrate and verify
expected information processing capabilities. The empirical investigation reveals a variety
of tissue and host based clonal selection systems capable of acquiring distributed informa-
tion via internal processes of controlled localisation and dissemination, in a decentralised
information exposure environment.
The general capabilities of Clonal Selection Algorithms as a Computational Intelli-
gence paradigm are defined in the context of a detailed assessment of the suitability of
the approaches to the important problem primitives of Function Optimisation and Func-
tion Approximation. The findings highlight the general capabilities of the approaches as
mutation-based parallel hill climbers for global optimisation and prototype quantisation
approaches to function approximation. Finally, an Integrated Hierarchical Clonal Selec-
tion Framework demonstrates the subsumed relationship between the cellular, tissue, and
host classes of algorithm, the dependent relationship with the complexity of the abstract
antigenic environment addressed by the system, and a general scaffold for a broader class
of distributed artificial immune systems.
i
Acknowledgements
I could not have completed this work without the help and assistance of a large number
of people and organisations.
Professionally, I would like to thank:

Tim Hendtlass for your wisdom, enthusiasm, encouragement and seemingly unbounded
patience and support through times easy and difficult. Your hands off approach has made
me both more mature as a person, and more confident as a researcher. I am fortunate to
have counted you as mentor and friend these past years. The Faculty of Information and
Communication Technologies, Swinburne University of Technology for providing me with
candidature and support throughout my PhD. The Australian Government, for providing
me with the financial assistance of an Australian Postgraduate Award (APA) during my
candidature.
Personally, I would like to thank:

Daniel Angus for his friendship, discussions, rants, debates, collaboration, competition,
and ultimately for defining my PhD experience. My officemates, past and present: Irene
Moser, Nigel Fisher, Gerard Murray, Clinton Woodward, Hans Pierrot, James Mont-
gomery, and Ben Prime for providing much needed fun, drinking partners, and most
importantly anchors for my sanity. Friends and Housemates for the good times, bad
times, and the much needed balance you brought to my life. Ying Liu for your inti-
mate understanding and patience for me whilst writing up my thesis. Finally, my parents
Lynette Brownlee and Derek Brownlee, and my sister Laura Brownlee for your uncon-
ditional continued support of my decision to quit a perfectly good job and pursue my
research interests.
ii
Declaration
I declare that: This thesis contains no material which has been accepted for an award to
myself for any other degree or diploma, except where due reference is made in the text
of the thesis. This thesis, to the best of my knowledge contains no material previously
published or written by another person except where due reference is made in the text of
the thesis. Where the work is based on joint research or publications, this thesis discloses
the relative contributions of the respective works or authors.
Jason Brownlee
Date:
Contents
Abstract i
Acknowledgements ii
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.5 Thesis Organisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 Background 8
2.1 Chapter Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 Artificial and Computational Intelligence . . . . . . . . . . . . . . . . . . . . 8
2.3 Natural and Artificial Immune Systems . . . . . . . . . . . . . . . . . . . . 11
2.4 Three Schools of AIS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.5 Distributed Artificial Immune Systems . . . . . . . . . . . . . . . . . . . . . 19
2.6 Open Problems in AIS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.7 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.8 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3 Clonal Selection Paradigm 35

3.2 Clonal Selection Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.3 Clonal Selection Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.4 Related Computational Paradigms . . . . . . . . . . . . . . . . . . . . . . . 55
3.5 Clonal Selection Adaptive Strategy . . . . . . . . . . . . . . . . . . . . . . . 64
3.6 Cellular Strategy and Beyond . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.7 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4 Cellular Clonal Selection 76

4.2 Abstract Cellular Paradigm . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.3 Realised Cellular Paradigm . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.4 Cellular Clonal Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.5 Spatial Repertoire . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
4.6 Response Mediation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
4.7 Cell-Cell Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
4.8 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
iv
5 Tissue Clonal Selection 145
5.2 Physiology of Lymphocyte Migration . . . . . . . . . . . . . . . . . . . . . . 145
5.3 Abstract Tissue Paradigm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
5.4 Realised Tissue Paradigm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
5.5 Lymphocyte Recirculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
5.6 Lymphocyte Homing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
5.7 Tissue Inflammation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
5.8 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
6 Host Clonal Selection 209

6.2 Populations of Interacting Immune Systems . . . . . . . . . . . . . . . . . . 209
6.3 Abstract Host Paradigm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
6.4 Realised Host Paradigm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
6.5 Population Elicited Immunity . . . . . . . . . . . . . . . . . . . . . . . . . . 228
6.6 Population Shared Immunity . . . . . . . . . . . . . . . . . . . . . . . . . . 239
6.7 Generational Maternal Immunity . . . . . . . . . . . . . . . . . . . . . . . . 246
6.8 Generational Evolved Immunity . . . . . . . . . . . . . . . . . . . . . . . . . 253
6.9 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
7 Hierarchical Frameworks 261

7.2 Clonal Selection Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
7.3 Antigenic Exposure Framework . . . . . . . . . . . . . . . . . . . . . . . . . 268
7.4 Integrated Hierarchical Framework . . . . . . . . . . . . . . . . . . . . . . . 273
7.5 Methodology and Hierarchical Framework . . . . . . . . . . . . . . . . . . . 278
7.6 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
8 Suitability and Application 290

8.2 Suitability of Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
8.3 Function Optimisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
8.4 Function Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
8.5 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318
9 Conclusions 320
9.1 Research Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
9.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
9.3 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
9.4 Further Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326
9.5 Final Words . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331
A Tables of Empirical Results 332

A.1 Tissue Clonal Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 332
A.2 Host Clonal Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336
B Works Arising from the Study 342

B.1 Peer-Reviewed Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . 342
B.2 Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342
B.3 Unpublished Technical Reports . . . . . . . . . . . . . . . . . . . . . . . . . 343
Bibliography 349
v
List of Figures
2.1 Depiction of the conceptual framework for devising Biological Inspired Al-
gorithms, taken from [373] . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.1 Depiction of the shape-space formalism taken from [322] (page 1225). Xi
represents antibodies with corresponding complementary recognition vol-
umes V!i . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2 Theory of receptor editing providing long jumps in the affinity landscape,
from [167]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.1 Depiction of external stimulation resulting in internal activation. . . . . . . 80

4.2 Example exposure regimes demonstrating some archetypical pathogen vir-
ulence behaviours. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.3 Flow diagram of the single and multiple exposure models. . . . . . . . . . . 83
4.4 Depiction of the trade-off between control of the environment and control
of the system. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.5 Box-and-whisker plot’s from the Cellular Empirical Study. . . . . . . . . . . 101
4.6 Example plots of Trs from the CCSA(N+N) at the end of the run on left
section on both plots, for both ACSP-1 and ACSP-10, where the problem
solution is represented on the right of each plot, and colour components
within each CSP. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.7 Box-and-whisker plot’s from the Replacement Empirical Study. . . . . . . . 105
4.8 Plots of the ordered state of the repertoire after the triggering of the stop
condition (left of each plot) for the assessed CCSA and RCCSA algorithms
on the ACSP-10, where the right of each plot provides the problem optima. 106
4.9 Depiction of a polyclonal activation and a polyclonal response. . . . . . . . 108
4.10 Depiction of a one-dimensional toroidal lattice. . . . . . . . . . . . . . . . . 113
4.11 Depiction of clones competing in the neighbourhood of their progenitor. . . 114
4.12 Depiction of the explicit versus implicit partitioning of input signals and
the effects on the repertoire. . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
4.13 Box-and-whisker plot’s from the Spatial Repertoire Empirical Study. . . . . 117
4.14 Example plots of the spatial repertoire from two variations of SCCSA at
the end of the run on ACSP-10 with BMC represented with circles. . . . . . 117
4.15 Depiction of inter-repertoire structures based on the mapping of cells. . . . 120
4.16 Depiction of the integration of feedback for supervised structure formation. 122
4.17 Depiction of the principle concerns of the MCCSA and the relationships in
terms of the selected measures. . . . . . . . . . . . . . . . . . . . . . . . . . 124
4.18 Example plots of the B- (top of each plot) and T-Cell (bottom of each plot)
repertoires from the ECCSA at the end of the run on ACSP-10 for two
different inter-repertoire mapping types. . . . . . . . . . . . . . . . . . . . . 125
4.19 Box-and-whisker plot’s for the relationship configuration with ECCSA. . . . 126
vi
4.20 Example plots of the B- (top of each plot) and T-Cell (bottom of each plot)
repertoires from the MCCSA at the end of the run on ACSP-10 for all four
inter-relationship types. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.21 Examples of more complex relationships that may be formed given receptor-
receptor recognition. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
4.22 Depiction of the two-string representation scheme. . . . . . . . . . . . . . . 131
4.23 Example of a simple two-receptor promotion scheme. . . . . . . . . . . . . . 132
4.24 Depiction of the example difference between a natural and remapping scheme.132
4.25 Depiction of the relationship between remapped antigen and receptors (ps
is primary and ss is secondary string). . . . . . . . . . . . . . . . . . . . . . 133
4.26 Box-and-whisker plot’s from the Recurrent NCCSA Empirical Study. . . . . 136
4.27 Example plots from of the response mappings and final repertoire from
R-NCCSA-PS on ACSP-10. . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
4.28 Box-and-whisker plot of the Average Cell Error for the DE-NCCSA schemes
on ACSP-10. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
4.29 Example repertoire plots of the four Dual Exposure NCCSA configurations
on ACSP-2 showing primary and secondary representations and antigen left
to right. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
5.1 Multiple tissue model and the situated spatial properties that may affect
antigenic exposure. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
5.2 Example of two different Tissue Exposure Regime with controlled selection
of discrete repertoires on exposure events. . . . . . . . . . . . . . . . . . . . 156
5.3 Depiction of the temporal memory effect in a single repertoire model over
time, where the arrows at the top represent information flow into the system
and the arrow along the bottom represents time. . . . . . . . . . . . . . . . 157
5.4 Depiction of the repertoire specialisation that results in the Spatial Self-
Organisation effect, where the arrows from above represent information
flow into the system and the arrows below represent dissipation of acquired
information. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
5.5 Depiction of the unbiased dissemination of information that results in the
Spatial Consistency of Response, where the arrow above represents infor-
mation flow into the system and the arrow below represents information
availability throughout the system. . . . . . . . . . . . . . . . . . . . . . . . 159
5.6 Depiction of information sharing and a cross-reactive response, where the
arrows above represent the flow of information into the system as well as
an implicit anticipation for the application of acquired information. . . . . . 160
5.7 Depiction of the trade-off between the general anticipation of exposure and
organised anticipation of exposure. . . . . . . . . . . . . . . . . . . . . . . . 161
5.8 Depiction of the relationship between primary and secondary lymphoid tis-
sues, where the arrows show the sharing of information between primary
tissues and the outward push of information from primary to secondary
tissues.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
5.9 Depiction of the relationship between secondary tissues and secondary and
tertiary tissues, where the arrows show the sharing of information between
secondary tissues and the outward push of information from secondary to
tertiary tissues. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
5.10 Depiction of the relationship between tertiary and secondary tissues and
tertiary tissues and the environment, where the arrows depict the flow of
information from the environment to the tertiary tissues and ultimately to
the secondary tissues. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
vii
5.11 Depiction of the layered tissue architecture and information flows between
the layers, the arrows show the flow of information from the environment
into the system as well as the outward push of new detectors from the centre
of the system. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
5.12 Depiction of the information streaming providing by lymph collecting in
lymph nodes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
5.13 Depiction of the courier-like behaviour of dendritic cell’s. . . . . . . . . . . . 165
5.14 Depiction of the layered information processing properties of the lymphoid
tissue architecture. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
5.15 Depiction of the simple filter model. . . . . . . . . . . . . . . . . . . . . . . 167
5.16 Depiction of the lymph node model. . . . . . . . . . . . . . . . . . . . . . . 167
5.17 Depiction of the host model. . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
5.18 Depictions of general and minimal configuration examples of the Infection
Antigenic Exposure Problem (IAEP). . . . . . . . . . . . . . . . . . . . . . 171
5.19 Depictions of the four different Tissue Exposure Regimes (TER). . . . . . . 172
5.20 Box-and-whisker plot of Host Diversity (HD) across all TER for the RTCSA
study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
5.21 Box-and-whisker plot of Host Error (HE) across all TER for the RTCSA
study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
5.22 Box-and-whisker plot of Average Tissue Diversity (ATD) across all TER
for the RTCSA study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
5.23 Box-and-whisker plot of Average Tissue Error (ATE) across all TER for
the RTCSA study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
5.24 Box-and-whisker plot of Host Diversity (HD) across all TER for the HTCSA
study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
5.25 Box-and-whisker plot of Host Error (HE) across all TER for the HTCSA
study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
for the HTCSA study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
the HTCSA study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
5.28 Box-and-whisker plot of Host Diversity (HD) across all TER for the ITCSA
study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
5.29 Box-and-whisker plot of Host Error (HE) across all TER for the ITCSA
study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
for the ITCSA study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
the ITCSA study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
6.1 Host Clonal Selection Model and the situated spatial properties that af-
fect antigenic exposure, where the arrows represent interaction between the
discrete components (hosts) in the system. . . . . . . . . . . . . . . . . . . . 216
6.2 Elaboration on figure 5.18b to depict the Minimal Habitat Antigenic Expo-
sure Problem (MAHEP) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
6.3 Box-and-whisker plot of Population Diversity (PD) across all HER for the
THCSA study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
6.4 Box-and-whisker plot of Population Error (PE) across all HER for the
THCSA study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
6.5 Box-and-whisker plot of Average Host Diversity (AHD) across all HER for
the THCSA study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
viii
6.6 Box-and-whisker plot of Average Host Error (AHE) across all HER for the
THCSA study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
SI-HCSA study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
6.8 Box-and-whisker plot of Population Error (PE) across all HER for the SI-
HCSA study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
the SI-HCSA study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
SI-HCSA study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
MI-HCSA study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
6.12 Box-and-whisker plot of Population Error (PE) across all HER for the MI-
HCSA study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
the MI-HCSA study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
MI-HCSA study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
EI-HCSA study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
6.16 Box-and-whisker plot of Population Error (PE) across all HER for the EI-
HCSA study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
the EI-HCSA study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
EI-HCSA study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
7.1 Depiction of the focus on systems and the compression of the environment
in the Hierarchical Clonal Selection Framework (HCSF). . . . . . . . . . . . 262
7.2 Depiction of the compression of both the system and the environment when
the HCSF is compressed to the primary objects of the adaptive system
formalism. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
7.3 Depiction of an expanded realisation of a hierarchical clonal selection system
with all three tiers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
7.4 Depiction of the focus on environment and the compression of the systems
in the Hierarchical Antigenic Exposure Framework (HAEF). . . . . . . . . . 269
7.5 Depiction of the focus on system and environment in the Integrated Hier-
archical Clonal Selection Framework (IHCSF). . . . . . . . . . . . . . . . . 273
7.6 Depiction of the integrated hierarchical clonal selection and hierarchical
antigenic exposure frameworks, showing subsumed relationships and com-
ponent cardinalities across the investigated paradigms. . . . . . . . . . . . . 276
8.1 Depiction that the suitability of application of a strategy to a problem is

defined by the overlap in their features, as well as the comparability with
other strategies, and transferability of results to other problems. . . . . . . 292
8.2 Depiction of the cellular algorithm approach to inductive learning with
holistic, sub-, and partial-models of a given problem. . . . . . . . . . . . . . 295
8.3 Depiction of the tissue algorithm approach to inductive learning with holis-
tic, sub-, and partial-models of a given problem. . . . . . . . . . . . . . . . 298
8.4 Depiction of the host algorithm approach to inductive learning with holistic,
sub-, and partial-models of a given problem. . . . . . . . . . . . . . . . . . . 301
ix
List of Tables
3.1 Clonal selection algorithm taxonomy overview. . . . . . . . . . . . . . . . . 46

3.2 CLONALG parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.3 BCA parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.4 SIA parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.5 Summary of primary objects in the adaptive systems formalism. . . . . . . 65
3.6 Summary of secondary objects in the adaptive systems formalism. . . . . . 66
3.7 Fundamental questions regarding adaptive systems, taken from [216] (page
29). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.1 List of goals for a colour-space domain. . . . . . . . . . . . . . . . . . . . . 87

4.2 Summary of terms for the colour space domain formalism. . . . . . . . . . . 88
4.3 Summary of the assessed configuration of CCSA for ACSP-1. . . . . . . . . 99
4.4 Summary of the assessed configuration of CCSA for ACSP-10. . . . . . . . 99
4.5 Summary of results for CCSA on ACSP-1 and ACSP-10. . . . . . . . . . . . 100
4.6 Summary of the assessed configuration for the CCSA-RCCSA empirical study.103
4.7 Summary of results from the CCSA-RCCSA empirical study on ACSP-10. . 104
4.8 Summary of results from the DCCSA empirical study on DCSP-10. . . . . . 110
4.9 Summary of results for SCCSA on AEP 10. . . . . . . . . . . . . . . . . . . 116
4.10 Summary of the configuration of the ECCSA with a series of inter-repertoire
selection relationships. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
4.11 Summary of results for ECCSA with two different mapping schemes on
ACSP-10. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4.12 Summary of results for ECCSA with four different inter-repertoire relation-
ship schemes on ACSP-10. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4.13 Exposure-activation scheme for a minimal implementation. . . . . . . . . . 132
4.14 Summary of the various replacement configurations for the mapped Recur-
rent Exposure Network Cellular Clonal Selection Algorithm (RE-NCCSA). 134
4.15 Summary of results for Recurrent Network Cellular Clonal Selection Algo-
rithm (R-NCCAS) on ACSP-10 . . . . . . . . . . . . . . . . . . . . . . . . . 135
4.16 Summary of the various replacement configurations for the Dual Exposure
Network Cellular Clonal Selection Algorithm (DE-NCCSA). . . . . . . . . . 139
4.17 Summary of results for the Dual Exposure Network Cellular Clonal Selec-
tion Algorithm (DE-NCCSA) on ACSP-10. . . . . . . . . . . . . . . . . . . 140
5.1 Summary of information processing needs, and proposed algorithms that

meet those needs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
5.2 Summary of the principle responsibility of each tier of the architecture. . . 166
5.3 Configuration of the architecture for the filter model. . . . . . . . . . . . . . 167
5.4 Configuration of the architecture for the lymph node model. . . . . . . . . . 168
5.5 Configuration of the architecture for the host model . . . . . . . . . . . . . 169
5.6 Summary of the difference in effect, strategy and mechanism between the
complementary concerns of tissue algorithms under discrete exposures. . . . 177
x
5.7 Summary of the exposure and re-exposure behaviour of the five defined tis-
sue exposure regimes and how they relate to system information availability
and anticipated need. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
5.8 Summary of general trends and their interpretation for Tissue Clonal Se-
lection Algorithms. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
5.9 Summary of the expected behaviours of the defined measures with regu-
lar and irregular exposure regimes on systems with mixing and isolated
information dissemination strategies. . . . . . . . . . . . . . . . . . . . . . . 179
5.10 Summary of the results for MTCSA and RTCSA-LRG on ATER and RTER
and the relative quantitative difference is measures. . . . . . . . . . . . . . . 185
5.11 Summary of confirmed behaviours of global and local information strategies
and exposure regimes, highlighting where expectations were incorrect. . . . 186
5.12 Summary of results from an empirical study into the HTCSA comparing
TER’s with the HTCSA-L information management strategy. . . . . . . . . 197
6.1 Summary of the different types of immunity. . . . . . . . . . . . . . . . . . . 210

6.2 Summary of the differences between the innate and acquired immune sys-
tems, taken from [257] (page 565). . . . . . . . . . . . . . . . . . . . . . . . 213
6.3 Summary of the difference in mechanism and strategy between the comple-
mentary concerns of host algorithms under discrete exposures. . . . . . . . . 226
6.4 Classification of Host Exposure Regimes in terms of general attributes. . . . 228
7.1 Mapping of the cellular, tissue, and host clonal selection perspectives onto
the secondary objects of the adaptive systems formalism. . . . . . . . . . . 263
7.2 Summary of the component cardinality across the principle abstract and
realised clonal selection algorithms, where ∗ represents a wild-card for de-
scendant algorithms of a given base algorithm. . . . . . . . . . . . . . . . . 265
7.3 Summary of the component interaction across the abstract and realised
clonal selection algorithms, where ∗ represents a wild-card for descendant
algorithms of a given base algorithm. . . . . . . . . . . . . . . . . . . . . . . 267
7.4 Summary of the component cardinality across the principle abstract and
realised antigenic exposure problems. . . . . . . . . . . . . . . . . . . . . . . 269
7.5 Summary of the bracketing of spatial exposure across the abstract and
realised antigenic exposure problems. . . . . . . . . . . . . . . . . . . . . . . 271
7.6 Summary of the bracketing of temporal exposure across the principle ab-
stract and realised antigenic exposure problems. . . . . . . . . . . . . . . . . 271
7.7 Summary of the suitability match between clonal selection and antigenic
exposure components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
7.8 Summary of the component cardinality for both the clonal selection and
antigenic exposure models. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
7.9 Summary of the Negative Selection AIS paradigm mapped onto the three
tiers of the IHCSF. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
7.10 Summary of the Immune Network AIS paradigm mapped onto the three
tiers of the IHCSF. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282
7.11 Summary of an interpretation of the field of Evolutionary Computation
in terms of the hierarchical reductionist methodology used to investigate
clonal selection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
7.12 Summary of an interpretation of the field of Connectionist Intelligence in
terms of the hierarchical reductionist methodology used to investigate clonal
selection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
xi
8.1 Summary of the projected applicability of Cellular Clonal Selection Algo-
rithms to Function Optimisation Problems. . . . . . . . . . . . . . . . . . . 308
8.2 Summary of the projected applicability of Tissue Clonal Selection Algo-
rithms to Function Optimisation Problems. . . . . . . . . . . . . . . . . . . 309
8.3 Summary of the projected applicability of Host Clonal Selection Algorithms
to Function Optimisation Problems. . . . . . . . . . . . . . . . . . . . . . . 310
8.4 Summary of the projected applicability of Cellular Clonal Selection Algo-
rithms to Function Approximation Problems. . . . . . . . . . . . . . . . . . 316
8.5 Summary of the projected applicability of Tissue Clonal Selection Algo-
rithms to Function Approximation Problems. . . . . . . . . . . . . . . . . . 316
8.6 Summary of the projected applicability of Host Clonal Selection Algorithms
to Function Approximation Problems. . . . . . . . . . . . . . . . . . . . . . 317
A.1 Summary of results from an empirical study into the RTCSA compared to
the MTCSA. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333
A.2 Summary of results from an empirical study into the RTCSA comparing
TER’s with the MTCSA local and RTCSA-LRG global information man-
agement strategies. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
A.3 Summary of results from an empirical study into the HTCSA. . . . . . . . . 335
A.4 Summary of results from an empirical study into the ITCSA. . . . . . . . . 336
A.5 Summary of results from an empirical study into the THCSA compared to
MP-HCSA. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
A.6 Summary of the pathogen transmission and vaccination transmission results
from an empirical study into the THCSA. . . . . . . . . . . . . . . . . . . . 338
A.7 Summary of results from an empirical study into the SI-HCSA compared
to MP-HCSA. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339
A.8 Summary of results from an empirical study into the MI-HCSA. . . . . . . 340
A.9 Summary of results from an empirical study into the EI-HCSA. . . . . . . . 341
xii
Chapter 1
Introduction
1.1 Motivation
Artificial Intelligence is concerned with the investigation of the information processing
mechanisms that underlie intelligent behaviour in systems that act rationally or like hu-
mans. The traditional approach toward designing and investigating such systems is a
rigorous top-down approach using a symbolic representation that provides clear indica-
tion why such systems work. A newer scruffy paradigm called Computational Intelligence
considers intelligence from a bottom-up perspective of strategies and outcomes. Exam-
ples of systems of this type are prescriptive indicating how they work, although they are
typically too complex to analyse to see why they work. The process for developing such
strategies, like many other fields of science and engineering, is to begin by patterning them
from biological and natural processes. Toward this end, the field of Artificial Immune Sys-
tems is concerned with abstracting and exploiting the computational patterns observed in
the structure and function of the immune system [109].
The immune system is vast in scale and complexity, comprised of specialised organs
and trillions of cells and molecules. The complexity of the system is commonly considered
in the same order as the central nervous system. It is responsible for maintaining the
homoeostasis of a host organism and in particular it is responsible for identifying and
eliminating invading micro-organisms called pathogen [81, 314]. The immune systems of
vertebrates has a capacity to adapt so as to acquire immunity and better respond in the
future to the same and similar infections. The immune system is both essential to a hosts
survival and a powerfully efficient garbage disposal. The precision mechanisms that exist to
keep a host alive can become corrupted and turn on and attack the host, an affliction that
underlies many types of autoimmune disease. The acquired immune system is responsible
for the manufacture of an enormous diversity of detector proteins called antibodies, such
that collectively a repertoire of these molecules can potentially detect many millions or
billions of different molecular structures. Interestingly the information required to prepare
such a potentially diverse army of protective protein conformations is far greater than the
number of genes encoded in a given vertebrates genome. This mismatch between the
limited information that a genome can represent and what the acquired immune system is
capable of detecting is achieved via an adaptive learning process described by the Clonal
1
Selection Theory [63]. This Darwinian inspired theory for antibody diversity begins with a
repertoire of general-purpose detectors that can partially detect most infections. A process
of selection of the fittest and cellular mitosis with mutations that affect what descendant
cells can identify, results in the specialisation of a host organisms detection repertoire to
its environment.
The learning, memory, and adaptive information processing properties of clonal selec-
tion are a ‘corner stone’ in the field of Artificial Immune Systems (AIS). The abstracted
clonal selection strategy has been embodied in optimisation, pattern recognition, and clas-
sification algorithms and provides a foundational improvement procedure for those AIS
that exploit the information processing attributes of other cell-centric immunological the-
ories. Beyond adaptation, clonal selection describes a process by which the acquired im-
mune system coordinates learning and memory of a lifetime of unknown infections without
central control or organisation. This adaptive process occurs autonomously and concur-
rently throughout the distributed tissues of a host organism. Interestingly the distributed
and decentralised attributes are an often cited advantage of patterning strategies from
the immune system, although much of the research in the field has focused on monolithic
systems and algorithms [417, 420]. There has been provoking work on the investigation of
distributed signature based computer security applications [212, 255], and in autonomous
navigation and group decision making in mobile robotics [238, 359, 268], although there
exists no clear framework to guide the realisation of such systems and little work on inves-
tigating the clonal selection paradigm toward this end. Finally, there is some consensus
among leaders in the field of a recent impasse in the general progress in the field of Artificial
Immune Systems. The same authors suggested a trend of turning back to the biological
system to abstract novel and more accurate models to replace the crude first generation
approaches with so-called second generation artificial immune systems [24, 389, 388, 401].
The thesis of this work is that the clonal selection paradigm can be elaborated to
exploit the distributed and decentralised information processing characteristics inherent
in the inspiring metaphor, and that the structure and function of the immune system
provide a path toward realising such computational models. Such information processing
characteristics are beneficial for classes of problems that can be addressed using parallel or
cooperative problem solving strategies, such as those problems partitioned along functional
or information availability axes. This work addresses the open problem of considering a
second generation of artificial immune systems, in this case distributed clonal selection
algorithms, by following the trend of devising novel and more accurate abstractions and
models from the inspiring biological system.
1.2 Hypothesis
The main hypothesis of this research is defined, as follows:
An in depth study of immunological structure and function, in particular the

distributed nature of the immune system, provides effective computational in-
spiration for the extension and improvement of existing clonal selection algo-
2
rithms.
The aim of this research is to improve the current understanding of the clonal se-
lection paradigm for adaptive information processing and more specifically the state of
distributed and decentralised approaches. This aim was pursued by using an understand-
ing of immunological structure and function to provide a conceptual framework for the
investigation. To verify the effectiveness of the investigated framework and clonal selection
algorithms the hypothesis was divided into a series of goals, defined in Section 1.3.
1.3 Goals
The research thesis is addressed in the context of five main research goals, as follows:
1. Identify a systematic methodology for realising a novel biological inspired

computational framework and models. Before the investigation of a novel immuno-
logical inspired framework and resultant models and algorithms, it is essential to identify
a systematic methodology to provide a set of well defined procedures as to how such a
framework may be realised and how the effectiveness of the models and algorithms may
be assessed.
2. Identify limitations with and elaborate upon the base clonal selection
paradigm. The clonal selection paradigm is a core information processing pattern in
cellular immunology and inspired computational models, therefore it is essential that both
the capabilities and limitations of existing adaptive and distributed computational models
are understood.
3. Identify immunological structures and/or functions which clearly exhibit

distributed information processing. The structures and functions of mammalian
immune system must be scrutinised through a lens of clonal selection for processes and
architectures that constitute distributed information processing, and formulated into ab-
stracted computational models.
4. Study the behaviours, capabilities, and limitations of devised computational

models. In order to assess the effectiveness of the devised abstracted models they must
be examined and understood with regard to their information processing capabilities,
limitations, and general behaviours.
5. Propose an integrated clonal selection framework that unites the base

and distributed variants. Finally, the adaptive clonal selection paradigm and realised
distributed models must be integrated into a framework such that their benefits and
limitations can be assessed given the success of its descriptive and predictive capabilities.
3
1.4 Contributions
This dissertation makes the following specific contributions:
1. The systematic elaboration and development of the existing clonal selection com-
putational paradigm as an explicit adaptive strategy and suggestion of additional
models based on the interaction of cells and molecules, called the Cellular Clonal
Selection Paradigm (Chapter 3 and Chapter 4).
2. The design and investigation of novel Spatial, Mediated, and Network Cellular Clonal
Selection Algorithms (Chapter 4).
3. The identification and abstraction of the lymphatic system and lymphocyte traf-
ficking as a framework for distributed Clonal Selection Algorithms called the Tissue
Clonal Selection Paradigm (Chapter 5).
4. The design and investigation of novel Recirculation, Homing, and Inflammation Tis-
sue Clonal Selection Algorithms (Chapter 5).
5. The identification and abstraction of host immunisation and evolution as a frame-

work for distributed Clonal Selection Algorithms called the Host Clonal Selection
Paradigm (Chapter 6).
6. The design and investigation of novel Transmission and Shared Immunity Popula-
tion Host Clonal Selection Algorithms, and the Maternal Immunity and Inherited
Immunity Generational Host Clonal Selection Algorithms (Chapter 6).
7. The integration of the three perspectives of the computational properties of clonal

selection (Cellular, Tissue, Host) into an Integrated Hierarchical Clonal Selection
Framework with both explanatory and predictive properties (Chapter 7).
8. The elicitation of salient features of the Cellular, Tissue, and Host Clonal Selection
Algorithms and systematic assessment of the suitability of application of the algo-
rithms to the foundational problem domains of Function Optimisation and Function
Approximation (Chapter 8).
1.5 Thesis Organisation

This section presents the organisation for this work, highlighting the objectives and setting
the scene for the remainder of the dissertation.
Chapter 2 provides background documentation on the general fields of study in which

this work is situated, specifically Computational Intelligence and Biologically Inspired
Computation. A clear definition of Artificial Immune Systems (AIS) is outlined highlight-
ing the computational and problem solving ends, and differentiating it from theoretical and
computational immunology. This chapter supports the contention of a cellular-theoretic
focus in the field of AIS, and the promise of patterns for distributed, decentralised and
4
autonomous systems. The three dominant paradigms of the field are reviewed as is the
state of distributed AIS, highlighting the lack of a suitable framework and limited elab-
oration of the clonal selection paradigm toward this end. Finally, this section outlines a
systematic research methodology for realising immune inspired algorithms, an experimen-
tal methodology to ensure systematic assessment of devised systems, and an integrated
methodology that focus on decomposition and an economy of models.
Chapter 3 provides a review of clonal selection theory providing a context for the effec-
tive interpretation of the state of the art of inspired computational models. A taxonomy
is presented of clonal selection algorithms highlighting the commonality, and more impor-
tantly open problems in the field. The placement and assessment of the field is considered
in the broader context of computational intelligence demarcating related research that con-
tributes to an improved understanding of the approach. The underlying computational
strategy is considered in the context of relevant adaptive systems theory demonstrating
the importance of the information environment of the strategy, interactions of compo-
nents, and emergent behaviours. The cellular perspective of the field is addressed and an
agenda is defined which involves the investigation and elaboration of the classical cellular
clonal selection perspective, and the consideration of a ‘host of tissues’ and a ‘population
of hosts’ as distributed cellular perspectives of the immunological theory.
Chapter 4 presents a conceptual framework for the clonal selection paradigm from a
cellular perspective called the Cellular Clonal Selection Paradigm. The framework includes
the principle components and base models that describe the current state of clonal selection
algorithms referred to as Quintessential Clonal Selection. Symmetry is provided to the
framework in the form of a complementary Antigenic Exposure Paradigm that encapsulates
the domain information to which a system governed by clonal selection is triggered and
to which it must respond. Exploiting the predictive properties of the cellular framework,
three clonal selection approaches are defined and investigated: a Spatial, Mediated, and
Network Clonal Selection Algorithm, that are empirically demonstrated on an antigenic
exposure inspired pattern recognition problem domain called Colour Space.
Chapter 5 considers the clonal selection paradigm in the context of the entire host
organism. The structure and function of the human immune system are reviewed, high-
lighting the important role of the lymphoid tissues and related systems in the trafficking
of immune cells around the body. The tissue architecture and cell mobility information
processing are abstracted into a Tissue Clonal Selection Paradigm that both describes the
cellular clonal selection paradigm as a single constrained instance of a tissue-based system,
and predicts a range of tissue-based clonal selection architectures and algorithms. Match-
ing the symmetry provided in the cellular paradigm, an Infection Exposure Paradigm is
defined to satisfy tissue-centric concerns of multiple concurrent points of exposure and
information acquisition. A Lymphocyte Recirculation, Lymphocyte Homing and Tissue
Inflammation Clonal Selection Algorithm are investigated and demonstrated on problem
instances drawn from the colour space pattern recognition domain. Generally, the empiri-
5
cal results confirm the expected localisation and dissemination of acquired information as
emergent behaviours from the decentralised cellular repertoires.
Chapter 6 considers the clonal selection paradigm in the context of a population of

host organisms with interacting immune systems. Immunisation and host evolution are
identified as the two primary facets of immune system interactions at the population level
that provide a different perspective on information dissemination and organisation than
the tissue paradigm. A Population and Generational Host Clonal Selection Paradigm are
proposed that are descriptive at both considering an instance of the tissue paradigm as
a constrained host system, and at integrating the related work on models of gene library
evolution. A Habitat Exposure Paradigm is defined to facilitate the asymmetric properties
of population-based exposure and transmission. The predictive qualities of the framework
are exploited in the design and empirical investigation of four host-based algorithms on
problem instances from the colour space domain. These include a Transmission and
Shared Immunity algorithms from the Population paradigm, and the Maternal Immunity
and Evolved Immunity algorithms from the Generational paradigm. Results confirm the
expectation of the higher level localisation and dissemination of acquired information
emergent behaviours from the decentralised systems using varied structures and functions
to those used in the Tissue paradigm.
Chapter 7 provides a systematic integration firstly of the computational clonal selec-

tion models into a Hierarchical Clonal Selection Framework, and secondly of the exposure
paradigm into a Hierarchical Antigenic Exposure Framework. Each framework provides a
one-sided perspective for the development and investigation of clonal selection approaches
and the mapping of complementary exposure problem domains. The integration of the
frameworks highlight both the specific design decisions that bound and control the com-
plexity of the investigated algorithms. In addition to explanatory power, the Integrated
Hierarchical Clonal Selection Framework suggests at the potential for the development and
investigation of additional models, algorithms, and even new paradigms inspired by the
structure and function of the immune system that constrain the computational concerns
of clonal selection. The generality of the immunologically-centric framework is considered
as a model for the broader study of Artificial Immune Systems, and the methodology
adopted to develop and investigate the framework is considered as a model for related
Computational Intelligence fields of research.
Chapter 8 considers the applicability of the developed clonal selection algorithms. A

feature-based methodology for addressing the suitability of an approach is proposed and
adopted to elicit the salient features from the algorithms from across the three clonal selec-
tion paradigms. Two foundational problem domains from Computational Intelligence are
considered and systematically reviewed: Function Optimisation and Function Approxima-
tion. The salient features of each domain are described and elaborated in the context of
specific application examples that demonstrate information processing properties relevant
to the clonal selection algorithms. The applicability of the algorithms from each clonal
6
selection paradigm is defined by the overlap in features between specific approaches and
problem instances, suggesting at an abundance of ongoing application-centric research.
Chapter 9 provides a detailed assessment of the research hypothesis by highlighting the

specific examples of how each research goal from Section 1.3 was addressed in this work.
Immunological structure and function are demonstrated as providing an effective path for
both the elaboration of the classical clonal selection computational paradigm, and as a
path toward realising decentralised and distributed extensions of the paradigm. The con-
tributions of the work are rephrased in terms of computational and conceptual artefacts,
highlighting the discrete and standalone products that may be differentiated and drawn
from this work. Three key methodological limitations of the project are considered as ac-
ceptable trade-off’s in addressing the specific research question that highlight alternative
and potentially future approaches of investigating the emerging sub-field of study. Finally,
the hierarchical perspective of the work is used to organise and systematically suggested
extensions to the project at three levels of abstraction including Framework, Paradigm,
and Algorithm, outlining an ongoing research agenda for the study of Clonal Selection
Algorithms.
7
Chapter 2
Background
2.1 Chapter Overview

This chapter provides background information and reviews the base motivations for the
work. Section 2.2 generally positions the work within the sub fields of Biologically In-
spired Computation and Computational Intelligence under the broader field of Artificial
Intelligence. Section 2.3 clarifies this further by outlining the specific field of Artificial
Immune Systems. Section 2.4 reviews the core paradigms in the field, highlighting the
cellular focus that inspires the computational approaches, and the general fundamental
role the adaptive clonal selection strategy has across the field. Section 2.5 motivates the
specific research hypothesis of this work in the development of distributed Artificial Im-
mune Systems, and how the this decentralised adaptive property is generally promoted
in the field, although remains underdeveloped. Section 2.6 reviews open problems in the
field, highlighting the general impasses and lack of satisfying progress and advocating the
specific approach taken in this work of turning back to the motivating metaphor toward
developing more sophisticated approaches. Finally, a systematic methodology is pieced
together in Section2.7 from best practices with regard to (1) biologically inspired algo-
rithm development, (2) experimental investigation, and (3) the economy of modelling and
integration of findings. This chapter sets the scene for the following chapter that out-
lines the specific elaborations and distributed realisations of clonal selection investigated
throughout the remainder of the dissertation.
2.2 Artificial and Computational Intelligence

2.2.1 Artificial Intelligence
The general field of study is the multi-disciplinary field of Artificial Intelligence (AI). Rus-
sell and Norvig provide a perspective that defines Artificial Intelligence in four categories:
(1) systems that think like humans, (2) systems that act like humans, (3) systems that
think rationally, (4) systems that act rationally [342]. In the definition, acting like a
human suggests that a system can do some specific things humans can do, this includes
fields such as the Turing test, natural language processing, automated reasoning, knowl-
8
edge representation, machine learning, computer vision, and robotics. Thinking like a
human suggests systems that model the cognitive information processing properties of
humans, for example a general problem solver and systems that build internal models of
their world. Thinking rationally suggests laws of rationalism and structured thought, such
as syllogisms and formal logic. Finally, acting rationally suggests systems that do rational
things such as expected utility maximisation and rational agents. Luger and Stubblefield
suggest that AI is a sub-field of computer science concerned with the automation of intel-
ligence, and like other sub-fields of computer science has both theoretical (how and why do
the systems work? ) and application (where and when can the systems be used? ) concerns
[276]. They suggest a strong empirical focus to research, such that although there may
be a strong desire for mathematical analysis, the systems themselves defy analysis due
to their complexity. The machines and software themselves are not black boxes, rather
analysis proceeds by observing the systems interactions with their environment, followed
by an internal assessment of the system to relate their structure back to their behaviour.
Artificial Intelligence is therefore concerned with investigating mechanisms that un-
derlie intelligence and intelligence behaviour. The traditional approach toward designing
and investigating AI (the so-called ‘good old fashioned’ AI) has been to employ a symbolic
basis for these mechanisms. A newer approach historically referred to as messy artificial
intelligence or or soft computing does not use a symbolic basis, instead patterning these
mechanisms after biological or natural processes. This represents a modern paradigm shift
in interest from symbolic knowledge representations, to inference strategies for adaptation
and learning, and has been referred to as neat versus scruffy approaches to AI. The neat
philosophy is concerned with formal symbolic models of intelligence that can explain why
they work, whereas the scruffy philosophy is concerned with intelligent strategies that
explain how they work [364].
2.2.2 Computational Intelligence

A modern name for the sub-field of AI concerned with sub-symbolic (messy, scruffy, soft)
mechanisms is Computational Intelligence. This name provides a banner which groups
four principle approaches: Fuzzy Intelligence, Connectionist Intelligence, Evolutionary
Intelligence, and Swarm Intelligence [135, 317]. A second popular and general name for the
strategy-outcome perspective of scruffy AI is Metaheuristics, which evolved from the neater
field of heuristics methods applied in Operations Research. A metaheuristic is defined as
a general algorithmic framework which can be applied to different optimisation problems
with relative few modifications to make them adapted to a specific problem [48]. Another
important perspective is that provided by the field of Machine Learning that focuses on
the learning properties of Artificial Intelligence. The term is commonly used to describe
inductive model building techniques (that generalise from specific observations) that are
applied to ‘learn’ relationships in data sets (the application of which is referred to as Data
Mining [433]), with or without supervision (corrective behaviour) [293]. The following
provides a summary of the four primary areas of study in Computational Intelligence:
9
Evolutionary Computation A paradigm that is concerned with the investigation of
systems inspired by the neo-Darwinian theory of evolution by means of natural selection.
Popular evolutionary algorithms include the Genetic Algorithm, Evolution Strategy, Ge-
netic and Evolutionary Programming, and Differential Evolution. The evolutionary pro-
cess is considered an adaptive strategy and is typically applied to search and optimisation
domains.
Swarm Intelligence A paradigm that considers collective intelligence emerges through

the interaction and cooperation of large numbers of lesser intelligent agents. The paradigm
consists of two dominant sub-fields (1) Ant Colony Optimisation that investigates prob-
abilistic algorithms inspired by the stigmergy and foraging behaviour of ants, and (2)
Particle Swarm Optimisation that investigates probabilistic algorithms inspired by the
flocking and foraging behaviour of birds and fish. Like evolutionary computation, swarm
intelligences are considered adaptive strategies and are typically applied to search and
optimisation domains.
Connectionist Intelligence An approach that is concerned with the investigation of

architectures and learning strategies inspired by the modelling of neurons in the brain
called Artificial Neural Networks. Learning strategies are typically divided into supervised
and unsupervised which manage environmental feedback in different ways. Neural network
learning processes are considered adaptive learning and are typically applied to pattern
recognition domains.
Fuzzy Intelligence An approach that is concerned with the investigation of fuzzy logic
which is a form of logic that is not constrained to true and false, but rather functions
which define degree’s of truth. Fuzzy logic is reasoning strategy and is typically applied
to expert system and control system domains.
2.2.3 Natural Computation

An important perspective on scruffy Artificial Intelligence is the motivation and inspira-
tion for the core information processing strategy of a given technique. Computers can
only do what they are instructed, therefore a consideration of Computational Intelligence
is to distil information principles and strategies from other fields of study, such as bi-
ology. The study of biologically motivated Computational Intelligence may be called
Biologically Inspired Computing [118], and is one of three related fields of Natural Com-
puting [145, 146, 313]. Natural Computing is an interdisciplinary field concerned with the
relationship of computation and biology, which in addition to Biologically Inspired Com-
puting also comprises of Computationally Motivated Biology and Computing with Biology
[316, 281, 2]. Therefore, the field of Artificial Intelligence, specifically the scruffy variety
of Computational Intelligence motivates this work from the perspective of an intelligent
problem-solving strategy in computer science, whereas the field of Natural Computing,
10
specifically the Biological Inspired Computation variety motivates the actual principles
and information processing capabilities of the strategy.
Biologically Inspired Computation (Computation inspired by biological metaphor )

The intent is to devise mathematical or engineering tools to address problem domains.
Biologically Inspired Computation fits into this category, as do other non-computational
areas of problem solving not discussed. At its simplest, its using solutions (a procedure
for finding solutions is considered a solution) found in the biological environment.
Computationally Motivated Biology (Biology with digital computers) The intent of

this area is to use information sciences and simulation to model biological systems in dig-
ital computers with the aim to replicate and better understand behaviours in biological
systems. The field facilitates the ability to better understand life-as-it-is and investigate
life-as-it-could-be. Typically, work in this sub-field is not concerned with the construction
of mathematical and engineering tools, rather it is focused on simulating natural phenom-
ena. Common examples include Artificial Life, Fractal Geometry (L-systems, Iterative
Function Systems, Particle Systems, Brownian motion), and Cellular Automata.
Computing with Biology The investigation of substrates other than silicon in which
to implement computation. Common examples include molecular or DNA Computing and
Quantum Computing.
2.3 Natural and Artificial Immune Systems

This section reviews the biological immune system which motivates the biologically in-
spired field of Artificial Immune Systems. The varied approaches in this field have lead
to some debate as to how it relates to sibling Computational Intelligence sub-fields, com-
prising features from Evolutionary Computation, Swarm Intelligence, and Connectionist
Intelligence. These relationships are further considered in Section 3.4.
2.3.1 Overview of the Immune System

This review of the immune system with a focus on the general problem to which the
biological system is attributed to addressing, which is to protect the host organism from
the threats posed to it from pathogens and toxic substances. Pathogens encompass a range
of micro-organisms such as bacteria, virus, parasites and pollen. This review provides
sufficient immunology such that a computer scientist may grasp the core architecture,
processes and general problem of addressed by the immune system. The focus is the
vertebrate immune system, specifically that of mammals given that disproportionally more
information is known about the system. For more general information, the reader is
directed to two addition reviews of immunology by and for Computer Scientists, de Castro
and Timmis [110] (Chapter 2), and Hofmeyr [213].
11
The traditional perspective regarding the role of the immune system is divided into
two primary tasks: the detection and elimination of pathogen1 . This behaviour is typi-
cally referred to as the differentiation of self (molecules and cells that belong to the host
organisms) from potentially harmful non-self. The architecture of the immune system is
such that a series of defensive layers protect the host. The most basic of these layers
provide physical barriers to prevent pathogen from entering the host such as skin, and
mucous membranes that trap and filter out pathogen. The next level is physiological
that provides inhospitable conditions for pathogen at entry points to host such as high pH
acidity in the digestive tract and varied temperature to inhibit growth of micro-organisms.
Once a pathogen makes it inside the host, it must contend with the innate and acquired
immune system. These interrelated immunological sub-systems are comprised of many
types of cells and molecules produced by specialised organs and processes to address the
self-nonself problem at the lowest level using chemical bonding, where the surfaces of cells
and molecules interact with the surfaces of pathogen.
Innate Immunity
The innate immune system is named such because the host organism is born with a set of
fast acting detection and elimination mechanisms that do not change over the hosts life-
time. In this regard, the mechanisms are reflexive providing a first line of defence against
pathogen. Innate mechanisms are the most primitive found in all classes of plant and
animal life improving in a species over evolutionary time. The complement system refers
to a mechanism where associated molecules bind to and coat specific types of bacteria.
The activation of this mechanism (a complement cascade) can lead to a rupturing of the
bacteria called lysis, or detection and subsequent killing of the bacteria by macrophage
cells called opsonisation. The process is fast acting, typically occurring within a the first
few hours of an infection. Macrophages are scavenger cells responsible for cleaning up
cellular debris, some bacteria, and micro-organisms that have been identified by the com-
plement system. These cells engulf material and neutralise it using a digestive process.
When a virus infects a cell, the cell may release chemicals called cytokines that provide an
indication to other cells of the infection. Natural Killer (NK) cells are a specialised class of
immune cell responsible for identifying virally infected cells via their chemical signals and
releasing chemicals that trigger apoptosis (programmed cell death) of these cells. These
NK cells are also responsible for identifying and eliminating tumour cells.
Adaptive Immunity
The adaptive immune system, also referred to as the acquired immune system, is named
such because it is responsible for specialising a defence for the host organism based on
the specific pathogen to which it is exposed. Unlike the innate immune system, the
acquired immune system is present only in vertebrates (animals with a spinal column).
The system retains a memory of exposures which it has encountered. This memory is
1
More recent perspectives on the role of the system include a maintenance system [81], and a cognitive
system [403].
12
recalled on reinfection exhibiting a learned pathogen identification. This learning process
may be divided into two types of response. The first or primary response occurs when
the system encounters a novel pathogen. The system is slow to respond, potentially
taking a number of weeks to clear the infection. On re-encountering the same pathogen
again, the system exhibits a secondary response, applying what was learned in the primary
response and clearing up the infection rapidly. The memory the system acquires in the
primary response is typically long lasting, providing pathogenic immunity for the lifetime
of the host, two common examples of which are the chickenpox and measles. White
blood cells called lymphocytes (or leukocytes) are the most important cell in the acquired
immune system. Lymphocytes are involved in both the identification and elimination of
pathogen, and recirculate within the host organisms body in the blood and lymph (the fluid
that permeates tissue). The behaviour of the adaptive immune system may be described
with regard to two general sub-systems: Humoral Immunity for addressing free-floating
pathogen (for example in blood), and Cellular Immunity for addressing virally infected
cells.
Humoral Immunity B-lymphocyte cells are a specialised class of white blood cell that
are created in the bone marrow and are predominantly responsible for the secretion of
antibodies. These are protein molecules that bind to parts or specialised features of
pathogen called pathogenic determinants. B-cells are activated by identifying pathogen via
their surface bound receptors which provide a complementary shape for pathogen surface
features, fitting like a lock and a key. Once activated a B-cell will produce vast quantities
of antibody which enter the blood stream and recirculate around the host organism to
clear up the infection.
Cellular Immunity T-lymphocyte cells are another class of white blood cell that are
matured in the thymus (an immune organ located in the chest). T-cells perform specific
roles both mediating the activation of B-cells (Helper T-cells) and in the identification and
neutralisation of virally infected cells (Killer T-cells). In this latter role, T-cells identify
processed (chopped up) pieces of pathogen on the surfaces of infected cells called histo-
compatibility complex (commonly referred to by its abbreviation: MHC). Cells responsible
for processing pathogen and presenting MHC are appropriately called antigen-presenting
cells, examples of which include some B-cells, macrophages and dendritic cells.
2.3.2 Artificial Immune Systems

History
Artificial Immune Systems (AIS) is a sub-field Computational Intelligence motivated by

immunology (primarily mammalian immunology) that emerged in the early 1990’s (for
example [40, 227]), based on the proposal in the late 1980’s to apply theoretical immuno-
logical models to machine learning and automated problem solving (such as [210, 138]).
The early works in the field were inspired by exotic theoretical models (immune network
theory) and were applied to machine learning, control and optimisation problems. The
13
approaches were reminiscent of paradigms such as Artificial Neural Networks, Genetic
Algorithms, Reinforcement Learning, and Learning Classifier Systems. The most forma-
tive works in giving the field an identity were those that proposed the immune system
as an analogy for information protection systems in the field of computer security. The
classical examples include Forrest, et al.’s Computer Immunity [151, 148] and Kephart’s
Immune Anti-Virus [248, 249]. These works were formative for the field because they
provided an intuitive application domain that captivated a broader audience and assisted
in differentiating the work as an independent sub-field.
Motivation
The motivation for the field of AIS is the immune system, specifically the architecture,
mechanisms, principles, and theories used to explain observed immunological function.
The authors de Castro and Timmis defend the immune system as an inspiration for com-
putational intelligence by providing a comprehensive listing of abstracted information
processing principles [110] (pages 55-56). This listing contains nineteen desirable compu-
tational attributes, the following of which specifically motivate this research:
• Uniqueness: Each individual with an immune system has distinct and different
capabilities and vulnerabilities.
• Disposability: There is no single cell or molecule that is essential to the functioning

of the immune system.
• Autonomy: There is no central or coordinating organ controlling the immune system
• Distributivity: The immune cells, molecules, and organs are distributed all over the
body and are not subject to centralised control.
• Fault Tolerance: The complementary roles performed by several immune components

allow the re-allocation of resources in the event of failure.
• Self-Organisation: There is no specific information as to how to respond to a given

pathogen, the system responds locally providing a global effect of defence.
Forrest and Hofmeyr take a similar approach in considering and abstracting the in-
formation processing principles of the immune system [147]. In their work, the authors
focus on three specific information-processing principles, although they highlight a num-
ber of general design principles that have much overlap with the de Castro-Timmis listing.
Perhaps the more important features that strongly entrench the field of Computational
Intelligence are Immune Learning (acquired information through interaction with the en-
vironment) and Immune Memory (persistence and ongoing application of acquired infor-
mation for short and long durations), both of which fit into a broad notion of intelligence.
Using this general description, the mammalian acquired immune system may be considered
rudimentary intelligent.
14
Definitions
Although the motivations of the field are easy to relate, it is important to clarify those
motivations that do not contribute to the specific field. This section focuses on some
standard definitions of what is and is not an Artificial Immune System. The authors
de Castro and Von Zuben [113] define an AIS as “a computational system based upon
metaphors of the biological immune system”. They continue by defining Immune Engi-
neering as the application of an AIS in “a meta-synthesis process that uses the information
contained in the problem itself to define the solution tool to a given problem, and then ap-
ply it to obtain the problem solution”. The distinction is that Immune Engineering is the
engineering process, differentiated from conventional engineering process via its biological
(immunological) motivation. Immune Engineering was later formalised into a framework
by de Castro and Timmis as a set of three principles of what an Artificial Immune System
should contain [109], as follows: (1) A representation for the components of the system,
(2) A set of mechanisms to evaluate the interaction of individuals with the environment
and each other, and (3) Procedures of adaptation that govern the dynamics of the system.
A further, more concise definition of AIS is provided in the same work as “adaptive
systems, inspired by theoretical immunology and observed immunological functions, prin-
ciples and models, which are applied to problem solving”. This definition clearly highlights
the role of theoretical immunology and immunological observations as an inspiration and
not the pursuit of research in AIS, and is the definition adopted in this dissertation. In
addition to the clarification it provides, the authors propose that a system to be referred
to as an AIS must incorporate a minimum level of immunology, such as an immune com-
ponent (for example cell, molecule, and organ), it has to be designed by incorporating
ideas from theoretical and/or experimental immunology, and it has to be directed toward
problem solving. They stipulate that attributing immunological terminology is insufficient
to call a system an artificial immune system.
Another perspective of AIS is provided by one of the pioneers in the field Ishida who
defines an Immunity-Based System (IMBS) as a “self-maintenance systems learned from
and inspired by the immune system” [229]. His IMBS is described as (1) a self-maintenance
system with monitoring of self and nonself, (2) a distributed system with autonomous com-
ponents capable of mutual evaluation, and (3) an adaptive system with diversity and se-
lection. The three properties of Ishida’s AIS complement the de Castro-Timmis definition
by highlighting core information processing characteristics as a baseline when abstracting
from the biological system that may be distilled to (autonomy, distributed, adaptive).
2.4 Three Schools of AIS

This section provides a summary of the field of Artificial Immune Systems. The review
presents a taxonomy that covers the three core paradigms employed that collectively cover
the majority of the work in the field: Clonal Selection, Negative Selection, and Immune
Network. Each paradigm is presented from the perspective of immunological theory of
immune cell interactions, the abstracted information processing principles, and the al-
15
gorithms and applications that have been proposed. As such, the review highlights the
central role the adaptive clonal selection theory plays across the principles paradigms in
the field.
2.4.1 Clonal Selection Paradigm

The clonal selection theory credited to Frank Macfarlane Burnet was proposed to ac-
count for the behaviour and capabilities of antibodies in the acquired immune system
[63, 64]. Inspired by the principles of Darwinian natural selection theory of evolution,
the theory proposes that antigens select-for lymphocytes (both B and T-cells). When a
lymphocyte is selected and binds to an antigenic determinant, the cell proliferates mak-
ing many thousands more copies and differentiates into different cell types (plasma and
memory cells). Plasma cells have a short lifespan and produce vast quantities of antibody
molecules, whereas memory cells live for an extended period in the host anticipating future
recognition of the same determinant. The important feature of the theory is that when
a cell is selected and proliferates, it is subjected to small copying errors (changes to the
genome called somatic hypermutation) that change the shape of the expressed receptors
and subsequent determinant recognition capabilities of both the antibodies bound to the
lymphocytes cells surface, and the antibodies that plasma cells produce.
The theory is interesting from an information processing perspective. It suggests that
starting with an initial repertoire of general immune cells, the system is able to change itself
(the compositions and densities of cells and their receptors) in response to experience with
the environment. Through a blind process of selection and accumulated variation on the
large scale of many billions of cells, the acquired immune system is capable of acquiring the
necessary information to protect the host organism from the specific pathogenic dangers of
the environment. It also suggests that the system must anticipate (guess) at the pathogen
to which it will be exposed, and requires exposure to pathogen that may harm the host
before it can acquire the necessary information to provide a defence.
The information processing principles of the clonal selection theory describe a gen-
eral learning strategy. The theory has inspired a suite of algorithms for optimisation,
classification, and other rudimentary machine learning problem domains. In each algo-
rithm, a population of adaptive information units (each representing a problem-solution
or component thereof) is subjected to a competition processes for selection, which to-
gether with the resultant duplication and variation, ultimately improves the information
units, and solves or approximately solves a problem. The seminal algorithm was proposed
by de Castro called the CLONal selection ALGorithm (CLONALG) [114, 117] and was
applied to Function Optimisation, the Travelling Salesman Problem (a combinatorial op-
timisation problem) and Binary Character Pattern Recognition. Other algorithms include
the B-cell Algorithm (BCA) [247, 246], and the Multi-objective Immune System Algo-
rithm (MISA) [79, 83] for optimisation, and the Artificial Immune Recognition System
(AIRS) by Watkins [418, 421] for supervised classification. This paradigm is elaborated
in Chapter 3.
16
2.4.2 Negative Selection Paradigm
An important consideration for Burnet in developing the clonal selection theory was the
integration of the fact that such a powerfully destructive process as acquired immunity
does not immediately turn against the host organism in the generation of self-antibodies
(autoimmune response). This ability of the clonal response to maturate lymphocyte cells
for pathogen and not self-tissues became known as self-nonself discrimination [86, 305],
and in conjunction with clonal selection forms paradigms by which the acquired immune
systems is understood in the field of immunology. Clonal selection accounts for this ob-
servation by proposing both that (1) the initial repertoire of immune cells are prepared in
such a way that none are autoimmune, and (2) that during the maturation process in the
ongoing interaction with pathogen, that autoimmune clones are destroyed. Immune cells
are continually being destroyed, therefore there is a constant stream of new immune cells
that must be created and integrated into the repertoire. This theory was confirmed by
the discovery of T-lymphocyte cells that are matured using both a positive and negative
selection process in the thymus.
Self-nonself discrimination primed with a negative selection process has interesting
information processing properties. The theory suggests that the anticipatory guesses made
in clonal selection are filtered by regions of infeasibility (protein conformations that bind
to self-tissues). Further, the self-nonself immunological paradigm proposes the modelling
of the unknown domain (encountered pathogen) by modelling the complement of what is
known. This is unintuitive as the natural inclination is to categorise unknown information
by what is different from what is known rather than guessing at the unknown information
and filtering guesses by what is known.
The information processing principles of the self-nonself discrimination via negative
selection are that of a anomaly and change detection system that model the anticipation
of variation from what is known. The seminal negative selection algorithm was proposed
by Forrest, et al. [151] in which a population of detectors are prepared in the presence of
known information, where those randomly generated detectors that match against known
data are discarded. The population of pattern guesses in the unknown space then monitors
the corpus of known information for changes. The algorithm was applied to the monitoring
of files for changes (corruptions and infections by computer viruses), and later formalised
as a change detection algorithm [129, 130]. It was applied to monitoring changes in the
execution behaviour of Unix processes [149, 215], and to monitor the changes in remote
connections of a network computer (intrusion detection) [211, 212]. The application of the
algorithm has been predominantly to virus host intrusion detection and their abstracted
problems of classification (two-class) and anomaly detection. More recently, the validity
of the application of negative selection algorithms in high-dimensional spaces has been
questioned, specifically given the scalability of the approach in the face of the exponential
increase in volume within the problem space [374].
17
2.4.3 Immune Network Paradigm
A concern of the clonal selection theory is that it presumes that the repertoire of reac-
tive cells remains idle when there are no pathogen to which to respond. Considering this
problem, Niels Jerne proposed an Immune Network Theory (Idiotypic Networks) where
immune cells are not at rest in the absence of pathogen, instead antibody and immune
cells recognise and respond to each other [233, 234, 235]. The theory proposes that anti-
body (both free floating and surface bound) possess idiotopes (surface features) to which
the receptors of other antibody can bind. As a result of receptor interactions, the reper-
toire becomes dynamic, where receptors continually both inhibit and excite each other
in complex regulatory networks (chains of receptors). The network theory suggests that
the clonal selection process may be triggered by the idiotopes of other immune cells and
molecules in addition to the surface characteristics of pathogen, and that the maturation
process applies both to the receptors themselves the idiotopes which they expose.
The immune network theory has interesting resource maintenance and signalling in-
formation processing properties. The clonal selection and negative selection paradigms
integrate the accumulative and filtered learning of the acquired immune system, whereas
the immune network theory proposes an additional order of complexity between the cells
and molecules under selection. In addition to cells that interact directly with pathogen,
there are cells that interact with those reactive cells and with pathogen indirectly, in suc-
cessive layers such that networks of activity for higher-order structures such as internal
images of pathogen (promotion), and regulatory networks (so-called anti-idiotopes and
anti-anti-idiotopes).
Early works, such as Farmer, et al. [138] suggested at the exploitation of the infor-
mation processing properties of network theory for machine learning. A seminal network
theory based algorithm was proposed by Timmis, et al. called the Artificial Immune Net-
work (AIN) [392] (later extended and renamed the Resource Limited Artificial Immune
System [393, 391] and Artificial Immune Network [258]). The algorithm maintains a popu-
lation of adaptive units that operate under a clonal selection process. Units are connected
to each other based on distance measures in the information space (for example Euclidean
distance if the units are real-valued vectors), such that adaptive units that are close in the
information space suppress the activation of each other. Therefore, the more connections
a given unit has with other nodes influences the units potential for being selected and
cloned under clonal selection. Timmis’ system was applied to feature extraction and data
clustering. Another seminal network based algorithm was proposed by de Castro and Von
Zuben called Artificial Immune Network (aiNet) which extended the principles of AIN
and CLONALG and was applied to clustering [115, 116]. The aiNet algorithm was further
extended to optimisation domains and renamed opt-aiNet [108].
2.4.4 Summary
The clonal selection, negative selection, and immune network paradigms represent the
majority of the influences for the information processing properties from immunology
exploited in the field of artificial immune systems. The proposed taxonomy is not complete,
18
for example an additional influence that is not discussed is the Danger Theory of Polly
Matzinger [287, 288] which proposes that the acquired immune system responds to signals
of damage, which opposes the fundamental self-nonself paradigm. This coarse taxonomy
of the field of Artificial Immune Systems clearly highlights the central role clonal selection
to both the cognitive properties assigned to the immune system (learning and memory),
and the three principle paradigms that underlie the field. This centrality and even reliance
on clonal selection suggests that developments and improvements to the understanding of
the paradigm will have follow-on effects to the other cell-based paradigms in the field.
2.5 Distributed Artificial Immune Systems

One of the often cited propositions of using the immune system as an inspiration for com-
putational systems is that it affords decentralised and distributed information processing.
This section reviews this promise of the field of Artificial Immune Systems, and reviews
systems and algorithms that have made progress toward this end.
2.5.1 Promise of Distributed Information Processing

The terms autonomous, decentralised, and distributed are used with abandon in the field
of Artificial Immune Systems to both describe information processing abstractions of the
immune system and the desired characteristics of inspired systems. For example, the mo-
tivations and definitions for the in Section 2.3.2 both feature the distributed aspects of the
immune system. Watkins, et al. described the “inherent distributedness” of the biological
immune system, and the “almost embarrassingly parallel nature” of the clonal selection al-
gorithm [417]. Watkins and Timmis further pushed this perspective and commented that
“Among the oft-cited reasons for exploring the mammalian immune system as a source of
inspiration for computational problem solving include the observations that the immune
system is inherently parallel and distributed ” [420]. Therefore, it is not unreasonable to
consider that a promise of the field of Artificial Immune Systems is a source of distributed
information processing paradigms, systems, and/or algorithms, imbued with notions of
autonomy, decentralisation, parallelism, concurrency.
In this work, the terms are used with explicit meaning. Autonomy refers to the self-
governance of discrete facets of the system such as the molecules and cells and to the
processes in which they are engaged for example the identification and neutralisation of
pathogen. Decentralised refers to the lack of centralised control, specifically the local-
isation of the self-governance such as at the spatial and temporal characteristics of a
pathogenic encounter. Finally, distributed is used in the systems sense of the integra-
tion and coordination of the localised self-governance of resources towards a holistic end,
such as homoeostasis and general organism defence from infection. The definitions used
align with the pursuit of Computational Intelligence systems that exploit the distributed
information processing with multiple interacting and intelligent agents referred to as Dis-
tributed Artificial Intelligence and Multi-Agent Systems [425]. From Weiss’ definition, the
autonomy, emergent cooperation and intelligence are preserved although the constraints of
19
rationality, local determinism, and goal or task-directed behaviour are relaxed in favour of
the biologically inspired bottom-up and stochastic mechanisms of Computational Intelli-
gence. Therefore, a distinction is made between centralised, distributed, and distributable,
where a system may be designed or re-engineered to be distributed although the benefits
of such a process are not limited to concurrent, parallel, or distributed implementation in
the computer hardware sense.
2.5.2 Algorithms, Systems, and Applications

This section reviews the efforts towards the realisation of distributed Artificial Immune
Systems. Considered in this review are the seminal works and main efforts toward such
realisation which are categorised as (1) distributed negative selection approaches for com-
puter security, (2) efforts toward autonomous and decentralised control systems and robotics,
(3) network systems and application domains, (4) parallel extensions of existent algo-
rithms, and (5) miscellaneous works.
Distributed Detectors for Computer Security
In their proposal of a distributed detection system based on the negative selection al-
gorithm Forrest, et al. comment on the characteristics of the acquired immune systems
pathogen identification and memory capabilities, highlighting that they are “highly dis-
tributed ”, and that responses are decentralised and “highly localized ” [148, 212]. Hofmeyr,
et al. extended the approach and applied it to the domain of computer network intrusion
detection [211]. In their system, a central negative selection process (central tolerisation)
prepares a private set of detectors for each machine on a computer network such that the
diversity of each machines detector set contributes to the protection of the entire network.
The system was simulated demonstrating the principle, and was implemented although
not in a distributed manner. The research was abstracted to a general framework for
distributed adaptive systems called the ARTificial Immune System (ARTIS), with the
application to intrusion detection renamed the Lightweight Intrusion Detection System
(LISYS) [214]. The detectors of LISYS involve a life-cycle that extends the negative selec-
tion algorithm including a naı̈ve state for untested detectors, a memory state for proven
detectors, and an activation threshold used to manage the sensitivity of raised intrusion
signals. Three specific distributed characteristics of acquired immunity are not realised in
LISYS, specifically (1) the detectors are not mobile although memory detectors are shared
with neighbouring systems, (2) the detectors do not undergo a clonal selection process
when activated, and (3) the system is not autonomous requiring the intervention of a
human operator in the promotion of activated detector to memory detector.
The Computer Defence Immune System (CDIS) proposed an elaborate multi-agent
system with much commonality to ARTIS and LISYS [432, 197, 17, 18]. Among the
many enhancements to LISYS, the CDIS intrusion detection system facilitated the hier-
archical dissemination of acquired information, specifically the signatures for detectors of
verified intrusion attempts. Kim, et al. proposed an intricate and integrated system of
artificial immune algorithms for intrusion detection [252, 253, 254, 255, 251]. The work
20
approached the problem of network intrusion detection in the same manner as LISYS and
CDIS, although it employed gene library evolution to improve detector generation during
centralised negative selection, and a dynamic clonal selection algorithm to improve detec-
tors successful at identifying anomalies. Although the system was designed for distributed
intrusion detection, it was not simulated or implemented in a distributed environment.
Finally, Dasgupta and Brian proposed a multi-agent system architecture for intrusion de-
tection with mobile agents on a network where anomaly detection was achieved through
a loose analogy of distributed immune cells [107].
Control, Robots, and Autonomous Navigation
Early contributions to the field of AIS were made by Ishida, et al. by phrasing distributed
fault diagnosis problems in the connectionist paradigm of Parallel Distributed Processing,
and exploring the capabilities of Jerne’s network theory on such problems [227]. The
models were graphs of sensor nodes that performed distributed information processing
(sensor fault diagnosis) without a central point of control [295]. Ishida expanded upon the
work to exploit the systems autonomous cooperative decision making and self-organising
capabilities in domains such as coin balancing and flow monitoring [228].
Lee and Jun, et al. proposed a Distributed Autonomous Robotic System that achieved
cooperative control of robots through emergent swarm behaviour for the selection of group
behaviour strategies in autonomous robots [270, 238]. The system was modelled such that
each robot is a B-cell that expresses behaviour strategies antibodies in an environment of
conditions (pathogens), with control parameters managed as T-cells. Robots select their
own behaviour strategies which are communicated with each other, propagating (clonal
selection proliferation) if useful, or suppressed if not, in a simulated immune network. The
robots were simulated in a dynamic environment to perform group tasks collectively se-
lecting action strategies. A similar approach inspired by Lee and Jun, et al. was employed
by Meshref and VanLandingham and applied to the benchmark dynamic control dog-and-
sheep problem, where dogs (B-cells) coordinate their behaviour to get sheep (pathogen)
into a pen [291]. Singh and Thayer proposed an architecture called the Immunology-
derived Distributed Autonomous Robotics Architecture for the large-scale coordination
and control of robot teams (many thousands of robots) simulated on a dynamic mo-
bile mine clearing domain [359]. Tasks in the domain were implements as pathogen,
whereas robot action strategies of varying generality were represented as immune cells
that are selected for a given robots current needs. A uni-directional beacon broadcast
inter-robot communication was used to coordinate behaviour inspired by a histamine-
based recruitment mechanism. The approach resulted in emergent coordination that the
authors claimed did not become apparent without large colonies of robots. The framework
was further applied to simulations of an exploration and mapping domain [384, 361], and
a search-and-rescue domain [360].
Lau, et al. proposed a multi-agent control system for fleets of autonomous guided vehi-
cles where vehicles are antibodies, and unassigned tasks they must perform are pathogen
to which they specialise [267, 266, 268]. Michelan and Von Zuben propose a decentralised
21
autonomous navigation system for a robot for garbage collection (locate, collect, obstacle
avoidance, limited resources), that although not distributed, provided an archetype AIS
application in robotics toward complex autonomous behaviour [294]. Ishiguro, et al. also
proposed a decentralised robotic application with behaviour arbitration for a mobile robot
that although not implemented in multiple robots, demonstrated complex autonomous
behaviour [230, 231].
Network Systems
Boudec and Sarafijanovic proposed an AIS for the detection of irregular dynamic routing
behaviour in mobile ad hoc networks [52]. Their negative selection based system iden-
tified misbehaving relays in a simulated network test bed by observing routing events
with detectors prepared off-line. Their approach was extended through the addition of a
virtual thymus used to maintain a dynamic representation of normal and abnormal relay
behaviour, distributed danger signals to govern the flow of information into the virtual
thymus, and memory detectors for improved secondary response [345, 346]. Ganguly, et al.
proposed an ImmuneSearch algorithm for a peer-to-peer network topology that was decen-
tralised using affinity-based proliferation and mutation to control message movement to
avoid flooding the network with messages [160]. Kephart proposes a vision for a scaled up
automated and adaptive virus signature system providing protection across the Internet
[250]. Mohr, et al. proposed exploiting information processing properties of the immune
system such as continual adaptation, distributedness, and context-dependant response for
a context-aware ubiquitous computing platform (also known as pervasive computing and
ambient intelligence) suitable for integration of hand-held devices [296]. Ishida proposed
the immune system from the perspective of the immune network theory provides the foun-
dational inspiration for developing an Autonomous Distributed Systems (ADS), which is a
general class of system of autonomous decision making agents organised in dynamic struc-
ture that cannot be designed beforehand [228]. Ishida highlights examples of adaptive,
autonomous, and decentralised artificial immune systems, although does not provide any
specific information as to how an ADS may be realised from an immunological basis.
Parallel Algorithms
The work by Watkins, et al. addressed the so-called unexploited inherent parallelism of
the CLONALG algorithm [417]. Their effort extended the algorithm such that pattern
recognition tasks are partitioned and executed concurrently (explicit divide and conquer),
and reassembled after execution. The same parallel strategy was applied to the AIRS
algorithm by Timmis and Watkins [420]. Coello Coello, et al. proposed a parallel im-
plementation of their constraint handling immune algorithm as a strategy for assessing
infeasible solutions [79, 78]. Their approach was implemented as an extension to the
standard genetic algorithm implemented in an multiple island population configuration.
22
Other
Hart and Ross explored the analogy between acquired immunity to addressable mem-
ory, specifically as Sparse Distributed Memories (SDM). These structures are a robust
content-addressable memory which are efficient at storing vast amounts of binary data
with a small number of physical addresses [365]. A hybrid method was applied to clus-
tering dynamic datasets that combined SDM, antigen-antibody matching for addressing,
and a co-evolutionary algorithm [199, 200]. Newborough and Stepney proposed a generic
framework for population based algorithms based on a general biologically-inspired con-
ceptual framework [304] (considered further in Section 2.7.1). Their framework provided
abstractions of common optimisation algorithms (clonal and negative selection, genetic al-
gorithm, swarm algorithms such as particle swarm and Ant Colony Optimisation), and a
proof-of-concept implementation on Field Programmable Gate Array demonstrated both
hardware parallelism and distribution. Toma, et al. proposed an Immune Distributed
Competitive Problem Solver that employed aspects of MHC and immune networks in a
multi-agent system to address division of labour problems such as n-Travelling Salesman
Problem [395].
2.5.3 Summary
The clonal selection, negative selection and immune network cellular immunological the-
ories describe decentralised processes capable of general autonomous (per-cell and per-
pathogen) decision making, suitable for distributed and parallel information processing.
The exploitation of these information processing attributes imposed on the biological sys-
tem is the promise of distributed Artificial Immune Systems, and is one of the principle
areas of research in the field of AIS, enshrined in definition. The three schools of AIS
demonstrated that the first-order realisation of Artificial Immune Systems is typically sim-
plistic and monolithic. The work on distributed detector-based computer security driven
by negative selection, and emergent cooperation in autonomous mobile robots are the two
most promising thrusts of research toward realising this goal. The work towards parallelis-
ing existent monolithic immunological algorithms and network applications may be fruitful
areas, although these are still immature and emergent areas of research. Distributed and
decentralised realisations of artificial immune systems represent the overarching ambition
of the field in general. Such information processing characteristics can be directly har-
nessed for classes of function optimization and function approximation problem domains
suited to parallel and cooperative solving strategies. What is lacking in these efforts is a
clear framework to guide the development of a distributed AIS, although Kephart’s vision
for cyberspace anti-virus, Mohr ubiquitous AIS, and Ishida’s vision for an immunological
ADS are thought provoking.
2.6 Open Problems in AIS

This section considers the open problems in the field of artificial immune systems, and the
current trends to address these problems. The review highlights the perceived impasse
23
in AIS and the trends of drawing on novel and more accurate models from the inspiring
biological system.
2.6.1 Open Problems

The authors de Castro and Timmis comment on the lack of and need for a adequate frame-
work for AIS design, interpretation, and application [111]. They propose a general AIS
framework for such purposes that includes three core facets, which are presented formally
in [109] (pages 60-61) as an AIS Engineering Framework, they are: (1) Representation
for the components of the system, (2) Affinity Measures and mechanisms to evaluate the
components of the system in the context of the domain (such as fitness scoring), and (3)
Algorithms and procedures for adaptation that govern the dynamics of the system. They
highlight both the need for more suitable real-world applications of AIS, and the need to
consider the accuracy of the inspiring metaphor, specifically the importance for computer
scientists to grasp the more subtle aspects of immunology [109] (page 310). They outline
four general areas of research as future trends in the field: (1) improvement and augmen-
tation of AIS, (2) new artificial immune systems, (3) application areas, and (2) extending
their general AIS framework.
Freitas and Timmis highlight the increasing trend in the field of applying generic
AIS algorithms to specific and well-defined benchmark domains without the required spe-
cialisation of such algorithms [155]. The authors consider the classification domain and
negative selection algorithms specifically, highlighting the need to consider standard ma-
chine learning principles such as over fitting and inductive bias which are not addressed in
such applications. Andrews and Timmis consider a path to next generation artificial im-
mune systems, commenting that an assessment of the current state of the field reveals the
majority of AIS are almost exclusively inspired by the theories of Jerne’s network theory
and Burnet’s clonal selection theory [24]. They propose to consider other and potentially
controversial immunological theories such as the Cognitive Theory proposed by Cohen,
and advocate the adoption of a guiding conceptual framework.
Timmis, et al. proposed that the first generation of biologically motivated systems are
crude analogies of biological systems, proposing that an inter-disciplinary approach (com-
puter scientists and biologists) is needed to define the second generation of such systems
[389]. They highlight both the importance of understanding the biological system being
exploited, and that high level abstractions of underlying biology are no longer sufficient
requiring a deeper level of understanding. They propose the path to addressing these
concerns is an inter-disciplinary approach. Their argument is made in the context of a
number of computational intelligence paradigms, including AIS. They comment on the
crudeness of AIS algorithms such as CLONALG that “whilst intuitively appealing, lacks
any notion of interaction of B-cells with T-cells, MHC, or cytokines”. Garrett proposes
a framework for assessing usefulness based on the distinctiveness and effectiveness of AIS
and applies the criteria to the three schools of AIS, showing that there is indeed some
useful contributions [164].
Timmis suggests that the field of AIS has reached an impasse [388], specifically with the
24
lack of theoretical advancement, limited application, and the adoption of naı̈ve biological
models. He proposes a number of challenges in the field to address the impasse: (1)
the development of novel and accurate metaphors that provide benefit to immunology,
(2) improved application of AIS models work toward a killer application for AIS, (3)
to develop a theoretical basis for AIS, and (4) to consider the integration of the immune
system with other biological systems. Hart and Timmis assess the field from an application
centric perspective, highlighting distinct lack and narrow focus of real-world applications
[201]. They consider the so called scatter gun approach to applications is caused by
poor foundational biological metaphors, suggesting three areas that may provide a fertile
future for the field: (1) the innate immune system, (2) embodied immune system working
towards host homoeostasis, and (3) life-long learning of the system. Twycross and Aickelin
consider Timmis’ proposed impasse in AIS, highlighting its causes as limited application to
challenging problems, lack of theoretical advances, and use of naı̈ve biological metaphors
[401]. They proposed to address this impasse by considering two addition unexploited
biological metaphors that are suggested as being less complex than those of the human
immune system: plant immunity and invertebrate immune systems.
2.6.2 Summary
The general open problems and trends to address them are common to the field of Compu-
tational Intelligence, specifically the need for a theoretical underpinning for the the field
and suitability and specialisation of algorithms to application domains. Beyond these
general concerns, there is a specific desire for a field-defining (so-called killer ) applica-
tion domains, and the trend to continue to assimilate more exotic immunological theory
and sub-fields towards this end. Further, there is a concerted general intent to devise
and employ general conceptual frameworks to unify existing algorithms, and in the in-
vestigation of novel Artificial Immune Systems. This work proposes to address the open
problem of a perceived impasse in the field of Artificial Immune Systems by following the
specific emerging trends in the field of (1) exploiting novel and more accurate immuno-
logical metaphor, (2) devising and exploiting a general framework that units the facets of
the novel metaphor, and (3) extending the existing AIS frameworks via integration. This
work does not address the open problems of an improved theoretical underpinning of the
field or killer applications, although it is believed that the work will provide insight toward
these questions.
2.7 Methodology
This section specifically address the research goal of identifying a systematic methodol-
ogy for the development and investigation of immunologically inspired algorithms. The
methodology proposed is comprised of the best practices regarding (1) the abstraction
of immunological properties and development of inspired algorithms in Section 2.7.1, (2)
experimental methodology for the qualitative empirical investigation of developed algo-
rithms in Section 2.7.2, and (3) the important considerations regarding the ‘economy of
25
modelling’ of developed algorithms and the need for patch-quilt integration of findings into
broader frameworks in Section 2.7.3.
2.7.1 Conceptual Framework

Conceptual Framework for Bio-Inspired Algorithms
Although a progression from inspiring biological system to inspired computation system

may appear to be an intuitive process, it can involve problems of standardisation of nomen-
clature, effective abstraction and departure from biology, and rigour. Stepney, et al. cau-
tion that by following a process that lacks the detail of modelling, one may fall into the
trap of reasoning by metaphor [400, 372, 373]. Besides the lack of rigour, the trap suggests
that such reasoning and lack of objective analysis limits and biases the suitability and
applicability of resultant algorithms. They propose that many algorithms in the field of
Artificial Immune Systems have succumbed to this trap. This observation resulted in the
development and application of a conceptual framework to provide a general process that
may be applied in the field of Biological Inspired Computation toward realising Biological
Inspired Computational Intelligence systems. The conceptual framework is comprised of
the following actors and steps:
1. Biological System: The driving motivation for the work that possess some innate
information processing qualities.
2. Probes: Observations and experiments that provide a partial or noisy perspective of

the biological system.
3. Models: From the probes abstract and simplified models of the information process-
ing qualities of the system are build and validated.
4. Framework : Build and validate analytical computational frameworks. Validation

may use mathematical analysis, benchmark problems and engineering demonstra-
tion.
5. Algorithms: The framework provides the principles for designing and analysing al-
gorithms that may be general and applicable to domains unrelated to the biological
motivation.
Immunology as Information Processing
Forrest and Hofmeyr summarised their AIS research efforts at the University of New
Mexico and the Santa Fe Institute as “immunology as information processing” [147].
They define information as spatio-temporal patterns that can be abstracted and described
independent of the biological system, and information processing as computation with
these patterns. They proposed that such patterns are encoded in the proteins and other
molecules of the immune system, and that they govern the behaviour of the biological
system. They suggest that their information processing perspective can be contrasted
26
Figure 2.1: Depiction of the conceptual framework for devising Biological Inspired Algo-
rithms, taken from [373]
with the conventional structural perspective of cellular interactions as mechanical devices.

They consider a simple four-step procedure for the investigation of immunology as in-
formation processing, transitioning from the biological system to a usable computational
tool:
1. Identify a specific mechanism that appears to be interesting computationally.
2. Write a computer program that implements or models the mechanism.
3. Study its properties through simulation and mathematical analysis.
4. Demonstrate capabilities either by applying the model to a biological question of

interest or by showing how it can be used profitably in a computer science setting.
The procedure is similar to the outlined in the conceptual framework for Biologically
Inspired Algorithms in that in addition to identifying biological mechanisms (input) and
demonstrating a resultant algorithms (output), the procedure (1) highlights the need for
abstraction involving modelling the identified mechanism, and (2) highlights the need to
analyse the models and abstractions. The procedure of Forrest and Hofmeyr can be used
to specialise Stepney, et al. conceptual framework by clearly specifying the immunological
information processing focus.
2.7.2 Experimental Methodology

Optimization is a generalized class of problem involving improvement of a structure un-
der a cost function. It is a one of a few standard perspectives for the investigation and
application for computational intelligence systems, specifically including clonal selection
algorithms. As such, this section considers the state of optimization algorithm experimen-
tal methodology, the concerns of which are applicable to the investigation of computational
intelligence algorithms in general.
When it comes to evaluating an optimisation algorithm, every researcher has their own
thoughts on the way in which to proceed. Unfortunately, many empirical evaluations of
optimisation algorithms are performed and reported without addressing basic experimental
design considerations. Perhaps before an experimental methodology can be adopted, a
27
researcher or practitioner may be paralysed by the perceived pessimism of the no free lunch
theorem that contends the futility of the benchmarking exercise. A pervasive problem in
the field of Computational Intelligence algorithms is the lack of meaningful and consistent
algorithm experimental and benchmarking methodology. This includes but is not limited
to issues of the selection of problem instances, the selection of algorithm specifications, the
algorithm configuration parameters, and interpretation of results. The intention of this
section is to summarise the general concerns of experimental methodology, with a focus
on benchmarking Computational Intelligence algorithms in the face of the ‘no free lunch’
theorem.
No Free Lunch
Wolpert and Macready’s No Free Lunch Theorem of search and optimisation has caused a
lot of pessimism and misunderstanding, particularly in related to the evaluation and com-
parison of computational intelligence algorithms [436, 435]. In simplest terms, the theory
indicates that when searching for an extremum of a cost function, all algorithms perform
the same when averaged over all possible cost functions. The implication is that the often
perused general-purpose optimisation algorithm is theoretically impossible. The theory
applies to stochastic and deterministic optimisation algorithms, and to algorithms that
learn and adjust their search strategy over time. It is invariant to the performance mea-
sure used as well as the representation selected. Perhaps the catalyst for benchmarking
cynicism is a comment accompanying the proof suggesting that: “. . . comparisons report-
ing the performance of a particular algorithm with a particular parameter setting on a few
sample problems are of limited utility” [436]. The theorem is an important contribution
to computer science, although its implications are theoretical. The original paper was
produced at a time when grandiose generalisations were being made as to algorithm, rep-
resentation, or configuration superiority. The practical impact of the theory is to bound
claims of applicability. Wolpert and Macready encouraged effort be put into devising
practical problem classes and the matching of suitable algorithms to problem classes. Fur-
ther they compelled practitioners to exploit domain knowledge in optimisation algorithm
application, now an axiom in the field.
Issues of Benchmarking Methodology
Empirically comparing the performance of algorithms on problem instances is a staple

for the fields of heuristics and Computational Intelligence, and the problems of effective
comparison methodology have been discussed since the inception of these fields. Johnson
suggested that the coding of an algorithm is the easy part of the process, that the difficult
work is getting meaningful and publishable results [236]. He provides a very through list
of questions to consider before racing algorithms, as well as what he describes as his pet
peeves within the field of empirical algorithm research. Hooker (among others) condemns
what he refers to as competitive testing of heuristic algorithms, calling it fundamentally
anti-intellectual [221]. Hooker continues by strongly encouraging a rigorous methodology
of what he refers to as scientific testing where the aim is to investigate algorithmic be-
28
haviours. Barr, et al. list a number of properties worthy of a heuristic method making
a contribution, which can be paraphrased as; efficiency, efficacy, robustness, complexity,
impact, generalisability and innovation [33]. Barr, Golden et al. specify a loose experi-
mental set-up methodology with the following intuitive steps; 1) define the goals of the
experiment, 2) select measure of performance and factors to explore, 3) design and exe-
cute the experiment, 4) Analyse the data and draw conclusions, and finally 5) report the
experimental results. They then suggest eight guidelines for reporting results, in summary
they are; reproducibility, specify all influential factors (for example code and computing
environment), be precise regarding measures, specify parameters, use statistical experi-
mental design, compare with other methods, reduce variability of results, ensure results
are comprehensive.
Peer, et al. summarise the problems of algorithm benchmarking (with a bias toward
Particle Swarm Optimisation) to the following points; duplication of effort, insufficient
testing, failure to test against state-of-the-art, poor choice of parameters, conflicting re-
sults, and invalid statistical inference [318]. Eiben and Jelasity site four problems with the
state of benchmarking evolutionary algorithms; (1) test instances are chosen ad hoc from
the literature, (2) results are provided without regard to research objectives, (3) scope
of generalised performance is generally too broad, and (4) results are hard to reproduce
[132]. The general problems with benchmarking methodology may be distilled into the
following sub-problems:
• Parameter Selection: Computational Intelligence algorithms are typically parame-

terised, although the mapping of parameter values to predictable effects on problem
domains is generally poorly understood. This is because typically unknown and non-
linear dependencies commonly exist between the variables resulting in the algorithm
being consider a complex system. Francois and Lavergne discuss the deficiencies of
the traditional trial-and-error and the experienced-practitioner approaches to pa-
rameter tuning, further suggesting that seeking general rules for parameterisation
will lead to optimisation algorithms that offer neither convergent or efficient be-
haviours [153]. There are many solutions not limited to the following: consulting
the literature although often ignored [132]), generalise from large numbers of ex-
periments [347], self-adaptive parameters, meta-algorithms for searching for good
parameters values, and statistical sensitivity analysis over parameter ranges [74, 45].
• Problem Selection: Problem instances should be selected to demonstrate specific

behaviours of the systems under study. Typically benchmarks are performed on
ad hoc lists of standard benchmark instances, that although offering the benefit
of comparability, typically devolve into racing contest. Eiben and Jelasity support
the division of problem instances into categories and encourage the evaluation of
optimisation algorithm over a large number of test instances [132]. In their paper on
understanding the interactions of Genetic Algorithm parameters Deb and Agrawal
propose four structural properties of problems for testing genetic algorithms; multi-
modality, deception, isolation, and collateral noise [125]. Yao, et al. divide their
large test dataset into the categories of uni-modal, multi-modal with many local
29
optima, and multi-modal with few local optima [441]. Whitley, et al. provide a
detailed study on the problems of selecting test instances for genetic algorithms, and
suggest that difficult problem instances should be non-linear, non-separable, and
non-symmetric [431].
• Measure Selection: There are many ways to measure the performance of an al-
gorithm for a problem instance, although the most common involves measures of
efficacy (problem specific costs) and efficiency (computation space and/or time).
Most Biologically Inspired Algorithms have a stochastic element, typically in their
random starting position(s) and in the probabilistic decisions made during sampling
of the domain. Therefore, the measuring of performance must be repeated a num-
ber of times to account for the stochastic variance2 , which also could be a measure
of comparison between algorithms. The most critical concern in the selection of
measures is that of standardisation for the purposes of comparison [44, 33].
• Statistical Significance: The scientific knowledge of Computational Intelligence is

predominantly accrued as empirical observations and conclusions. Therefore, there
is a need for methods that permit interpretation and conclusions to be drawn be-
yond the standard reporting of results. Statistical hypothesis testing and related
methods provides such tools, although unfortunately are rarely employed. Peer,
et al. [318] and Birattari and Dorigo [43] provide a basic introduction (suitable for
an algorithm-practitioner) into the appropriateness of various statistical tests for al-
gorithm comparisons. Parametric statistical methods are used for interval and ratio
data (like a real-valued performance measure), and non-parametric methods are used
for ordinal, categorical and rank-based data. Interval data is typically converted to
ordinal data when salient constraints of desired parametric tests (such as assumed
normality of distribution) are broken such that the less powerful non-parametric tests
can be used. The use of non-parametric statistical tests maybe preferred as some
authors (for example [75]) claim the distribution of cost values are very asymmetric
and/or not Normal.
2.7.3 Designing and Investigating Adaptive Systems

Complex and adaptive systems are difficult to design and investigate, therefore the need
for a coherent and proven methodology for such activities is paramount. Inspired by an
interpretation of the achievements of the Wright brothers in achieving the first powered
flight, this section provides a summary of a methodology for investigating and designing
conceptual machines proposed by David Goldberg [177].
For the purposes of clarity, a model is a conceptualisation that provides a description
of a system that accounts for its properties and may be used to study the characteristics
of the system. For the remainder of this thesis, conceptual models are instantiated as
algorithms for the purposes of computational implementation, simulation, and study. The
2
Typically ≥ 30 according to the central limit theorem such that the underlying distribution can be
meaningfully summarised.
30
algorithms provide procedures for problem-solving when the behaviour of the instantiated
model suitably match the characteristics of a problem.
Engineers and Mathematicians
Goldberg describes the airplane and other products of engineering as material machines,
and distinguishes them from the engineering of genetic algorithms and other adaptive sys-
tems as conceptual machines. He argues the methodological distinction between the two is
counter-productive and harmful from the perspective of conceptual machines, specifically
that the methodology of the material is equally applicable to that of the conceptual [174].
The obsession of mathematical rigour in computer science, although extremely valuable,
is not effective in the investigation of adaptive systems given their complexity. Goldberg
sights the airplane as an example where the engineering invention is used and trusted
without a formal proof that the invention works (that an airplane can fly3 ). This de-
fence leads to what Goldberg refers to the economy of design which is demonstrated with
a trade-off that distinguishes ‘model description’ (mathematician-scientists) that is con-
cerned with model fidelity, and model prescription (engineer-inventor) that is concerned
with a working product. In descriptive modelling the model is the thing (of interest)
whereas in ‘prescriptive modelling’, the object is the thing (of interest). In the latter, the
model (and thus its utility) serves the object, in the former model accuracy may be of
primary concern. This economy of modelling provides a perspective that distinguishes the
needs of the prescriptive and descriptive fields of investigation.
The mathematician-scientist is interested in increasing model accuracy at the expense
of the speed (slow), where as the engineer may require a marginally predictive (inaccurate)
model relatively quickly. This trade-off between high-cost high-accuracy models and low-
cost low-fidelity models is what may be referred to as the modelling spectrum that assists
in selecting an appropriate level of modelling. Goldberg proposes that the field of genetic
algorithms expend too much effort at either ends of this spectrum. There is much work
where there is an obsession with blind-prototyping many different tweaks in the hope of
striking it lucky with the right mechanism, operator, or parameter. Alternatively, there
is also an obsession with detailed mathematical models such as full-blown differential
equations and Markov chains. The middle ground of the spectrum, what Goldberg refers
to as little models is a valuable economic modelling consideration for the investigation of
conceptual machines to do good science through good engineering.
Methodology
The methodology has been referred to as post-modern systems engineering and is referred
to by Goldberg as a methodology of innovation [175]. The core principles of the process
are as follows:
1. Decomposition: Decompose the large problem approximately and intuitively, break-

ing into quasi-separate sub-problems.
3
Goldberg is quick to point out that sets of equations do exist for various aspects of flight, although no
integrated mathematical proof for airplane flight exists.
31
2. Modelling: Investigate each sub problem separately (or as separate as possible) using
empirical testing coupled with adequately predictive, low-cost models.
3. Integration: Assemble the sub-solutions and test the overall invention, paying atten-
tion to unforeseen interactions between the sub-problems.
Decomposition Problem decomposition and decomposition design is an axiom of re-

ductionism and is at the very heart of problem solving in computer science. Therefore,
it is not worth dwelling on the topic other than to comment as to its meaning within
the context of adaptive systems. One may consider the base or medium on which the
system is performing its computation mechanisms, the so-called building blocks of infor-
mation processing. A structural decomposition may involve the architecture and data
structures of the system. Additionally, one may also consider a functional breakdown of
mechanisms such as the operators applied at each discrete step of an algorithmic process
or mechanisms. The reductions achieved provide the basis of investigation and modelling.
Small Models Given the principle of the economy of modelling presented as a spectrum,
one may extend the description of each of the five presented model types. Small Models
refers to the middle of the spectrum, specifically to the application of dimensional and
facet-wise models. These are mid-range quantitative models that make accurate prediction
over a limited range of states at moderate cost. Once derived, this class of models generally
requires a small amount of formal manipulation and large amounts of data for calibration
and verification. The following summarises the modelling spectrum:
• Unarticulated Wisdom: (low-cost, high-error) Intuition, what is used when there is

nothing else.
• Articulated Qualitative Models: Descriptions of mechanisms, graphical representa-

tions of processes and/or relationships, empirical observation or statistical data col-
lection and analysis.
• Dimensional Models: Investigate dimensionless parameters of the system.
• Facet-wise Models: Investigation of a decomposition element of a model in relative

isolation.
• Equations of Motion: (high-cost, low-error) Differential equations and Markov chains.
Facet-wise models are an exercise in simple mathematics that may be used to inves-
tigate a decomposition element of a model in relative isolation. They are based on the
idea of bracketing high-order phenomena by simplifying or making assumptions about the
state of the system. An example used by Goldberg from fluid mechanics is a series of
equations that simplify the model by assuming that a fluid or gas has no viscosity, which
matches no known substance. A common criticism of this modelling approach is “sys-
tem X doesn’t work like that, the model is unrealistic”. The source of such concerns with
adaptive systems is that their interactions are typically high-dimensional and non-linear.
32
Goldberg’s response is that for a given poorly understood area of research, any useful
model is better than no model. Dimensional analysis or the so-called dimensional reason-
ing and scaling laws are another common conceptual tool in engineering and the sciences.
Such models may be used to investigate dimensionless parameters of the system, which
may be considered the formalisation of the systemic behaviours.
Integration Integration is a unification process of combining the findings of various

models together to form a patch-quilt coherent theory of the system. Integration obviously
is not limited to holistic unification, one may address specific hypothesis regarding the
system resulting in conclusions about existing systems, and design decisions pertaining to
the next generation of systems.
Application In addition to elucidating the methodology, Goldberg specifies a series of

five useful heuristics for the application of the methodology as follows (taken from [174],
page 8):
1. Keep the goal of a working conceptual machine in mind. Experiments commonly

get side tracked by experimental design and statistical verification; theoreticians get
side tracked with notions of mathematical rigour and model fidelity.
2. Decompose the design ruthlessly. One cannot address the analytical analysis of a
system like a genetic algorithm in one big ‘gulp’.
3. Use facet-wise models with almost reckless abandon. One should build easy models
that can be solved by bracketing everything that gets in the way.
4. Integrate facet-wise models using dimensional arguments. One can combine many
small models together in a patch-quilt manner and defend the results of such models
using dimensional analysis.
5. Build high-order models when small models become inadequate. Add complexity to
models as complexity is needed (economy of modelling).
2.8 Chapter Summary

This Chapter positioned the investigation broadly in the field of Biologically Inspired Com-
putational Intelligence, a sub-field of Artificial Intelligence, and specifically in the field of
Artificial Immune Systems. Clonal Selection was highlighted as an adaptive paradigm in
the field of Artificial Immune Systems that underlies much of the general field. Distributed
Artificial Immune Systems including distributed Clonal Selection Algorithms is an over
promised and under delivered, although potentially beneficial field of study. An effective
realisation of distributed clonal selection algorithms is expected to facilitate the broader
application of such approaches to classes of problem that may benefit from parallel and
cooperative problem solving strategies, such as problems with functional or information
availability decompositions. Turning back to the structure and function of the immune
33
system to develop new AIS approaches including distributed approaches is currently an
advocated approach in the field given the apparent impasse experienced in the progress
of such approaches. A systematic framework was adopted for the development of new
immunologically inspired approaches, with careful empirical investigation, and awareness
of the economy of models and need for the re-integration of findings. This chapter both
positioned and motivated the research hypothesis outlined in Chapter 1, and defined the
systematic methodology adopted for the development and investigation of adaptive and
distributed clonal selection algorithms throughout the remainder of the dissertation. To-
ward this end, Chapter 3 provides an in depth investigation into the clonal selection
paradigm, both reviewing and elaborating the cellular perspective of such approaches
which is later pursued in Chapter 4, and outlining the agenda for realising distributed
approaches in Chapters 5 and 6.
34
Chapter 3
Clonal Selection Paradigm

This chapter provides an in depth review of clonal selection theory and inspired computa-
tional algorithms, highlighting the open problems in the state of the field that specifically
motivate the work. Section 3.2 reviews the clonal selection theory and relevant immuno-
chemistry and theoretical immunology providing a detailed foundation to effectively inter-
pret the motivations for the state of clonal selection algorithm research. The field of clonal
selection algorithms is reviewed in Section 3.3 providing a compressive taxonomy of ap-
proaches and focusing on the abstracted immunological principles and application domains
for resultant systems. The review identifies three limitations that motivate the remainder
of the chapter, specifically (1) the lack of placement of clonal selection approaches in a
broader context, (2) the lack of explicit investigation of clonal selection as an adaptive sys-
tem, and (3) the cellular focus of abstracted computational principles. These general open
problems are investigated in turn. Section 3.4 reinforces the relationship of the paradigm
with evolutionary algorithms and instance based learning, and highlights the relationships
with binary hill climbing and competitive learning paradigms. Section 3.5 investigates
clonal selection in the context of adaptive systems theory, highlighting the need to focus
on the system-environment relationship of the paradigm, and the general short-sightedness
of selection-based adaptive systems. Finally, Section 3.6 outlines an agenda for investi-
gating the clonal selection paradigm for the remainder of the dissertation, firstly at the
classical adaptive cellular level, as well as the inherently distributed ‘host of tissues’ and
‘population of hosts’ perspectives of immunology.
3.2 Clonal Selection Theory

3.2.1 Theory of Acquired Immunity
This section provides an overview of the theory including its inception, main processes,
and the mechanisms employed for diversity. This discussion of clonal selection is limited to
the discussion of B-lymphocytes and humoral immunity given that the development and
much of the experimental and theoretical work on clonal selection has been completed in
35
this context.
Theory Development
Paul Ehrlich observed the exponential generation of antibodies triggered by the primary
exposure to antigen in blood. He proposed a side-chain theory (circa 1900) in which
blood cells naturally made side chains (now known as receptors) capable of binding to all
antigens. His theory proposed that the information and capability to detect all antigen was
already possessed by the host (for example in the genome) and expressed by the antibodies.
This model was ultimately rejected because it was demonstrated that antibodies could be
produced in response to synthetic substances of which the antibodies could have no prior
knowledge. Further, the theory could not explain the improvement in the specificity of the
response in the secondary and subsequent exposures to the same antigen. The antigen-
template theory was proposed and suggested that antigen direct the response [315]. In the
theory antibodies use the antigen as a template or model and modify themselves to provide
a more effective complement. The emerging field of genetics suggested the invalidity of this
theory as the specificity of an antibody could not be changed after its creation given that
the process of gene to amino acids to protein conformation is one-way. Niles Jerne had
great difficulty with the template theory and listing a number of observations to which the
theory could not account [232]. The template theory proposed a one-to-one relationship
between antibodies and antigen, although experimental observations suggested that far
more antibodies were produced in the primary response than there were antigen. Further,
the template theory suggested the role of antibodies ended after neutralising the antigen,
whereas observations indicated that antibodies have a short lifespan, therefore failing to
account for observed immunological memory exhibited in the improved specificity of the
secondary response.
Jerne proposed an alternative called the natural selection hypotheses of antibody diver-
sity. His theory proposed that the host contains a small pre-existing collection of randomly
generated antibodies. In generating this pre-existing pool, self-antibodies where proposed
to be somehow neutralised before causing harm to the host. The theory proposed that
antigen naturally select antibodies of high affinity (a good match) from which more anti-
bodies of the same affinity are produced. The result of the antigen-to-antibody selection
mechanism the production of antibodies specific to the antigen in large numbers, which
presented an improved response of the host to the antigen. Finally, he suggested that
copying errors during the triggered production of antibodies may result in an improved fit
with the antigen and account for the observed immunological memory effect. In his work
on antibodies and allergy, David Talmage extended upon the natural selection hypothesis
suggesting that Jerne’s theory was an extension of Ehrlich’s side-chain theory, and that
replicating cells, not antibodies, were the main feature of the immune response [380]. Tal-
mage suggested the notion that immune cells must specialise in producing antibodies with
the same specificity as the receptors of the surface of the cell that trigger its proliferation.
Frank Macfarlane Burnet also extended Jerne’s theory at the same time as Talmage
(and subsequently received priority) focusing on a clone perspective and the self-replicating
36
cellular mechanisms during the immune response [63]. He called his theory the clonal se-
lection hypothesis of antibody diversity. The ‘missing link’ for Burnet was the pre-existing
naı̈ve antibody population that was proposed in Jerne’s theory. As with Talmage’s the-
ory, antigen select the best matching cells, that results in their subsequent replication. He
suggested that for the selectionist theory to hold, that each cell must produce antibody of
one type (all with the same receptor configuration). His theory attributed the exponential
rise in the number of antibodies after infection to the exponential rise in the number of
antibody producing cells due to clonal expansion of selected cells. It also explained the
improved secondary response given the specialisation of receptors during expansion, al-
though the mechanisms by which receptors specialised was only speculated. The following
provides a summary of the characteristics of Burnet’s clonal selection, from [65]:
1. Initial Repertoire: Randomly generated initial population of antibody producing

immune cells.
2. One-to-one: An immune cell can detect and release antibody of one specificity, which
is passed on to descendent cells.
3. Elimination: Those initial cells that are reactive to self-tissues are selected against
(killed).
4. Proliferation: Generation of antibody and the proliferators of the cell after contact
(selection) with antigen.
5. Forbidden: The elimination of progeny that can select for self tissues.
Talmage and Burnet both followed up two years later with monographs elaborating
on the theory, providing a more detailed account of the clonal selection process [381, 64].
Burnet continued his work on the study of immunological tolerance, to which his theory
supported, and with Peter Medawar won the Nobel Prize in 1960 for work that provided
the foundation for viable tissue and organ transplant. It is also interesting to note that
it was Burnet who conceptualised the self-nonself abstraction of the immune system in
his work on immunological tolerance, which like his clonal selection theory has remained
a foundational concept of modern immunology1 . Nossal and Lederberg supported the
one-cell-one-antibody assertion of clonal selection providing first test and confirmatory
evidence for the theory [306]. As mentioned, Burnet speculated as to the mechanisms for
antibody diversification and specialisation in his theory. His prediction of the existence
of such a mechanism was confirmed approximately 20 years later by Tonegawa et al. who
investigated the fine tuning of antibody receptors called somatic hypermutation for which
he later won the Nobel Prize [223, 397].
Theory Detail
An antigen is a molecule which can elicit an immune response, an antibody (immunoglob-

ulin) is a molecule which can bind to and neutralise an antigen, and a B-lymphocyte is an
1
For an overview of the history of self-nonself paradigm see Crist [86]
37
immune cell which can bind to antigen as well as produce antibodies. The physical surface
of an antigen has a number of different structural features called antigenic determinants.
An antibody and B-cell have receptors which have a specificity (ability to bind) to only
one surface feature of an antigen, although the feature an antibody binds to may be rep-
resented on more than one different antigen. A host possesses a base pool of randomly
generated lymphocytes that define the hosts natural immunity, and are prepared during
the embryonic stage of development. During this time, the immune system learns a tol-
erance for the tissues of the host organism and those lymphocytes that are self-reactive
are removed from the pool. The repertoire of natural lymphocytes possesses the ability
to bind to any antigenic stimulus (natural or synthetic) with some general affinity. The
introduction of an antigen triggers an immune response. The lymphocytes that are most
suited (have the best complementary shape) for the antigen bind to it triggering cell di-
vision and differentiation. Therefore, it is the features of the antigen that select the best
matching lymphocytes from the hosts cellular repertoire. After a lymphocyte binds to an
antigen it moves to lymphatic tissue (germinal centre) to begin the differentiation process.
The function of this process is the preferential proliferation to replicate a clone of cells to
produce antibody capable of neutralising the triggering antigen. The lymphocyte prolifer-
ates and differentiates into two different cell types: plasma cells that continue to replicate
and which release large numbers of antibody, and memory cells that like the parent cell
act as a catalyst for response to subsequent infections.
During the proliferation and differentiation process copying errors (somatic mutations)
may occur with a probability of mutation much higher than the genetic mutations that
occur in typical cell divisions. The result is a clone of lymphocytes that have receptors
similar to the parent cell, although vary slightly in their specificity to the triggering antigen.
This means that on subsequent exposures to the antigen, the host possesses a larger
repertoire of lymphocytes and antibodies capable of neutralising the antigen. Given the
somatic mutations introduced during replication, some of the lymphocytes will contain an
improved match for the antigen, improving their chance of being selected. An antibody
producing lymphocyte will only produce antibodies of the same affinity. This means
that after the cell has formed (after mitosis) it may be considered an antibody factory
where all antibodies produced have the exact same complementary shape for a specific
antigen surface feature. The fact that a surface feature may be represented by more than
one antigen provides the basis for vaccinations. This commonality of shape facilitates
the capability for generalisation, where lymphocytes and antibody may be raised on one
harmless antigen and be effective on other more harmful antigens that possess the same
structures. The memory cells generated during the differentiation process may live for
months and years which is much longer than typical lymphocytes that lives for hours or
days, providing long term memory of the learned antigenic defence.
The problem of antibody diversity was considered for at least 100 years before a co-
herent theory was proposed, although it is important to highlight that the theory in its
original inception was not perfect. Specific details, such as the mechanisms of diversity
where not known for many years after the theory’s acceptance. Silverstein suggested that
38
with regard to the clonal selection theory, like Darwin’s theory of evolution by natural
selection, much of the specifics have been out-dated given scientific progress [358]. Rather
than abandon the core principles, Silverstein emphasises the need to embrace the contin-
ually updated version of the theory. A pioneer himself in immunology, Lederberg suggests
similar sentiments: “Don’t let conflicting and awkward ‘facts’ stand in the way of an es-
thetically satisfying theory whose fundamentals are consistent with the world model and
with one another! And be suspicious of ‘facts’ that seem in the way of any coherent theory”
[269] (page 181).
Finally, the clonal selection process of acquired immunity is not without problems.
An observed effect known as Original Antigenic Sin 2 refers to the commitment of the
immune response and subsequent memory to specific surface characteristics of a pathogen,
such as a virus, then having the virus mutate and change those surface characteristics
[152, 122]. The result of the commitment is that the previously high-affinity lymphocytes
and antibody become low-affinity within the context of the change to the virus. Given the
clonal selection and expansion of lymphocytes in the first exposure, the now low affinity
receptors dominate the subsequent exposures, effectively out-competing other lymphocytes
that may be better suited to the changed virus.
Diversity Mechanisms
The genetics and chemistry of antibody diversity is beyond the scope of this work. This
section provides a summary of the main mechanisms that lead to the diversity of the B-
lymphocyte repertoire, which includes gene recombinations, somatic mutations, receptor
editing, class switching, and gene conversion.
• Genetic Recombination: The process that accounts for the initial and on-going high
diversity of new lymphocytes is called secondary V(D)J rearrangements, which refers
to the recombination of the genetic code during the development of B-lymphocytes.
This genetic recombination process accounts for the diversity of the initial lympho-
cyte repertoire and in the creation of new (known as naı̈ve) B-lymphocytes [396, 397].
• Somatic Hypermutation: Along with genetic recombination was the first discovery
made regarding the genetic component of antibody diversity [396]. It refers to the
point mutations of the genome that occur during the clonal expansion phase when the
cell moves to the germinal centre for differentiation (for more information regarding
cell maturation see [37, 38]). The mutations are targeted at the genes in the genome
that code for the receptors and occur at a rate about one million times higher than
the mutation rates in other genes. Mutations include substitutions, additions, and
deletions of genetic code. Specific sequences (motifs) within the receptor genes are
targeted with high frequency, referred to as ‘hot spots’.
• Receptor Editing: Clonal selection proposes that those cells with receptors that bind
to self-tissues (auto-reactive) are clonally deleted before they can cause damage.
2
“Original Sin” is a term from Christian theology describing the condition of sin that marks all humans
as a result of Adam’s first act of disobedience.
39
Receptor editing is an additional process for re-programming auto-reactive receptors
by recombining their genes [332, 307]. Those cells that have receptors that are still
auto-reactive after this editing process are eliminated from the repertoire. Receptor
editing was proposed to occur before somatic mutation and in response to contact
with antigen (similar to template theory). It has also been proposed that receptor
editing may provide an additional general receptor diversity mechanism for so-called
‘long jumps’ (large changes) in receptor affinity [167].
• Class Switch Recombination: The genetic recombination that results in antibod-

ies changing their behaviour by becoming different immunoglobulin types, although
maintaining the same specificity for antigen [198, 273]. Gene conversion is another
genetic recombination process and the primary mechanism for diversification of im-
mune receptors in chickens (and some other species) that is employed instead of
somatic mutation used in mammals [438, 273].
3.2.2 Shape Space and Affinity Landscape

The shape-space formalism and its sibling the affinity landscape are two common geometric
paradigms from theoretical immunology, both of which have had a recent revival and
reapplication in the related field of Artificial Immune Systems. This section briefly surveys
the history of these terms in the context of computational immunology, and consider their
pervasive application in AIS.
Shape-Space
Perelson and Oster introduced the shape space formalism in their theoretical investigation
of antibody repertoire size and reliable recognition [321]. Given antibody ab and anti-
gen ag, the authors discretised the physical aspects of the antibodies combining site and
the antigenic determinant into a vector of n so called shape parameters. The parame-
ters included the set of intermolecular forces important to the ab − ag recognition. The
abstraction of a molecule was later referred to as its generalised shape, where the shape pa-
rameters for a ag and ab molecule may be defined as points in an n-dimensional Euclidean
space called shape-space S [320]. The authors suggested that given an adequate method
of defining molecule shape, the dimensionality of such a space could be relatively small (as
low as five dimensions). Further, the formalism simplified the complementarity nature of
ab − ag interactions, thus rather than the matching convention of ab #= ag, the formalism
considers ab = ag. Given this simplification, the formalism defines distances between ab
and ag in the Euclidean shape space as the degree of complementarity or affinity of the
interaction. A generalised affinity function may be defined as cij = f (xi , xj ), where cij
is the affinity between the molecules i and j, and xi and xj are vectors representing the
molecules in shape space, f is an appropriate function for the selected representation. The
shape space is defined as a finite hypercube, with a volume V of uniform density. Each
antibody was defined with a local recognition region in the shape space !, with a Gaussian
falloff. An ab-ag interaction only occurs if the ag is above an affinity threshold, within the
40
triggering hyper-sphere of the ab. In their application of the model, random distributions
of ab and ag were used and self-antigens were ignored.
Figure 3.1: Depiction of the shape-space formalism taken from [322] (page 1225). Xi
represents antibodies with corresponding complementary recognition volumes V!i .
The shape-space formalism provides a geometric way to consider molecule interactions

in the context of the immune system, it was not the first of such a geometric abstraction
of molecular interaction. For example, Write considered a genotypic-space for possible
gene combinations and a mapping onto the now ubiquitous fitness landscape in theoretical
genetics [437]. This fundamental ag − ab interaction paradigm was used throughout the
late 1980’s and 1990’s predominantly in theoretical works modelling clonal selection and
various aspects of Jerne’s idiotypic network theory (for example work on a one-dimensional
Euclidean shape-space by Segel [355], and an investigating the stability of the idiotypic
model by Perelson [320]). Carneiro and Stewart criticised the shape-space formalism, fo-
cusing on the simplicity of the abstracted space and the shortcomings of the simple affinity
functions [70]. They cautioned against the extrapolation of results from the abstraction
to the real shape-space, highlighting that the dimensionality of the space should be much
higher (at least 10, and as high as 20), and that the affinity mapping function would have
to be irregular and discontinuous. They support their claims with observations from simu-
lated docking of molecules based on crystallographic structures. They propose a splitting
of the formalism into a realistic shape-space in which real molecules can be evaluated, and
an abstracted inversion space as a tool for modelling.
Affinity Landscape
It is typical in physics for systems to minimise some utility in the context of a response
surface, whereas in biology it is common for systems to maximise some utility. This is
highlighted by the already mentioned evolutionary fitness landscape paradigm introduced
by Wright in conceptualising gene combinations [437]. In the immune system, one may
consider the utility of an antibody as its affinity in binding to a specific antigen. A related
concern called avidity, refers to strength of the binding between antigen and antibody.
41
Among the first to employ the conceptualisation of an affinity landscape were Kauffman
and Weinberger who used this geometric paradigm in the field of theoretical immunol-
ogy to investigate and describe the theoretical effects of affinity maturation [242, 243].
They investigated adaptive walks using Kauffman’s NK3 model on the affinity landscape
based on antibody sequences or so called antibody-space in response to a given antigen.
Kauffman and others used the NK model to investigate various molecular sequence spaces
and the corresponding response surfaces, for example [137, 240, 241]. One may define an
affinity landscape as a degree of complementarity response surface for a given collection of
antigen detecting agents (lymphocytes and antibody), for a given specific antigenic deter-
minant. Abstractions of this surface considered in theoretical immunological are typically
continuous, the topology of which consists of many local optima. Detecting agents may
navigate this response surface through the process of affinity maturation, which involves
manipulations to genotypic sequence information (such as the clonal selection diversity
mechanisms discussed in Section 3.2.1).
The affinity landscape is a simple yet important conception, and although Kauffman
et al. were perhaps the first to apply such a name, the geometric formalism was no doubt
in the mind of Perelson and others in the early days of theoretical and computational
immunology. The formalism has played an important role in conceptualising and for-
malising affinity maturation of the immune response by hypermutation (for example see
[325, 356, 385, 386]), which as has been demonstrated is a critical aspect of the clonal selec-
tion theory of antibody diversity. George and Gray used the affinity landscape as a crutch
in describing their theories on receptor editing, which were generalised very effectively on
a simple two-dimensional affinity graph in Figure 3.2 [167, 168].
Figure 3.2: Theory of receptor editing providing long jumps in the affinity landscape, from
[167].
Artificial Immune Systems
The shape-space and affinity landscape paradigms have been widely used in the field of
Artificial Immune Systems. Some examples include the quintessential binary shape-space
used in negative selection (for example [375]) and immune network algorithms, the recogni-
3
Not to be confused with Natural Killer cells, N and K are parameters of the Kauffman’s model.
42
tion region and threshold used in data mining and classification algorithms [422, 421, 259],
and the affinity landscape cost functions navigated by optimisation algorithms [427, 117].
The usage of these paradigms is unified in the framework for AIS of Engineering de Cas-
tro and Timmis, and listed in Section 2.6.1. The authors presented abstract models of
immune cells, molecules and their interactions including the affinity landscape and shape-
space paradigms in an effort to elaborate the three facets of their framework (represen-
tation, affinity measures, and immune algorithms). One may distil the discussion of the
AIS Engineering Framework and related abstraction by de Castro and Timmis, to a triad
relationship between representation in a shape-space, a mapping function between points
in that space, and an affinity landscape of scores assigned by the mapping function, as
follows:
• Representation: Extending the shape-space paradigm, the sequence, or search space

is considered essentially a representation domain. Examples provided include real-
valued, binary, integer and symbolic.
• Mapping: A function that translates from the representation shape-space to an

affinity landscape surface. Distance and matching within the shape-space are the
mapping functions discussed with a number of appropriate suggestions for the various
shape-spaces listed. A point in one of these spaces may have a recognition region,
and a cross-reactive threshold. A mapping function may be complemented with an
additional threshold binding function, perhaps to convert a scalar affinity into a
decision variable.
• Affinity: Affinity may be further abstracted from the degree of complementarity in

an antibody-antigen interaction, to a more general scoring assigned to the mapping
between points in the shape-space. Affinity may be a measure of quality of an agent
in a specific environment or in the context of a specific requirement.
The proposed triad provides a geometric interpretation and relation of shape-space to

affinity landscape paradigms and provides an open framework for investigating immuno-
logical principles in domains, not limited to the pattern recognition problem. One may
consider further geometric conceptions related to affinity maturation in immune system
that lack a suitable formalism, let alone an abstraction for use in Artificial Immune Sys-
tems. Two examples may include: (1) the aggregation of multiple affinity landscapes, that
is the affinity of a repertoire in the context of an antigenic environment not limited to one
antigen, and (2) the temporal and spatial aspects to affinity throughout the host organism
in the context of lymphocyte and antibody mobility.
3.3 Clonal Selection Algorithms

The previous section outlined in detail the abstraction, immunochemistry, and theoretical
immunology that inspired the field of clonal selection algorithms. The review highlighted
that the main actors in clonal selection are cells, antibodies, and antigen, and the main
43
processes are selection, proliferation, and genetic mutations. Further, it showed that the
common measures used to describe the relationship between antibody and antigen are
affinity and avidity, and that the main conceptual formalisms used to describe the con-
straints and effects of such interactions are shape-space and the affinity landscape. This
section reviews the so called clonal selection principle which is the abstraction that moti-
vates the field of clonal selection algorithms. A taxonomy of such algorithms is presented,
highlighting the open problems in the state of the art.
3.3.1 From Theory to Principle

In the glossary of their seminal book on Artificial Immune Systems, de Castro and Timmis
provided a definition of the clonal ‘selection principle’ as “the prevalent theory stating
that the specificity and diversity of an immune response are the result of selection by
antigen of specifically reactive clones from a large repertoire of preformed lymphocytes,
each with individual specificities”, not distinguishing between the theory and information
processing abstraction [109] (page 322). In their treatment of Burnet’s theory, the authors
comment that lymphocytes may be considered to undergo a process similar to natural
selection and that only those cells that contact an antigen may be selected for proliferation,
strongly highlighting the Darwinian connection made by Jerne and Burnet. The authors
focused on the clonal expansion and variation aspects of the process, referring to the
abstraction of the theory as affinity maturation of the immune response. They
propose two important computational features of the theory in their abstraction (page
80): (1) An antigen selects multiple cells to proliferate where each cell has an individual
clonal expansion rate proportional to its affinity to the antigen (higher affinity equals more
clones), and (2) the mutation of each cell during clonal expansion is inversely proportional
to the affinity of cell to the antigen (higher affinity equals less point mutations). The
affinity proportional rates imposed on clonal expansion and mutation in their general clonal
selection abstraction may be traced back to an earlier work of de Castro and Von Zuben
on the seminal clonal selection algorithm CLONALG [114] (discussed in Section 3.3.2), in
which they cited the work of Berek and Ziegner [38], commenting that a controlled cloning
and mutation process could improve the efficiency of the process, and may be employed
by the immune system.
In their treatment of the clonal selection and inspired algorithms Cutello and Nicosia
proposed that clonal selection allows “the learning of patterns during the primary response,
and retrieval of previous knowledge during the secondary response or the cross-reactivity
process” [97] (page 112). They defined the abstraction of the theory called the clonal
selection principle as involving both a learning or training phase followed by an applied
or testing phase that may continue for the lifetime of the system, an approach which is
common to the field of machine learning and pattern recognition. They proposed two key
features of the clonal selection theory be taken into account in realising such a principle:
(1) the hypermutation mechanism which they consider a local search procedure that leads
to a fast maturation during the learning phase, and (2) clonal expansion mechanism that
triggers the growth of a new population of high-value (affinity) cells. They consider all
44
algorithms based on the clonal selection principle to be population based, where each
member of the population represents a candidate solution belonging to the combinatorial
fitness landscape of a given computational problem.
What is distinctly common between these two sets of principles is (1) the absence of
concerns of cell and antibody preparation for the task of pathogen identification associ-
ated with the negative selection paradigm, and (2) the absence of the concerns of cell-cell
(inter-repertoire) interactions such as those associated with the immune network paradigm.
Neither sets of principles explicitly exclude these concerns, rather they are not the focus of
the information processing. The Cutello-Nicosia and the de Castro-Timmis abstractions of
the clonal selection theory are the two main sets of principles inspiring the IA and CLON-
ALG families of clonal selection algorithms respectively (Section 3.3.2). Both abstractions
consider clonal selection to be an adaptive process that operates on a finite set of discrete
units, the composition of which is changed though the iterative application of selection
and clonal expansion with variation. Both abstractions also value the diversity provided
through bind mutation that results in pre-committed configurations, some of which offer
a relative selective advantage. The Cutello-Nicosia abstraction considers distinct training
(change) and test (fixed) phases of adaptation, whereas the de Castro-Timmis abstraction
considers affinity-biased controls on a single adaptive phase.
3.3.2 Algorithm Taxonomy

This section presents a taxonomy of clonal selection algorithms that divides the scope of
such algorithms into a genealogical tree of five groups, with a sixth miscellaneous group4 .
The presented taxonomy only considers those algorithms that explicitly exploit (or claim
to exploit) the clonal selection principle, which as outlined above requires the integration
of the abstract actors and processes of the the clonal selection theory. Also listed in the
miscellaneous category are examples of those AIS that although do not fit the adopted
definition of a clonal selection algorithm, exploit considerable aspects of the clonal selection
principle. Each linage is reviewed with regard to the core principles, exemplar algorithms,
and summary of general application. Table 3.1 provides a summary of each lineage and
classified algorithms.
CLONALG
Hidden at the back of a technical report surveying the applications of Artificial Immune
Systems, de Castro and Von Zuben proposed the Clonal Selection Algorithm (CSA) as a
computational realisation of the clonal selection principle for pattern matching and opti-
misation [112]. This algorithm which has become perhaps the most popular in the field of
AIS, was later published and represented [114], and again [117] where it was renamed to
CLONALG (CLONal selection ALGorithm). The general CLONALG model involves the
selection of antibodies (candidate solutions) based on affinity either by matching against
4
This is the scope of clonal selection algorithms to the author’s knowledge. It is not expected to be
complete given the vastness and disparate nature of related publications, nor is it required to be complete
for the purposes of this dissertation.
45
Lineage Algorithms Primary Application
CLONALG CSA, CLONALG, CLONALG (1,2), ACS, Optimisation
CLONCLAS, RCSA, MOCSA, IMCSA,
AISMM, SACSA, ECA
AIRS AIRS, AIRS2, Parallel AIRS Classification
BCA BCA Optimisation
IA IA, SIA, I-PAES, CLIGA, CLIGA+, NC- Optimisation
IA, READ-Alg, opt-IA, opt-IMMALG, Par-IA,
Dyn-IMMALG
MISA MISA Multi-Objective Optimisation
Other Too large and varied to classify Varied
Table 3.1: Clonal selection algorithm taxonomy overview.
an antigen pattern or via evaluation of a pattern by a cost function. Selected antibod-

ies are subjected to cloning proportional to affinity, and the hypermutation of clones
inversely-proportional to clone affinity. The resultant clonal-set competes with the exis-
tent antibody population for membership in the next generation. In addition, low-affinity
population members are replaced by randomly generated antibodies. The pattern recogni-
tion variation of the algorithm includes the maintenance of a memory solution set which in
its entirety represents a solution to the problem. A binary-encoding scheme is employed
for the binary-pattern recognition and continuous function optimisation examples, and
an integer permutation scheme is employed for the Travelling Salesman Problem (TSP).
Table 3.2 summarises the algorithms parameters, and Algorithm 3.1 presents a general
formulation of the CLONALG.
The work by Watkins, et al was proposed to exploit the inherent distributedness of
the CLONALG [417], as discussed in Section 2.5.2. In the work, the pattern recognition
variation of the CLONALG was modified such that each memory cell is partitioned to
different processes and evolved independently or in small groups, the results from which
are collated at the end of the algorithm run and returned as the algorithm result. White
and Garret also investigated the pattern recognition version of CLONALG and generalised
the approach for the task of binary pattern classification renaming it Clonal Classification
(CLONCLAS) where their approach was compared to a number of simple Hamming dis-
tance based heuristics [427]. Walker and Garrett investigated CLONALG and Evolution
Strategies (ES) on dynamic function optimisation, showing that although CLONALG can
achieve better results faster than ES on low dimensional dynamic functions, ES consis-
tently outperforms CLONALG on the two high-dimensional problems tested [415]. In
an attempt to address concerns of algorithm efficiency, parameterisation, and representa-
tion selection for continuous function optimisation Garrett proposed an updated version of
CLONALG called Adaptive Clonal Selection (ACS). The mutation parameter, the number
of antibodies selected for cloning, and the number of clones produced for each antibody
were changed to automatic parameters, controlled in a similar way to those in ES [163].
CLONALG is perhaps the most popular basis for elaboration and application with regard
46
Parameter Description
P Repertoire of antibodies.
N The fixed antibody repertoire size.
n The number of antibodies to select for cloning.
L Bitstring length for each antibody.
Nc Number of clones created by each selected antibody. Originally expressed as
a function of the repertoire size (for optimisation) Nc = round(β · N ) (where
β is a user parameter), although a direct integer specification of Nc is simpler.
A rank-based (affinity proportionate) variation of the parameter is presented
for pattern recognition.
d Number of random antibodies to insert at the end of each generation. Random
antibodies replace the d lowest affinity antibodies in the repertoire.
Stop condition Typically a specified number of generations or function evaluations or epochs
of exposure to patterns.
affinity Solution evaluation, typically the solution is decoded into a domain specific
representation and assigned a quality scoring.
clone Duplication of a bitstring.
hypermutate Modification of a bit string where the flipping of a bit is governed by an affinity
proportionate probability distribution. Originally p = exp(−ρ · f ), although
the opt-aiNET variant is also popular p = ( ρ1 ) · exp(−f ) (where ρ is a user
parameter and f is the normalised affinity scoring).
replace The set of the d lowest affinity clones in the population (P ) are replaced with
randomly generated solutions.
Table 3.2: CLONALG parameters.
to clonal selection algorithms, specifically for function optimisation problem instances, the
extent of which is excluded in the interest of brevity.
Artificial Immune Recognition System
After CLONALG, the Artificial Immune Recognition System (AIRS) algorithm is perhaps
the second most popular clonal selection algorithm, although the approach was designed
for and has only been applied to the supervised classification problem domains. The
earliest work on AIRS was in Watkins Masters work [422] that was later published [423].
The approach is a supervised learning algorithm for classification that uses the idea of an
Artificial Recognition Ball (ARB) (introduced in earlier works on the Artificial Immune
Network algorithm) to represent clones (groups) of identical B-cells. The AIRS procedure
involves cloning and somatic hypermutation for preparing a set of real-valued exemplar
vectors suitable for classifying unobserved cases, using a single iteration (epoch is the
machine learning term) over a set of training data. Watkins and Boggess applied the
AIRS to a suite of benchmark classification problems [418], and Goodman and Boggess
problems did the same, comparing to a conceptually similar approach called Learning
Vector Quantization (LVQ) [178].
Given the rapid popularity of the approach Marwah and Boggess investigated the algo-
rithm seeking issues that affect the algorithms performance [286]. They compared various
47
Algorithm 3.1: CLONALG for Function Optimisation.
Input: N, n, L, d, β, ρ
Output: P
1 P ← CreateRandomCells(N, L);
2 while ¬StopCondition() do
3 foreach pi ∈ P do
4 Affinity(pi );
5 end
6 Pselect ← Select(P, n);
7 Pclones ← 0;
8 foreach pi ∈ Pselect do
9 Pclones ← Clone(pi , β);
10 end
11 foreach pi ∈ Pclones do
12 Hypermutate(pi , ρ);
13 Affinity(pi );
14 end
15 P ← Select(P, Pclones , N);
16 Prand ← CreateRandomCells(d, L);
17 Replace(P, Prand );
18 end
19 return P;
variations of the algorithm with modified resource allocation schemes, tie-handling within
the ARB pool, and ARB pool organisation. AIRS was again raced against LVQ by Boggess
and Hamaker on datasets that contained irrelevant features to assess the algorithms ability
to handle noise [49]. Greensmith and Cayzer applied AIRS to hierarchical document clas-
sification [186], that culminated in Greensmith’s masters work [185]. Watkins and Timmis
proposed a new version of the algorithm called AIRS2 which became the replacement for
AIRS1 [419]. The updates reduced the complexity of the approach while maintaining the
accuracy of the results. An investigation by Goodman, et al. into the so called ‘source of
power ’ in AIRS indicated that perhaps the memory cell maintenance procedures played
an important role [179]. A follow-up empirical investigation by Goodman and Boggess
supported the original finding indicating that the process by which new memory cells are
admitted into the ARB pool is critical to the success of the approach [180]. The work
by Watkins, et al., already mentioned briefly in Section 2.5.2 proposed a parallel version
of AIRS permitting the division of training patterns and memory pool suitable to ex-
ploit parallel hardware [420]. The extent of further elaborations and applications of the
approach are excluded in the interest of brevity.
B-Cell Algorithm
Kelsey and Timmis proposed the B-Cell Algorithm (BCA) as an AIS designed for continu-
ous function optimisation [246]. The algorithm maintains a pool of B-cells (binary-encoded
candidate solutions) that are subjected to cloning and mutation. An elitist replacement
population maintenance scheme is applied that ensures only improved cells are admitted
48
into the pool. The mutation operator, which is called contiguous somatic hypermutation
selects a random sub-string of a solution to vary in a probabilistic manner, in what the
authors claim as ‘hot-spot’ mutation. Kelsey, et al. applied the BCA to multi-modal
dynamic chaotic test functions [247]. Empirical algorithm tuning by the authors revealed
that small population sizes (3-5) show better results. In an investigation of AIS applied to
optimisation, Hone and Kelsey provide a case study investigation of the BCA and showed
fractal structures on the complex plane suggesting the potential usefulness of studying
AIS as non-linear dynamical systems [219]. In a further empirical study, Timmis, et al.
compared the BCA to opt-aiNET5 and a Hybrid Genetic Algorithm attributing the partial
success of BCA to the mutation scheme, speculating it results in the escaping of local-
optima search behaviour [390]. The parameters of BCA are summarised in Table 3.3, and
the general procedure for the approach is listed in Algorithm 3.2.
P Repertoire of antibodies.
N Antibody population size.
Nc Number of clones to create of each antibody.
Nr Number of random antibodies to create and insert each generation.
Stop Condition Typically if no progress is made for a number of generations.
hypermutation Uses a processes called contiguous hypermutation, a random location in the
bit string is selected, and a random bitstring length is selected. Each bit in
the sub-string is flipped with the probability ρ.
replace A parent is replaced only if a member of its clone has a higher affinity (greedy
replacement).
Table 3.3: BCA parameters.
A proof of convergence for the BCA was proposed by Clark, et al. using a Markov Chain
model [76]. The proof simplifies the algorithm to an elitist search with a single population
member suggesting that the members of the population can be treated independently
given the lack of interaction during the optimisation procedure. Further, they speculate
that the introduction of inter-solution interactions in the BCA will have a detrimental
effect on results of a search. Finally, in an empirical study Bull, et al. applied the BCA to
what they refer to as less-smooth test problem instances (Diophantine equations) seeking
empirical convergence heuristics [62]. Four variants of the algorithm were compared: an
approach that used an elitist selection mechanism to introduce inter-solution interactions,
and three of what the authors refer to as mega-mutation schemes that attempt to introduce
further diversity into the search. These less-greedy modifications of the approach achieved
a better final result compared to the classical BCA on the test functions chosen, perhaps
suggesting the use of diversity introduction approaches in further BCA applications.
5
(opt-aiNET) is an immune network algorithm for optimisation specified in [108]
49
Algorithm 3.2: B-Cell Algorithm (BCA).
Input: N, Nc , Nr , ρ, L
Output: P
1 P ← CreateRandomCells(N, L);
4 Affinity(pi );
5 end
7 Pclones ← Clone(pi , Nc );
8 Pclones ← CreateRandomCells(Nr , L);
11 Affinity(pi );
12 end
13 pi & ← SelectBest(Pclones );
14 if pi &.affinity ≤ pi .affinity then
15 pi ← pi &;
16 end
17 end
18 end
19 return P;
Immunological Algorithm Family
A simple clonal selection inspired algorithm was proposed by Cutello and Nicosia called
Immunological Algorithm (IA) [95, 96]6 later renamed to Simple Immunological Algorithm
(SIA) [97]. The algorithm maintains a population of B-cells that are exposed to a clonal
expansion process each iteration. This expansion process involves the cloning of cells and
the application of a hypermutation operator, and was demonstrated on the Minimum
Hitting Set Problem (MHSP) and the 3-Satisifiability Problem (3-Sat). The SIA was
extended and an applied to the Graph Colouring Problem (GCP) [100]. The extensions
involved the introduction of a local-search procedure that operated upon each B-cell after
the clonal expansion phase. In addition, rather than an elitist selection method of maintain
the population size after each expansion, an aging operator was introduced for each B-
cell. B-cells are probabilistic deleted from the population using an equation inspired by the
biological literature. Two variations of the aging operator were applied, an elitist version
that ensured the best B-cell’s were not deleted, and a pure strategy that probabilistically
deleted irrespective of the elitist concerns. A birthing operator was also added to ‘top-
up’ the population to the configured size, and an information gain (a stabilisation in the
measure of information discovered by the algorithm) measure was used as the termination
criteria for the algorithm. Given the enthusiastic renaming and tweaking of this family of
approaches, the SIA (the simplest variation of the family) is presented in Table 3.4 and
6
The Immunological Algorithm (IA) is renamed and represented many times by its authors. Other
names include Simple Immune Algorithm (SIA), Cloning Information Gain Aging (CLIGA), and Opti-
mization Immune Algorithm (opt-IA, opt-IMMALG). This linage is named the IA family for simplicity.
50
Algorithm 3.3.
P Antibody population.
L Length of binary string representation.
d Population (repertoire) size.
dup The number of clones created for each antibody.
clone Duplication of the bitstring.
hypermutation Probabilistic modification of a bit string (bit flipping), requires the specifica-
tion (ρ) of the probability of flipping each bit.
Table 3.4: SIA parameters.
Algorithm 3.3: Simple Immune Algorithm (SIA).

Input: d, dup, ρ, L
Output: P
1 P ← CreateRandomCells(d, L);
3 Affinity(pi );
4 end
6 Pclones ← 0;
8 Pclones ← Clone(pi , dup);
9 end
12 Affinity(pi );
13 end
14 P ← SelectBest(P, Pclones , d)
15 end
16 return P;
The probabilistic aging operator was replaced with a simplified generational aging op-
erator by Cutello, et al. in an application to the 2DHP protein folding problem [102]. In
a more detailed study on different varieties of the same protein folding domain, the aging
operator was further tweaked to facilitate longer life spans on some B-cells deemed useful
to the search process [104]. The clonal expansion aspect of the algorithm (cloning and
hypermutation) was grafted into to an existing evolutionary multiple-object optimisation
technique and called I-PAES [91] . The transformed algorithm (generational aging, in-
formation gain stopping criteria) was reviewed and renamed to the ‘Cloning, Information
Gain, Aging’ (CLIGA) algorithm [97]. A modified version called CLIGA+ was proposed
in which each B-cell contains more than one receptor (pattern), permitting application of
the algorithm to pattern recognition tasks. Also proposed in this work is a Noisy Channel
variation of SIA (NC-IA), and a Reaction-Diffusion variation of SIA (READI-Alg) both of
which were applied to instances of the GCP. Cutello, et al. again renamed the approach
51
to Optimisation Immunological Algorithm (opt-IA) and applied the approach to instances
of binary trap functions [94]. An additional fitness inversely-proportional hypermutation
referred to as hypermacromutation was proposed and compared to the traditional static
approach. This algorithm was evaluated again in a larger study involving a number of
machine learning domains [92], and again on a large number of continuous function opti-
misation instances [93]. Cutello, et al. investigated the hypermutation operators of opt-IA
[101]. Cutello, et al. applied opt-IA to the 3DHP protein folding problem [90]. Cutello
and Nicosia applied opt-IA to graph colouring, MHSP, and satisfiability [98]. Work by
Anile, et al. hybridised opt-IA with a direct search method [27]. An extension of opt-IA
called Aligner was proposed by Cutello, et al. and applied to multiple sequence alignment
of DNA [89].
An extension to opt-IA was proposed by Cutello, et al. called the parallel immune
algorithm (Par-IA) which is a master-slave version of the algorithm applied to numerical
function optimisation [99]. Cutello, et al. renamed the approach to Optimisation Immune
Algorithm (Opt-IMMALG), applying the approach to continuous function optimisation
using a real-valued representation as opposed to the binary representation used in pre-
vious works [103]. Also stated in this work was the use of fitness inversely proportional
hypermutation as the standard mutation operator for the approach. This algorithm was
extended and renamed dynamic immune algorithm (dyn-IMMALG) by Cutello, et al. who
propose a dynamic rather than static clonal operator [88]. The approach was applied to
binary trap functions and compared to opt-IA and variations of CLONALG.
Multi-objective Immune System Algorithm
Coello Coello and Cruz Cortes introduced an AIS called the Multi-objective Immune Sys-
tem Algorithm (MISA), and as its name suggests was designed as a population-based
approach for constrained and unconstrained multi-objective optimisation [77]. In the ap-
proach a repertoire of solutions is split into antigens (Pareto non-dominated and feasible
solutions) and antibodies (Pareto dominated and infeasible solutions). A bit-string rep-
resentation is used and antigens are selected at random and matched against antibodies
using Hamming distance. After the selection step, antibodies are cloned, mutated the
population is unioned and reduced back to the configured size, culling the lower quality
solutions. An external (elitist) memory repertoire is maintained of non-dominated feasible
solutions. Solutions are added to the memory set if they are non-dominated by the current
memory set population and sufficiently diverse as determined by a grid-based maintenance
structure. MISA was extended by the same authors and further compared to state of the
art evolutionary approaches for multi-objective optimisation [83]. An EC-AIS hybrid ter-
minology was adopted and an EC-based crossover mechanism was adopted within the
memory set. The main algorithm was simplified such that all population members were
consider antibodies, and only the lower score solutions was selected for cloning and hyper-
mutation. The modified algorithm was shown effective, although it demonstrated rapid
convergence behaviours on benchmark problem instances. Finally, Villalobos-Arias, et al.
proposed a convergence proof for the update MISA showing that the elitism within the
52
algorithm was needed to guarantee convergence [408].
Other
Research into the five core lineages of clonal selection algorithms revealed a large number
(>70 works) on new or extended clonal selection algorithms, predominantly in non-western
conference proceedings. The vast majority of these works involved the application of a
CLONALG, SIA, or derivative algorithms to benchmark and engineering problem domains.
The extent of these works are omitted here for brevity, although the dominant applica-
tions are listed as follows: feature selection for models, parameter tuning for models or
controllers, parameter tuning for a PID controllers, anomaly and/or intrusion detection,
pattern recognition, multi-objective optimisation, function optimisation, general optimi-
sation, multi-user detection, and finally hybridisation with other algorithms.
As discussed in Section 2.4, the clonal selection principle underlies the adaptive quali-
ties of all three of the major paradigms of Artificial Immune Systems. As such, many works
investigating the negative selection and immune network employ B-cells and antibodies
that are modified using processes of selection and expansion with variation. Given that
these approaches do exploit the clonal selection principle they may be considered clonal
selection algorithms, although do not fit into the presented taxonomy as the clonal selec-
tion principle is not the primary focus of the works. This section provides some examples
of such research. Weinand propose a dynamical systems computational model of somatic
mutation of B cells to evaluate the effect of somatic mutation of affinity maturation of the
immune response [424]. Zhang and Hou propose a Niching Clonal Selection Algorithm
(NCSA) that combines negative selection and refinement applied to the pattern matching
problem of anomaly detection [445]. Yu and Hou proposed an extension of CLONALG
called CsAL (Clonal Selection Algorithm) to investigate the negative selection approach
to virus detection [443]. In the body of work on an integrated intrusion detection sys-
tem discussed in Section 2.5.2, the authors propose a static and dynamic clonal selection
algorithms (DynamiCS) which are an augmentation to T-cell inspired negative selection
system that uses clonal selection mechanisms to improve detectors.
3.3.3 Criticisms and Open Problems

The review of the clonal selection principle demonstrated that abstractions of the theory
are generally simplistic, considering the information processing properties and mechanistic
procedure at a high-level, correlating the general AIS observations made in Section 2.6.
The authors de Castro and Timmis reduced clonal selection to a single antigen optimi-
sation system using affinity maturation with controlled affinity-proportionate cloning and
inversely affinity proportional mutation. Cutello and Nicosia reduced clonal selection to
the more general life-long acquisition and retrieval for application of patterns. The com-
mon features to clonal selection abstractions are the properties of (1) population based
information processing (2) clonal expansion of selected high affinity population members,
and (3) affinity maturation via mutation of clones. The commonly executed concerns of
the abstractions include (1) concerns of self and nonself discrimination (cell preparation),
53
(2) distinction between cell types and/or antibody molecules, and (3) cell-cell interactions.
The realised clonal selection algorithms are monolithic, considering antibody confor-
mations in binary and real value shape-spaces, mostly against a single affinity landscape.
As such, the dominant application domains are optimisation (function optimisation), and
supervised classification with procedures that closely resemble genetic algorithms (CLON-
ALG, IA, BCA, MISA) and vector quantisation procedures (AIRS), and the dominant
conceptual and quantitative comparisons have been made with evolutionary algorithms.
The review demonstrated that the trend with these approaches have been to (1) enhance
the algorithm procedure through domain specific assumptions (limited resources in the
case of ARB for classification, and affinity proportionate descent with modification in
CLONALG for optimisation), (2) tweaking and parameter tuning (numerous additional
and alternative operators in the IA family, sensitivity analysis of CLONALG, BCA and
AIRS), and (3) application, benchmarking and racing (the majority of the published re-
search). Based on the detailed review of the state of the art of clonal selection algorithms
in the context of clonal selection and related theory, this section identifies three specific
limitations that represent open problems in the field, as follows:
1. Related Approaches: Clonal Selection Algorithms have generally been related and
compared to Evolutionary Algorithms including the Genetic Algorithm and the Evo-
lutionary Strategy, although broader connections with related computational intel-
ligence fields are typically brief and/or non existent. Systematically placing and
contrasting Clonal Selection as a Computational Intelligence metaphor with related
paradigms provides both a context for comparison of behaviour and capability, and
suggests at the specific overlapping features and findings that may be exploited as
well as those features that differentiate the approach and should be pursued.
2. Implicit Adaptive System: Given the adaptive qualities of the clonal selection theory,
computational models are implicitly assumed to be instances of adaptive systems.
Whether or not a CSA is an adaptive system is not in question, rather the more
important concerns of what it means for computational models of clonal selection to
be classified as adaptive systems. Unlike the Genetic Algorithm that emerged from
Holland’s investigation and formalisation of adaptive systems, the inception and in-
vestigation of clonal selection algorithms has been focused on addressing benchmark
instances of difficulty computational problems. The consideration of the clonal se-
lection as a strategy in the context of an adaptive systems formalism will provide a
stronger foundation of understanding, and insights into the broader and unrealised
information processing capabilities of the approach.
3. Cellular Focus: The scope of clonal selection algorithms are concerned with abstrac-
tions of lymphocyte interactions with antigen, specifically with little or not detail
of the molecular basis of such interactions, nor the broader consideration of where
such interactions occur or the effect they have on the host organism. This cellular
focus is reasonable for the preliminary realisation of computational models inspired
by the theory embodied in the present state of the field, although elaboration and
54
development of so-called ‘second generation’ clonal selection algorithms requires a
holistic consideration of the acquired immune system and the restrictions and con-
straints such a broader perspective imposes on improvement-centric adaptive clonal
selection strategies.
3.3.4 Summary
The detailed treatment of the Clonal Selection Theory in Section 3.2 provided the context
for effectively appraising the state of clonal selection algorithms, specifically with regard to
the task of identifying limitations which may be exploited toward elaborating the general
approach. The taxonomy presented in Section 3.3.2 unified disparate works by general
algorithm lineages, that ultimately highlighted both the deep commonality between the
lineages and the algorithms that belong to them, and the simplicity of the overarching
computational abstraction that underlies such approaches. Importantly this insight pro-
vided the context to identify three limitations and open problems with the clonal selection
paradigm that motivate the remainder of this Chapter. Specifically, Section 3.4 reviews
machine learning approaches to clonal selection highlighting the findings that may be ex-
ploited and the features that differentiate the approach. Section 3.5 considers the clonal
selection strategy in the context of adaptive systems theory, highlighting critical features
of the theory excluded and underrated by existing computational abstractions. Finally,
Section 3.6 address the cellular focus of the paradigm, and outlines an agenda for extend-
ing the approach beyond a discrete repertoire of cells toward distributed clonal selection
algorithms.
3.4 Related Computational Paradigms

This section considers some related machine learning paradigms and their relationship
with the clonal selection principle and inspired algorithms. This review includes both
related paradigms that have been previously considered such as Evolutionary Computation
and Reinforcement Learning, and additional newly considered paradigms such as Lazy
Learning, Competitive Learning and Hill Climbing. The intent of this investigation is to
highlight findings from related computational paradigms that may provide insight into
the behaviour, function, and/or capability of the current state and potential elaborations
of clonal selection algorithms. An interesting aside is the work by Newborough, et al.
[304] provide a framework that suggested that all population algorithms (including clonal
selection) may be treated in a similar manner.
3.4.1 Evolutionary Computation

Evolutionary Computation (EC) is a field of study of Evolutionary Algorithms (EA’s)
that has much in common with AIS, although draws its inspiration for computation from
Darwinian (theory of natural selection) and neo-Darwinian (findings of modern genetics
also referred to as the new synthesis) theories of evolution. Clonal selection algorithms
have a superficial similarity to some EC algorithms such as the Genetic Algorithm (GA),
55
and to the properties of modern Evolution Strategies (ES). See [142, 173] for a classical
treatment of EC and [31, 32] for a modern treatment of the field of EC.
In their work on CLONALG, de Castro and Von Zuben consider the similarity be-
tween a GA and their approach, particularly in regard to the binary representation used
and the stochastic-Darwinian processes employed by both algorithms [112]. They suggest
the differences between the algorithms, include the vocabulary used (genetics and evolu-
tion verses the shape-space formalism and antibody-antigen cellular interactions), and the
somatic mutation and receptor editing used to explore the shape space. The authors also
claim that CLONALG can be categorised as an evolutionary-like algorithm, although they
maintain the same arguments of inspiration, vocabulary, and formalism and the primary
differences [117]. In their book de Castro and Timmis again acknowledge the similarity of
CLONALG to an EA [109]. They are quick to point out that a major difference between
the inspirations of the two approaches is that mutation in evolution is random, whereas the
hypermutation process of clonal selection is controlled and directed, proportional to the
receptors affinity with the triggering antigen (specific to their abstraction). They suggest
that work on EA’s can be leveraged by CSA’s indicating that research on selection oper-
ators (e.g. tournament and roulette wheel selection) may be exploited. The shape-space
formalism is presented as a CSA representation abstraction where non-binary shape-space
schemes and corresponding mutation mechanisms are discussed, also leveraging from re-
search from representation and mutation in EA’s. The authors provide a treatment of
evolution and the clonal selection of the acquired immune system. They suggest an im-
portant difference between the two theories is the fact that in the clonal process expansion
occurs through cell cloning, that there is no sex or genetic recombination, rather only
affinity inversely-proportionate somatic hypermuatation.
In work on the MISA, Coello Coello and Cruz Cortes claim that their approach is not
a genetic algorithm because it does not use recombination (crossover operator) [77], they
later adopt a crossover procedure in their approach, as well as adopt a hybrid of EC and
AIS terminology [83]. In their work evaluating the BCA on function optimisation Timmis,
et al. suggest that BCA is not a GA, based on the empirical performance of the approach
on a small suite of test problem instances, although they are very quick to point out the
limitations of their small study and the requirement for further research [390]. Forrest,
et al. investigated the pattern recognition properties of the immune system [150]. They
used a binary coded GA to model antibody-antigen matching in the immune system, which
included the clonal selection mechanism, claiming “The GA without crossover is a rea-
sonable model of clonal selection, while the GA with crossover models genetic evolution”.
Hightower, et al. use a Binary GA model of somatic hypermutation of clonal selection to
investigate the Baldwin Effect and evolution [207]. Fukuda, et al. [157] and Mori, et al.
[297] used a GA to investigate clonal selection properties and immune network algorithms
for scheduling and resource allocation. The niching-like properties (an EA property in-
spired by theories of population genetics and ecology) were observed by de Castro and
Von Zuben with CLONALG on multi-modal function optimisation [117], and empirically
compared to the fitness sharing approach of Goldberg, et al. [173, 176, 126]. The nich-
56
ing search properties were proposed to occur given the hill-climbing like behaviour of
the independent and semi-independent evolution of B-cells of the various clonal selection
algorithms.
The principle similarity between evolutionary algorithms like the genetic algorithm
and clonal selection algorithms is the Darwinian inspired theories that inspire both fields
(selection and descent with modification). Natural selection accounts for the diversity
of species, whereas clonal selection accounts for the diversity of antibodies. As such,
both processes are imperfect, stochastic process that operate on pre-committed discrete
individuals with short terms goals (relative selective advantage).
The differences between the sub-fields are subtle, related to the functional specifics of
the inspired theories. The theory of natural selection encompasses that of clonal selection,
as the genetics of the organism define the fitness of that organism, one aspect of which is the
structural and function aspects of the organisms adaptive immune system. Related to this
point is that of the time-scales involved, evolution of species such as mammals involves
the accumulation of small changes over geological time, where as the accumulation of
changes in clonal selection are exploited by a single host over its lifetime. Although in
both cases selection, decent, and modification occur on genetic material, the hereditary
of evolution is the entire genome, where as the hereditary of the B-cell lines are those
genes which affect antibody structure. The time-scales also highlight the properties of
learning within the host, evolution involves learning at the species level with a single host
representing a pre-committed unit of adaptation, whereas acquired immunity involves
learning within the host, like the learning that occurs within the brain, so called lifetime
or meta-ontogenetic and meta-epigenetic (non-inheritable changes to DNA) adaptation.
The differences go beyond nomenclature to the specifics of the inspired processes. Selection
in the immune system as a discrete event that triggers (or does not trigger) an immune
response resulting in decent with modification. Selection or survival of the fittest is an
abstraction that involves relative differences in reproductive success not contingent on
anticipated triggered events. The clonal expansion process involves mass replication via
cell division where modification is introduced through high-probability point mutations
in related genes. Darwinian evolution employs both sexual and asexual reproduction,
although evolutionary algorithms typically exploit the crossover of genetic material and
point mutations to introduce variation.
3.4.2 Binary Hill Climbing

A realisation of the clonal selection principle includes the choice of strings of bits as a
first-order representation. This choice of representation is (1) amenable to analysis and
(2) amenable to remapping, and (3) provides a genetic metaphor facilitating genetic-like
operations such as copying errors during cell division. The use of binary strings as a
first-order representation is pervasive in computational intelligence, particularly in the
field of evolutionary computation. Section 3.4.1 demonstrated that there has been much
consideration of the relationship between EC and CSA, although little attention has been
given to strong relationship between optimisation-driven CSA’s and binary hill climbing.
57
This section considers this relationship by providing a brief overview of the state of the
field.
The genetic adaptive plan which became the genetic algorithm operates on a popula-
tion of bit-strings. This representation facilitated the genetic operators of the approach,
and the analysis of the behaviour of the system as processing building blocks (schemata
theorem). One of the generalised genetic operators is mutation, the probabilistic flipping
of bits in the bitstring during genetic replication. Holland proposed that mutation is de-
termined “by a random process where each position has a small probability of undergoing
mutation, independently of what happens in other positions” [216] (page 109). In using
random mutation alone Holland suggested the use of small mutation probabilities such that
the process provides a high probability of maintaining history, which is the dependence of
the approach. Holland pointed out the use of high mutation values results in little depen-
dence between successive generations of individuals, calling such an approach essentially
enumerative. Low probability mutation facilitates the dependence between observations
and new trials, although is suggested to be very unsophisticated compared to genetic
crossover. He proposed the use of low probability mutation as an improvement operator
for existing solutions that rather than used in the construction of new trials. He proposed
the operator should be used to complement the primary genetic operator by restoring lost
genetic material. Further, the application of the mutation operator was described to be
disrupting to the building blocks moved around by the crossover operator, further reinforc-
ing the need for a low probability of occurrence. Holland called mutation a background
operator, and thus the effects of the operator were largely ignored, and the probability
of point-mutations was conventionally kept very small (for example Pm ∈ [0.001, 0.01])
[173, 31]. The probability of a bitstring being mutated is Pm = 1 − (1 − pm )L where Pm
is the mutation rate and L is the string length. Goldberg differentiated between mutation
and crossover by suggesting that selection+mutation provides continual improvement with
limited scope, whereas selection+crossover provides innovation via jumping and remixing
of components [177].
Eiben and Schippers summarised the role of the mutation operator as a uninary (or
asexual) and unbiased operator, highlighting its role as a main operator in evolutionary
strategies (traditionally real numbers) and the primary operator in evolution programming
(traditionally operates on finite state machines) [133]. Spears highlighted the complemen-
tary nature of the two primary genetic algorithm operators suggesting that mutation the
focus if optimality is important and the crossover is useful accumulated payoff is impor-
tant [367]. He suggested that crossover and mutation are two versions of the same general
operator that perturbs genetic representations based on available information. In his dis-
sertation, Spears proposes that mutation has a high exploratory power (in that it is not
dependent on the composition of the population) and no positional dependence (in that
it has the potential to transform a given string into any other string in the search space)
[368]. Work by Schaffer, et al. in the late 1980’s and early 1990’s highlighted the potential
of the mutation operator, suggesting the operators power had been underestimated in ge-
netic algorithms [348, 347]. These works attempted to formalise an optimal mutation rate
58
for genetic algorithms, work that was extended experimentally in [347] and theoretically
[206]. Later work suggested the use of an initially large mutation rate that exponentially
decreased with generational running time of the algorithm. A dynamic and decreasing
mutation rate was further supported by others investigating the potential of the operator
(not limited to [298, 141]). The reason for this is “the probability that a mutation will
give a better string decreases with the number of bits which are correct” [30]. A simplified
genetic algorithm was proposed for investigating the strengths and limitations of the mu-
tation operator, which became known as a mutation hill-climber, (1+1,m) [298] or simply
(1+1) [30]. The algorithm has a static mutation probability (Pm ) applied uniformly and
independently to each bit in the bitstring. The algorithm was called a hill climber because
without crossover, there is no inter-population interaction of genetic information, thus if
a population is used, it is a collection of unrelated parallel hill climbers. The (1+1) is
taken from the Evolution Strategy naming convention ES(µ+λ) where µ parents create λ
offspring, and the best from both ‘+’ are selected. The scheme was reduced to a single
parent and offspring for ease of analysis. The algorithm has the interesting property as a
hill climber that there is no well-defined neighbourhood for a given string. Rather, a prob-
abilistic neighbourhood is defined by the mutation probability. This neighbourhood covers
the entire search space such that it is possible to move from a given string to any other
string in the space, a probability that decreases with the increase in Hamming distance
between strings and the current string. Mühlenbein analysed the algorithm and proposed
an approximate optimal mutation rate (the standard mutation rate) as Pm = L.
1
The CLONALG, BCA and the basis of the IA family (SIA for example) are elaborations
of either a parallel Random Mutation Hill Climbing Algorithm (RMHCA) or the (1+1,m)-
Algorithm. As such, it is reasonable to consider the state of (optimisation-based) clonal
selection algorithms as investigations into parallel hill climbers (GA’s without crossover as
commented by Forrest, et al. [150]). A fair assessment of algorithms such as CLONALG,
BCA, and the IA family requires the consideration of the decades of empirical and theoret-
ical findings from this related field, such as mutation rates, test problems, and existing hill
climbing and mutation-based hill climbing algorithms for comparison. This perspective
may also highlight a severe weakness in the state of the art of optimisation-centric clonal
selection algorithms.
3.4.3 Lazy Learning

Instance-based learning is a non-parametric and supervised machine-learning paradigm
where a model is constructed from domain data instances at query time. The complexity
increases with the quantity of data, and the more data available, the more specific the
model [8, 342]. Instance-based learning is typically referred to as Lazy Learning, which
is a machine learning paradigm where generalisation occurs as required. Lazy learning
may be contrasted with Eager Learning in which generalisation occurs before a query is
received. In addition to the deferral of generalisation, lazy learning is also typified by
the combination of stored data used to reply to requests, and the fact that the response
is discarded after it is delivered [6, 7, 426]. The thee principle characteristics of lazy
59
learning are: Defer : Store all training data and defer processing until queries are given
that require replies, Demand-Driven: Queries are answered by combining training data,
using a local learning (neighbourhood) approach. Instance are (1) identified in as points
space, (2) a similarity measure is used to define the neighbourhood, (3) prediction function
uses information to answer a query. Discard the answer and any intermediate results after
answering the query.
Instance-based learning may be considered a type of local learning algorithm, where
reasoning is based on models (kernels) constructed from instances similar to the input
instance. A typical concern with such models is the capability of the system to gener-
alise. This represents a trade-off between the capacity (size) of the model and the number
of samples to consider in the model. Bottou and Vapnik rephrased this local learning
trade-off to be between the capacity of the model and the locality of the samples, where
locality is the shape and size of the instance selection region (neighbourhood) [51]. This
locality property of model construction defines how local an algorithm is, such that an
instance-based approach such as k-Nearest Neighbour is highly-local with a low capacity,
and a connectionist approach such as a neural network is non-local with high capacity.
Case-based Reasoning (CBR) (or memory-based reasoning), is a generalised application of
instance-based learning applied to reasoning such as expert systems [5]. Aamodt and Plaza
provide a concise description of the various different identities of instance based learning
from a case-based reasoning perspective [1]. Their taxonomy highlights the perspective
each identity places on this style of learning which covers exemplar, instance, memory,
case, and analogy based learning and reasoning. Further, Aamodt and Plaza propose a
descriptive framework for CBR, which includes a general algorithmic cycle for the appli-
cation of the approach, as follows: Retrieve the most similar case or cases, Reuse the
information and knowledge retrieved to solve the problem, Revise the proposed solution,
and Retain the parts of the experience that may be useful for future problem solving.
There are two primary classes of lazy learning approaches: Nearest Neighbour Meth-
ods, and Kernel Methods. Nearest neighbour methods are function approximation al-
gorithms that search through a database for similar instances (cases) to a given input
instance and make a prediction [84]. The similar points to consider in the prediction
are referred to as the neighbourhood, and the number of points is typically denoted k,
thus the scheme may be referred to as k-nearest neighbour or kNN. A natural extension
to nearest neighbour approaches is that each case generates a local density function or
kernel. Predictions may be made by summing together densities functions. The ker-
nel function depends upon the distance function between cases, where the Gaussian is
a typical function employed. Some techniques include distance weighted regression such
as Linear Weighted Regression methods [29], and a neural-network like approach called
Radial Basis Function [59, 61, 328]. Much of the work on instance based learning algo-
rithms is focused on reduction of storage requirements (thus increasing the efficiency of
the approach), and improving the algorithms robustness to with regard to attribute noise
and irrelevant attributes. Perhaps the most popular concern of nearest neighbour based
approaches is what has been referred to as the curse of dimensionality. This refers to the
60
fact that in common datasets, the nearest neighbour is approximately the same distance
away from a given point as the furthest neighbour. This effect is observed in dimensions
as low as 10-15 [41]. The problem is caused by the exponential increase in volume of a
space by the adding of additional dimensions7 .
The clonal selection principle does share some commonality with the instance based
learning, particularly in the context of the continual adaptation to antigen and Aamodt
and Plaza’s four R’s of CBR. The specific knowledge about the antigen is discarded after
the infection is addressed, although is retained implicitly in the selected cognate receptors.
The AIRS is a realisation of clonal selection that considers the the repertoire as a model
of cases that are compressed (quantised) to a set of exemplars. For example, a recent
work by Seeker and Freitas [353] make the argument that AIRS is nothing more than an
instance-creating algorithm for use by kNN. Exemplar-based modelling is used in a vari-
ety approaches for function approximation problems (such as the LVQ algorithm), which
represent an important application perspective on clonal selection algorithms (considered
further in Section 8.4). A similar approach was taken on the pattern recognition variation
of CLONALG, although the number of patterns was known and not compressed as in the
case of AIRS.
3.4.4 Competitive Learning

Intrator and Edelman commented that competition between functional units is a widespread
biological phenomenon, particularly in the vision systems, the brain, and other sensory
systems [226]. Competitive learning systems are based on the idea of competition for
activation, and lateral (same level) inhibition. Without inhibition, the models result in
an averaged or muddied converged state. They define competitive learning as a dynamic
redistribution of responsibilities of various units over parts of the representation space and
propose a global, local, and a hierarchical perspective. The authors suggest that resource
allocation may be addressed through selective inhibition, such that those units or mod-
ules that have not been activated for some time are inhibited less. Further, they suggest
investigation into two types of competitive learning: (1) competitive learning via lateral
inhibition, (2) competitive learning via the top-down separation of flows of information.
Three perspectives of competitive learning are as follows [226]:
• Global Competition Competition over the entire representation space, with a strong
requirement for inhibition via lateral connections between neurons and the distinc-
tion between hard (single winner) and soft (multiple winner) competition
• Local Competition The division of the representation space such that local experts
(modules of units) compete (known as mixture of experts). Assumes that different
processes generate different parts of the representational space, thus hard competi-
tion assume a one-to-one matching between modules and hidden generator processes
that they model.
7
It should be noted that the curse of dimensionality has been considered in the context of negative
selection algorithms by Stibor, for example [374]
61
• Hierarchical Competition A mixture of global and local competition, where the rep-
resentation space is partitions or split into a hierarchical tree structure.
Competitive learning is a connectionist machine-learning paradigm where an input

pattern is matched to the node with the most similar input weights, and the weights
are adjusted to better resemble the input pattern. This is called the Winner-Take-All
(WTA or maximum activation) unsupervised learning method where the input pattern is
compared to all nodes based on similarity. The nodes compete for selection (or stimula-
tion) and ultimately adjustment (or learning) [276]. Some popular competitive learning
algorithms include Kohonen’s Self-Organising Map (SOM), Adaptive Resonance Theory
(ART), and Counterpropagation Networks, and Neural Gas. SOM is a seminal achieve-
ment in the field of competitive learning inspired by the spatial order and organisation
of brain functions [261]. The Self-Organizing Map is an unsupervised learning algorithm
(clustering) that embodies the WTA principle and facilitates topological preservation in
the face of dimensionality reduction. The Learning Vector Quantization (LVQ) is a su-
pervised learning algorithm (classification) and is a zero-order SOM with no connectivity
between nodes. Without the connectivity, there is no geometry or topological neighbour-
hood effects, although the unstructured collection of exemplars compete for stimulation
using the winner-take-all principle, and winners that miss-classify during the training pro-
cess are suppressed. A problem with the winner-take-all learning principle is that units
may win too much, and thus dominate selection and response. The result is that some
units may never win and the units are considered dead (resources are underutilised) [341].
A relatively simple solution is to dampen the competitiveness of the principle by keep-
ing track of how often units win competitions and using this information to reduce the
chance of wining in the future. This may be achieved by introducing a conscience to the
units [128], or a similar method called frequency-sensitive competitive learning [9]. This
effect may also be achieved by using a generalised version of the suppression mechanism of
LVQ2 called rival-penalised competitive learning, where the second winner (rival) in each
competition is suppressed [439].
Grossberg with his Adaptive Resonance Theory (ART) [189, 188, 190], along with
von der Malsburg [412] were among the first to propose theories of self-organisation and
competitive learning of neural cells. Grossberg exploited ART as the basis of a network
architecture and unsupervised competitive learning model [71, 191]. The model addressed
the instability of unsupervised competitive learning in the face of unexpected irregular-
ities in the input signals. The ART models (ART1, ART2, and others) are in effect
self-regulating control structures for autonomous learning and recognition. Hecht-Nielsen
combined Kohonen’s feature map with Grossberg’s outstar network structure calling it
a counterpropagation network. This network produces a mapping (like backpropagation)
that functions as a statically optimal self-programmable look-up table [202, 203]. Competi-
tive Hebbian Learning is a combination of competitive learning and Hebbian-style learning
in which nodes specialise their response to inputs [428, 430, 429]. Martinetz employed com-
petitive Hebbian-learning rule to construct topology preserving graphs by inserting edges
between nearest-neighbour nodes distributed across feature space based on input signals
62
from the domain [282]. The Neural Gas algorithm of Martinetz and Schulten, like SOM is
another self-organising and unsupervised competitive learning algorithm [285, 283, 284].
Unlike SOM (and more like LVQ), the nodes are not organised into a lower-dimensional
structure, instead the competitive Hebbian-learning like rule is applied to connect, order,
and adapt nodes in feature space. Martinetz calls this the winner-take-most learning rule
distinct from Kohonen’s winner-take-all rule. The result of the Neural Gas algorithm is a
set of points that are distributed across feature space in relatively proportion to the input
signal density, with a graph topology that preserves structures in the feature space.
Lymphocytes compete with each other for binding to an antigen. This process has
stochastic aspects (the units are mixed in a diffuse substrate), and deterministic aspects
(physics of the chemical bonding). Generally, the population is scanned and the highest
affinity receptor wins (is proliferated and adapted). Clonal selection is an application of
the winner-take-all principle which is manifest in the clonal dominance of receptors for
antigen. This effect may also account for the antigenic sin, where the greedy-nature of
the winners (winners keep winning) results in lateral inhibition of cells that may be better
suited for an antigen. A second interesting point is that the inhibitory effect of clonal
selection is attrition and dominance not only of the competition, but also with regard to
resource allocation. Cells have a finite lifetime before they are discarded, and the cell
population is continually turned-over. Therefore, those receptors that are not used, are
not preserved in memory (memory cells) and are lost.
3.4.5 Reinforcement Learning

Reinforcement Learning is an area of study in the field of machine learning characterised
by system (or agent) learning to perform a task through a process of trial-and-error in an
environment that provides minimal feedback in the form of delayed rewards. Sutton and
Barto provide a seminal treatment of reinforcement learning, commenting that problems
and not learning methods characterise the field [376].
In the proposal of the CLONALG, de Castro and Von Zuben commented on the com-
monality between the clonal selection process and reinforcement learning. The learning
task of the immune system may be the identification and neutralisation of antigen with
the goal of host survival, where the clonal selection process reallocates resources to an-
tibodies (B-cell lines) that are demonstrated as effective. They further comment that
feedback is received by the process in what may be interpreted as an unsupervised man-
ner through the re-introduction of previously encounter antigen. They highlight the de-
sire to efficiently allocate limited resources, and the importance of generalisation via the
cross-reactive response. The de Castro-Von Zuben interpretation is simplified from the
conventional Sutton-Barto definition in which a delayed numerical scoring is given to the
entire system given the systems overall performance on the whole uncertain environment.
This simplification highlights the question as to how to concisely phrase clonal selection,
acquired and even the immune system as a reinforcement learning problem. Sutton-Barto
point out that in such a phrasing that a system must perceive the state of its environment,
take actions to effect this state, and have internal goals that relate to this state (sensa-
63
tions, actions, goals). An adjustment to de Castro and Von Zuben’s description of clonal
selection better relating it to reinforcement learning is as follows: Sensations collectively
are the summation of all antigen perceived by the immune system (for example the ex-
ogenous pathogenic, and/or internal antigenic environment), Actions collectively are the
mechanisms of the clonal selection process applied discretely to each antigenic exposure,
and Goals collectively is the systems efficiency and effectiveness at pathogen identification
and neutralisation toward host survival, the feedback of which may be perceived through
injury to host tissues from infection or the holistic performance of the system for antigen
identification. A reinforcement learning approach that combines aspects of pattern match-
ing and adaptation using a genetic algorithm to evolve a set of binary encoded rules, called
Learning Classifier Systems. The relationship of this approach has been made between
negative selection algorithms (such as ARTIS), which is discussed further in Section 3.5.3.
3.4.6 Summary
This investigation reinforced the important association of clonal selection algorithms with
the field of Evolutionary Computation, specifically with genetic algorithms as they are
generally both stochastic adaptive process that operate on pre-committed discrete units,
motivated by short term relative goals. The review of the related field of binary hill
climbing algorithms demonstrated that the bit-string based CLONALG, BCA and much
of the IA family may be considered population-based elaborations of the Random Mutation
Hill Climbing Algorithm (RMHCA) or variations of the greedy (1+1,m)-Algorithm. The
clonal selection principle and the exemplar-based AIRS share common properties with the
field of lazy learning, specifically the deferred and demand driven properties of instance-
based learning and the retrieve, revise, retain principles of case-based reasoning. Finally,
the principle and algorithms exhibit the competition for selection and winner-take-all
properties of competitive learning, where clonal dominance and proportional allocation of
resources results in an implicit lateral inhibition of less competitive cells.
3.5 Clonal Selection Adaptive Strategy

This section explicitly considers the computational abstraction of clonal selection in the
context of adaptive systems theory. An adaptive systems formalism is reviewed and used
to phrase a clonal selection adaptive plan, that is generally similar to Holland’s classical re-
productive and genetic plans. A modern complex adaptive system approach is considered,
providing insights into both the acquired immune system and into assessing the emergent
properties of clonal selection based approaches.
3.5.1 Adaptive Systems

Adaptive systems are difficult to define, compare, and evaluate. The use of formal tools is
necessary for the investigation of systems of this type. Holland proposed a formalism in
his seminal work on adaptive systems that provides a general manner in which to define
an adaptive system [216]. Phrasing systems in this way provides a framework under which
64
adaptive systems may be evaluated and compared relative to each other, the difficulties
and obstacles of investigating specific adaptive systems are exposed, and the abstracted
principles of different system types may be distilled. This section provides a summary
of the Holland’s seminal adaptive systems formalism and considers clonal selection as an
example of an adaptive plan.
Adaptive Systems Formalism
This section presents a brief review of Holland’s adaptive systems formalism taken from
[216] (Chapter 2). This presentation focuses particularly on the terms and their descrip-
tion, and has been hybridised with the concise presentation of the formalism by De Jong
[120] (page 6). The formalism is divided in two sections: (1) Primary Objects summarised
in Table 3.5 and (2) Secondary Objects summarised in Table 3.6. Primary Objects are
the conventional objects of an adaptive system: the environment e, the strategy or adap-
tive plan that creates solutions in the environment s, and the utility assigned to created
solutions U .
Term Object Description

e Environment The environment of the system undergoing adaptation.
s Strategy The adaptive plan which determines successive structural modifications
in response to the environment.
U Utility A measure of performance or payoff of different structures in the envi-
ronment. Maps a given solution (A) to a real number evaluation.
Table 3.5: Summary of primary objects in the adaptive systems formalism.
Secondary Objects extend beyond the primary objects providing the detail of the for-
malism. These objects suggest a broader context than that of the instance specific primary
objects, permitting the evaluation and comparison of sets of objects such as plans (S),
environments (E), search spaces (A), and operators (O).
A given adaptive plan acts in discrete time t which is a useful simplification for analysis
and computer simulation. A framework for a given adaptive system requires the definition
of a set of strategies S, a set of environments E, and criterion for ranking strategies X. A
given adaptive plan is specified within this framework given the following set of objects:
a search space A, a set of operators O, and feedback from the environment I. Holland
proposed a series of fundamental questions when considering the definition for an adaptive
system, which he rephrases within the context of the formalism (see Table 3.7).
Some Examples
This section provides a few examples of phrasing adaptive systems using the formalism
and examples as to how the formalism may be used as a framework in the investigation of
adaptive systems. Holland provides a series of illustrations rephrasing common adaptive
systems in the context of the formalism [216] (pages 35-36). Examples include: genet-
ics, economics, game playing, pattern recognition, control, function optimisation and the
65
Term Object Description
A Search Space The set of attainable structures, solutions, and the domain of action
for an adaptive plan.
E Environments The range of different environments, where e is an instance. It may
also represent the unknowns of the strategy about the environment.
O Operators Set of operators applied to an instance of A at time t (At ) to transform
it into At+1 .
S Strategies Set of plans applicable for a given environment (where s is an instance),
that use operators from the set O.
X Criterion Used to compare strategies (in the set S), under the set of environ-
ments (E). Takes into account the efficiency of a plan in different
environments.
I Feedback Set of possible environmental inputs and signals providing dynamic in-
formation to the system about the performance of a particular solution
A in a particular environment E.
M Memory The memory or retained parts of the input history (I) for a solution
(A).
Table 3.6: Summary of secondary objects in the adaptive systems formalism.
central nervous system. Holland applies the formalism to investigate his schemata the-
orem, reproductive plans, and genetic plans, the foundational models became the field
of Evolutionary Computation. From working within the formalism, Holland makes six
observations regarding obstacles that may be encountered whilst investigating adaptive
systems [216] (pages 159-160): (1) High cardinality of A: makes searches long and storage
of relevant data difficult. (2) Appropriateness of credit: knowledge of the properties about
‘successful’ structures is incomplete, making it hard to predict good future structures from
past structures. (3) High dimensionality of U on an e: performance is a function of a large
number of variables which is difficult for classical optimisation methods. (4) Non-linearity
of U on an e: many false optima or false peaks, resulting in the potential for a lot of wasted
computation. (5) Mutual interference of search and exploitation: the exploration (acqui-
sition of new information), exploitation (application of known information) trade-off. (6)
Relevant non-payoff information: the environment may provide a lot more information in
addition to payoff, some of which may be relevant to improved performance.
Cavicchio provides perhaps one of the first applications of the formalism (after Holland)
in his dissertation investigating Holland’s reproductive plans [72] (and to a lesser extent
in [73]). The work summarises the formalism, presenting essentially the same framework,
although he provides a specialisation of the search space A. The search space is broken
down into a representation (codes), devices (solutions), and a mapping function from
codes to devices. The variation highlights the restriction the representation and mapping
have on the designs available to the adaptive plan. Further, such mappings may not be
one-to-one, there may be many instances in the representation space that map to the
same solution (or the reverse). Although not explicitly defined, Holland’s specification
of structures A is clear in pointing out that the structures are not bound to a level of
abstraction, that his definition covers structures at all levels. Nevertheless, Cavicchio’s
66
Question Formal
To what parts of its environment is the organism (system, organisation) adapt- What is E?
ing?
How does the environment act upon the adapting organism (system, organisa- What is I?
tion)?
What structures are undergoing adaptation? What is A?
What are the mechanisms of adaptation? What is O?
What part of the history of its interaction with the environment does the organ- What is M ?
ism (system, organisation) retain in addition to that summarised in the structure
tested?
What limits are there to the adaptive process? What is S?
How are different (hypotheses about) adaptive processes to be compared? What is X?
Table 3.7: Fundamental questions regarding adaptive systems, taken from [216] (page 29).
specialisation for a representation-solution mapping was demonstrated to be useful in

his exploration of reproductive plans (early genetic algorithms). He proposed that an
adaptive system is first order if the utility function U for structures on an environment
encompasses feedback I. Cavicchio described the potential independence (component-
wise) and linearity of the utility function with respect to the representation used. De Jong
also employed the formalism to investigate reproductive plans in his dissertation research
[120]. He indicated that the formalism covers the essential characteristics of adaptation,
where the performance of a solution is a function of its characteristics and its environment.
Adaptation is defined as a strategy for generating better-performing solutions to a problem
by reducing initial uncertainty about the environment via feedback from the evaluation of
individual solutions. De Jong used the formalism to define a series of genetic reproductive
plans, which he investigated in the context of function optimisation.
3.5.2 Clonal Selection Adaptive Plan

Holland’s work on adaptive systems, along with collaboration with colleagues at the
Santa Fe Institute evolved into the field (and formalism) of Complex Adaptive Systems
(considered in the next Section). Nevertheless, the adaptive systems formalism Holland
proposed in 1975 is distinct and remains a useful tool, particularly in the phrasing and
investigation of the information processing properties of biologically inspired adaptive
plans. This section considers clonal selection as an adaptive plan, similar to Holland’s
reproductive and genetic adaptive plans.
Primary Objects A primary objects interpretation of clonal selection reveals an adap-

tive plan that encompasses the extent of clonal selection algorithms. The environment e is
comprised of one or a set of antigen, the strategy s is the successive structure modification
of a repertoire of antibody using antigen-driven selection and expansion with variation.
The utility U is the antigen-wise affinity. This mapping provides little insight, other than
the explicit identification of the environment as an actor in the process. As was demon-
67
strated in the review of the abstraction of the theory and resultant algorithms, the focus
lies with the clonal selection principle itself, where as the environment, which defines the
scope of the required adaptation, is ignored, and/or simplified to a single antigen (cost
function in the case of optimisation).
Secondary Objects This section considers the detail of the formalism, the secondary
objects, by addressing the questions proposed by Holland (listed in Table 3.7).
Environments (E) The acquired immune system adapts to antigen. Antigen may
be endogenous, such as self-tissues and misbehaving cells (cancer), and exogenous such
as pathogen. The tradition with clonal selection abstractions is to ignore the self-nonself
discrimination problem which is delegated to the negative selection paradigm, thus one
may may either consider a general antigenic environment (exogenous and endogenous), or
just a pathogenic environment (exogenous). Further, each host has a distinct antigenic
environment given the unique internal composition and exposure with the environment.
Feedbacks (I) The environment acts upon the adapting system by explicitly se-
lecting those structures to adapt. The environment is exposed discretely to parts of the
system, where the internal repertoire of cells and molecules is compartmentalised and
distributed throughout the host, and antigen of various sources have discrete entry and
contact points with the repertoire within the host organism. The antigenic-wise feedback
of affinity and avidity is relative to the space and time of the specific antigen and the
specific sub-repertoire. This is an elaboration on the simplified clonal selection models in
which antigen-antibody interaction is holistic with regard to both sets (all antigen interact
with all antibody). In addition, different antigen may have different effects on the host,
such as pathogen virulence or the speed of cancer metastasis. These antigen behaviour
may effect the amount of damage or potential damage a given antigen may cause to the
host organism, perhaps perceived by the acquired immune system as danger signals (Dan-
ger Theory), or through the release of specialised chemicals such as cytokines. Finally,
environmental feedback may be assessed implicitly where the receptor affinity or avidity
for given antigen may be a measure of the responsiveness, and thus the utility of the
system.
Search Space (A) The search space has been well considered in the abstractions
of the theory. The structures undergoing adaptation are the cell receptors, which are
defined by receptor genes of the genome within each cell, and are manifest in protein
conformations. As discussed, the common abstraction for this space is combinatorial
shape-space formalism. One may consider the breakdown of A proposed by Cavicchio
into that of representation, devices, and the mapping which clearly correlates with the
framework proposed by de Castro and Timmis, discussed in Section 3.2.2.
Operators (O) The mechanisms of adaptation have also been well considered in
abstractions of the theory. Beyond the generic selection, expansion, hypermutation, there
68
is an array of genetic diversity mechanisms outlined in Section 3.2.1, there are also inter-
cell interactions such as Helper T-cells, and receptor-receptor recognition described in the
immune network theory. Although the scope of the simple clonal selection operators is well
considered, those operators which may influence such first-order operators are employed
have not (such as cellular interactions during clonal selection).
Memory (M ) The history of the interaction between receptors and antigen is recorded
using two memory mechanisms. Implicit memory occurs through the change in a given
receptors density given the exponential increase in cells that occurs after an exposure.
Explicit memory, which is exploited by some clonal selection algorithms (CLONALG for
example) is achieved through creation and retention of long-lived memory cells.
Strategies (S) The limits to the adaptive process have not been investigated. The
adaptive process may be mediated by helper cells or un-mediated, it may be triggered
by an exogenous or endogenous antigen. Beyond combinations of the operator set O, the
clonal selection is localised to a space and time within the repertoire. Such constraints
on antigen-centric selection have not been considered. Therefore, the consideration that
influences the process suggested in the operators set provide much opportunity for different
strategies beyond simple reflexive clonal selection.
Criterion (X) The adaptive processes may be compared by their general capabilities
of identifying and neutralising antigen in an environment. Some adaptive plans may be
suited to general or specific identification, others may be suited to endogenous or exogenous
antigenic environments. Given that S considers that there are multiple clonal selection
strategies, and E considers that there are multiple of such environments for strategies, the
consideration of general clonal selection centric assessment criterion has not previously
been considered.
Summary The primary objects consideration reveals, a first-order adaptive system

where a simple abstraction of the clonal selection procedure considers a payoff that en-
compasses the extent of the interaction with the environment, where typically (in the case
of optimisation) an environment is composed with one antigen and one affinity landscape.
The secondary objects consideration of clonal selection provides an explicit decomposition
and elaboration of the procedure, highlighting a number of weak points not considered,
including: the environment as an actor in the process, a range of clonal selection envi-
ronments, numerous explicit and implicit feedbacks beyond a single affinity landscape, an
implicit memory mechanism, the potential for a class of clonal selection adaptive plans,
and clonal-selection centric criterion for comparing such plans. In addition, the elabora-
tion highlighted those aspects that have been considered in existing abstractions including:
the search space formalism, the operators used in the procedure, and the core (simplest)
arrangement of the operators. Therefore, the explicit phrasing of clonal selection as a
Hollandian adaptive system provided a clear framework for communicating (1) the sub-
elements of the strategy, and (2) the extent to which those sub-elements have or have not
69
been investigated.
3.5.3 Complex Adaptive Systems

Adaptive strategies are typically complex because they result in irreducible emergent be-
haviour that occur as a result of the non-linear interactions of systems components. The
study of Complex Adaptive Systems is the study of high-level abstractions of natural
and artificial systems that are generally impervious to traditional analysis techniques.
Macroscopic patterns emerge from the dynamic and non-linear interactions of the systems
low-level (microscopic) adaptive agents. The emergent patterns are more than the sum
of their parts. As such the traditional reductionist methodology fail to describe how the
macroscopic patterns emerge. Rather, holistic and totalistic investigatory approaches are
applied that relate the simple rules and interactions of the simple adaptive agents to their
emergent effects in a ‘bottom-up’ manner. Some relevant examples of CAS include: the
development of embryos, ecologies, genetic evolution, thinking and learning in the brain,
weather systems, social systems, insect swarms, bacteria becoming resistant to an antibi-
otic, and the function of the adaptive immune system. The field of Complex Adaptive
Systems was founded at the Santa Fe Institute (SFI), in the late 1980’s by a group of
physicists, economists, and others interested in the study of complex systems in which
the agents of those systems change [21]. Perhaps one of the largest contributors to the
inception of the field from the perspective of adaptation was Holland. He was interested in
the question of how computers could be programmed so that problem-solving capabilities
are built up by specifying: “what is to be done” (inductive information processing) rather
than “how to do it” (deductive information processing). In the 1992 reprint of his book
he provided a summary of CAS with a computational example called ECHO [216]. His
work on CAS was expanded in a later book which provided an in depth study of the topic
[218]. There is no clear definition of a Complex Adaptive System, rather sets of parsimo-
nious principles and properties, many different researches in the field defining their own
nomenclature. Popular definitions beyond Holland’s work include that of Gell-Mann [166]
and Arthur [28].
Immune System as a Complex Adaptive System The acquired immune system

of vertebrates is an often cited example of a natural complex adaptive system. Some
examples of the acquired immune system phrased as a complex adaptive system include:
A simulation of the immune system and the AIDS HIV using Cellular Automata [187]. A
simulation of vaccination in the immune system using agents [10], a general summary of the
immune system and simulation approaches in the context of CAS [11], and the simulation
of an immune system and HIV using elements of classifier systems, genetic algorithms
and cellular automata [383]. The connection between CAS and the field of Artificial
Immune Systems has been observed. Hofmeyr and Forrest considers the relationship
between the negative selection-based ARTIS, the acquired immune system, and CAS [214].
They proposed that representation is fine-grained where peptides are the basic unit of
representation, and that coordination is emergent from the interaction of trillions of cells
70
and molecules. They also considered the importance of the systems tight-coupling with
its environment, and specifically the immune systems ability to respond to perpetually
novel stimuli. Vargas, et al. consider the immune system and Artificial Immune Systems
as examples of complex adaptive systems [405], and phrase an AIS inspired Learning
Classifier System [406, 404].
Beyond explicitly investigating immunology as a CAS, the formalisation of such sys-
tems can provide insights into the general properties of the immune system as a complex
adaptive system. For example, Holland suggested that such systems respond instant-
by-instant, and the importance of the systems ability to balance exploration (acquisition
of new information or capabilities) with exploitation (efficient use of information or ca-
pabilities already available) [217]. Arthur proposed a definition of CAS in the context
of economics with six properties [28], all of which resonate with the information pro-
cessing properties of the acquired immune system. Generally these properties are: the
dispersed interaction of possibly heterogeneous agents acting in parallel, the lack of a
global controller, the cross-cutting interaction of units across hierarchical concerns, con-
tinual adaptation in the presence of perpetual novelty, and finally dynamics where the
system operates far from global optimum and equilibrium. Finally, Jost suggests that the
environment in which a CAS exists is more complex than the CAS itself and that CAS
depends on regularities in its environment [237]. Jost provides a rigorous assessment of
CAS in the context of internal and external complexity. External complexity is defined
as the amount of input, information, energy the system obtains from the environment.
Internal complexity is the complexity of the internal representation of the information it
takes as input (model complexity). The goal of the system is to handle as much input as
possible with as simple a model as possible, to attempt to increase the external complexity
and reduce the internal complexity of the system.
Clonal Selection The immune system may be an instance of clonal selection, although
the computational properties of clonal selection alone are not. Clonal selection describes
the structure and function of an opportunistic improvement mechanism for specialising a
repertoire of molecular pattern recognition cells. A broader perspective of clonal selection
that takes into account more than receptors, lymphocytes, and antigen introduces com-
plexity into the adaptive system that will make effective assessment difficult. A complexity
perspective on adaptive systems may be exploited in the investigation of computational
models of clonal selection by focusing effort on a top-down analysis emergent or systemic
effects rather than a bottom-up analysis of the discrete components. Specifically, two
top-down clonal selection centric concerns include: (1) the information composition of the
system which may be explicitly aggregated and assessed in the case of distributed system
(for example information in the holistic repertoire), and (2) the holistic system capabil-
ity in a given environment potentially aggregated across distributed contact points (for
example capability across the holistic repertoire).
71
3.5.4 Summary
Clonal selection is a theory that defines the adaptation of a general repertoire of lympho-
cytes with surface bound immunoglobulin, toward a repertoire with specialised capability
for a hosts antigenic environment. Holland’s adaptive system formalism provided a frame-
work for defining the sub-elements of clonal selection as an adaptive plan, highlighting
the generally underestimated role of the environment and the feedbacks and interactions
between the components of the system. Specifically, the scope of reviewed clonal selection
algorithms may be considered first order adaptive systems where the extent of the feed-
back to the system is presented in the utility function. Modern adaptive systems theory
embodied in the study of Complex Adaptive Systems, highlighted the inherent complexity
of the environment in which a system operates, and the difficulty in using bottom-up ap-
proaches to analyse such systems. The perspective suggested at the consideration of the
immune system as an instance of a CAS, where clonal selection alone provides an adaptive
quality to the system and is in and of itself not an CAS.
3.6 Cellular Strategy and Beyond

This section considers the cellular-focus criticism of clonal selection algorithms made in
Section 3.3.3. Specifically, given that cellular clonal selection may be considered a mo-
tivating metaphor for adaptive information processing, elaborations of this foundational
approach are considered including explicit investigation in the context of related immuno-
logical metaphors as well as related cellular interaction concerns. The problem of dis-
tributed clonal selection is considered in the context of the structure and function of the
broader immune system. An agenda is outlined toward distributed clonal selection al-
gorithms that suggests: (1) the exploitation of immune physiology with regard to the
lymphatic system and lymphocyte migration, and (2) the exploitation of populations of
interacting immune systems with regard to immunisation and evolutionary immunology.
3.6.1 Adaptive Cellular Strategy

This chapter has reviewed clonal selection from an immunological (Section 3.2), compu-
tational (Section 3.3), and adaptive systems perspectives (Section 3.5) providing a abun-
dance of support to suggest at the information processing properties of the motivating
metaphor. The theory is adaptive by definition as it describes how a general repertoire
of lymphocyte receptors are specialised (adapted through iterative improvement) in the
context of an environment. The adaptive theory relies on the replication properties of
immune cells and iterative improvement depends upon the blind mutation-based affinity
maturation to antigenic signals. Receptor interaction with antigen is a pattern recognition
process involving the ‘fitting’ or ‘matching’ of pre-committed molecular structures, where
each attempt at match with antigen requires a localised proximity and physical contact
(trials within trials). The geometric paradigms from theoretical immunology phrase the
scope and preparation of receptor configurations as a finite combinatorial search problem.
Pair-wise relationships between configurations can be assessed in the context of a speciali-
72
sation of Wrights fitness landscape formalism that stresses the adaptation of the ‘goodness
of fit’ of structures.
A first-pass exploitation of the underlying information processing qualities of the the-
ory elicited a suite of adaptive algorithms that superficially show a strong resemblance
to parallel mutation hill climbing algorithms for the purposes of global optimisation and
vector quantisation. A deeper assessment of the underlying adaptive strategy in the con-
text of adaptive systems theory provided insights into the behaviour of the approach, and
motivations for elaboration. Regarding behaviours, the strategy uses adaptation toward
information acquisition in a population memory structure using accumulated trial-and-
error with pre-committed structures, each of which is individually redundant. More inter-
estingly, this redundancy may promote degeneracy in contextual emergent specificity in
addition to general fault tolerance. Two specific areas were highlighted for elaboration, as
follows:
Cellular Interactions The interactions between cells conventionally belongs to the

domain of the immune network theory rather than clonal selection, although cellular in-
teractions represent a fundamental concern of clonal selection as an adaptive strategy and
an immunological theory. Specifically, the cells interaction when competing for selection
and ultimately for limited resources in the repertoire. In the immune system, lymphocytes
compete in a spatial environment, and varied immune cell types interact with each other
to coordinate a response such as in the case of the two-signal activation of B-lymphocytes
by antigen and helper T-cells. Cell interactions add an additional level of complexity of
acquired information and behaviour to the adaptive strategy beyond raw improvement of
structures. This approach of elaborating clonal selection with so-called ‘more detailed’
immunological inspired models aligns with the general approach advocated in the field
(Section 2.6). Specifically, the added detail provided by cellular interactions was advo-
cated by Timmis, et al. in the context of CLONALG, where it was suggested that “whilst
[CLONALG is] intuitively appealing, [it] lacks any notion of interaction of B-cells with
T-cells, MHC, or cytokines.” [389].
Antigenic Environment The explicit phrasing of clonal selection as an adaptive sys-

tem highlighted the distinct lack of a well defined environment in which the system is
situated and interacts with. Antigen provide the catalyst for adaptation in clonal selec-
tion, and more generally the extent of endogenous and exogenous antigen the system may
be exposed to over its lifetime of usefulness defines the scope of information with which
the system must be concerned. The classical primary and secondary response to anti-
gen highlighting the improvement capability of the clonal selection process highlights the
exposure-centric nature of this environment (environmental trigger and system response).
Without an antigenic environment, the information processing qualities of clonal selection
have little meaning, highlighting the importance in elaborating this fundamental aspect
of the adaptive strategy.
73
3.6.2 Toward Distributed Clonal Selection
The discrete population basis of the clonal selection strategy and trigger-based adaptation
provide an inherently parallel process suitable for concurrent implementation. This was
shown to have been highlighted and demonstrated explicitly with AIRS for classification,
although parallisation alone fails to capitalise on the promise of distributed realisations
of the underlying adaptive strategy. Toward this end, the structure and function of the
broader immune system provide motivating metaphors for extending and constraining
the clonal selection strategy for decentralised adaptation in a distributed environment.
Specifically, two broader perspectives of clonal selection are proposed as paths towards
such approaches, as follows:
Host of Tissues A clone of lymphocytes possess the same receptors and are therefore
suited to the same antigen, where a repertoire (pool or population) of cells defines an
aggregation of such groups. In an organism (such as a mammal), this repertoire is spa-
tially distributed throughout the tissues of the host, where diverse groups of lymphocytes
provide antigen surveillance by migrating around the body via the vascular system, whilst
other groups position themselves in tissues awaiting the arrival of pathogen. As such,
the structure and function of the immune organs, tissues, and their connectivity through-
out the general mammal organism provide a template for organising and constraining a
tissue class of clonal selection approaches. Lymphatic tissues and lymphocyte migration
has been proposed as a model for distributed Negative selection specifically in Hofmeyr’s
ARTIS (specifically as read-only broadcast of negative detectors, see Section 2.5.2) and
more generally as an inspiration for decentralised Artificial Immune Systems by Dasgupta
[106], although to the authors knowledge such approaches have not been implemented or
investigated in the context of distributed adaptive clonal selection.
Population of Hosts A natural extension of the tissue perspective, is to scale the

adaptive concerns of clonal selection to that of a population of hosts with interacting ac-
quired immune systems. Although conceptually, a host perspective is removed from the
concerns of receptor interactions with antigenic determinants, the general principles of
population immunology depend on such consistent low-level adaptation. Specifically, the
a population of hosts is concerned with immunisation of all types such as vaccinations,
epidemics, and maternal immunity. Unlike tissues of a body, individual hosts in a pop-
ulation are redundant in the same manner as individual cells are, with life cycles, and
evolution. The consideration of genetic evolution combined with ontogenetic evolution of
an immune gene library using genetic and clonal selection was considered by Hightower,
et al. [208, 207], although there has been little follow-up work, specific with regard to the
role and relationship of clonal selection in such systems.
74
3.7 Chapter Summary
The detailed review of the clonal selection theory in Section 3.2 provided the context to
assess the motivations and expected behaviours embodied in the state of the field of clonal
selection algorithms. This insight aided in outlining the three proposed criticisms of the
state of clonal selection algorithms in Section 3.3.3. The first regarding the placement
and assessment of the field in a broader context of related Computational Intelligence
approaches was addressed in Section 3.4 strengthening the relationship with Evolutionary
Computation, and outlining the important relationships with Hill Climbing algorithms,
Instance-Based Learning, and Competitive Learning. The second limitation suggested
that the lack of explicating investigating clonal selection as an adaptive strategy which
was addressed in Section 3.5 highlighting the need for: (1) a clearly defined environment
to complement the adaptive response of the strategy, (2) the importance of interactions
between discrete components of the system, and (3) the focus on emergent behaviours.
Section 3.6 addressed the third and final limitation of the cellular focus of Clonal Se-
lection Algorithms, and in so doing outlined a research agenda for both (1) elaborating
upon the existing cellular clonal selection algorithms, and (2) provided a basis for inves-
tigating distributed clonal selection algorithms. This agenda occupies the remainder of
this dissertation. Specifically, Chapter 4 investigates the Cellular Clonal Selection adap-
tive strategy including both the definition of an Antigenic Exposure Paradigm, as well as
algorithms that incorporate additional forms of intra-repertoire interactions. Chapter 5
defines and investigates the Tissue Clonal Selection Paradigm inspired by the structure
of the lymphatic system and behaviour of migrating lymphocytes, and Chapter 6 defines
and investigates the Host Clonal Selection Paradigm inspired by the immunisation and
evolution of the acquired immune system.
75
Chapter 4
Cellular Clonal Selection

This chapter consolidates the findings of the previous chapter and proposes the first of
three constrained perspectives of the clonal selection strategy. Specifically, this chapter
is concerned with clonal selection as constrained by the antigenic, cellular, and molecu-
lar interactions of cells within a repertoire called the Cellular Clonal Selection Paradigm.
Importantly the perspective of clonal selection as an adaptive strategy considered in this
chapter both encapsulates the state of the art in clonal selection algorithms, and pro-
vides a bedrock of understanding for the integrated hierarchy of perspectives considered
in the remainder of this dissertation. Section 4.2 considers the cellular perspective from
an abstract perspective, dividing the concerns of the paradigm into that of system and en-
vironment, clearly delineating the responsibility of the adaptive strategy and the problem
domain that provides the context for adaptation. The paradigm is realised in Section 4.3
with regard to specific algorithm and problem definitions and empirical measures for as-
sessing the composition and capability of a given repertoire governed by clonal selection.
The behavioural expectations of an archetype of the strategy are assessed and confirmed
in a series of three empirical studies in Section 4.4. These findings provide a foundation
from which three extensions of the strategy are proposed and investigated using empirical
study to confirm behaviour expectations, including (1) a spatial context for the repertoire
in Section 4.5, (2) mediation of response provided by the interaction different cell casts
in Section 4.6, and (3) the promotion of higher-order cross-exposure structures in the
repertoire via network inspired interactions in Section 4.7.
4.2 Abstract Cellular Paradigm

The Cellular Clonal Selection Paradigm is a rephrasing and restricting of the existing field
of algorithms that realise the computational attributes of the clonal selection theory of
acquired immunity. This section considers the abstract concerns of the paradigm focusing
on (1) a quintessential cellular clonal selection adaptive strategy, (2) a system-environment
phrasing of the strategies interaction with a given domain called the antigenic exposure
paradigm, and (3) a generic problem domain for conceptualising and investigating cellular
76
clonal selection algorithms and their elaborations.
4.2.1 Quintessential Clonal Selection

This section considers an archetype of clonal selection adaptive strategy as the consolida-
tion of the related immunological theory, computational principles and approaches, and
adaptive systems theory from Chapter 3. Clonal selection is constrained to the concerns
of a repertoire of cells and antigen called the Cellular Clonal Selection Paradigm that
focuses on the separation and interaction of system-environment, and clonal selection as
an adaptive strategy.
Constrained Scope
The clonal selection theory is defined at the cell, antibody, and antigen level of immunobi-
ology and immunochemistry to describe what in complex system theory are described as
emergent effects of host immunity. Therefore one may constrain the investigation of the
information processing concerns of clonal selection at this level and consider the emergent
immunity effects within the context of a repertoire of cells with regard to antigen. In this
work, the study of clonal selection under this constraint is referred to as Cellular Clonal Se-
lection, distinct from the consideration of clonal selection under multiple of such repertoires
in Tissue Clonal Selection (Chapter 5), and the consideration of clonal selection across
multiple whole immune systems in Host Clonal Selection (Chapter 6). Cellular Clonal Se-
lection is primarily concerned with the management of a repertoire of discrete cells and its
interaction with antigen. The scope of control for this management is the repertoire in its
entirety. This suggests that the repertoire of cells in cellular clonal selection is centrally
governed with access to complete information regarding the information content of the
repertoire. Therefore the scope of cellular clonal selection algorithms considered in the
reviewed taxonomy (Section 3.3) may be considered Cellular Clonal Selection Algorithms.
As discussed in that section, this level of clonal selection algorithms are strongly related
in mechanism and effect to the Genetic Algorithm (GA) and the Mutation Hill Climbing
Algorithm (MHCA), as well as the related fields of Lazy and Competitive Learning.
System and Environment
A critical perspective on computational models of clonal selection highlighted by the adap-

tive systems theory in the previous chapter was the need for a clear delineation between
system and the environment that the system interacts with. The system is defined as be-
ing comprised of a cellular repertoire managed by clonal selection and related information
processing mechanisms. The environment is defined as the scope of endogenous antigen
and exogenous pathogen to which the system may be exposed. This delineation is easily
related back to the existing state of clonal selection algorithm research as a the separation
between clonal selection algorithm and problem domain to which the algorithm is applied.
The information content of the system and the environment may therefore be considered
in the context of the shape-space and affinity landscape geometric paradigms from the-
oretical immunology as sets of discrete cells and antigen patterns that may be assessed
77
(specificity and antigenicity) against each other as a cost surface. The antigenic environ-
ment defines the scope of the information a system may acquire and apply (immunity),
the full importance to clonal selection of which is considered in the Antigenic Exposure
Paradigm in Section 4.2.2.
Clonal Selection as Strategy
The management of a cellular repertoire in the context of an antigenic environment is the

responsibility of the clonal selection adaptive strategy. The focus of such a strategy is
the acquisition and application of information (immunity) from an unknown information
(antigenic) environment. The selectionist, adaptationist, and satisficing lens of the clonal
selection strategy considered in the previous chapter highlighted the following computa-
tional principles of the theory:
1. Information is acquired through cumulative generate-and-test of pre-committed dis-

crete structures.
2. Resource allocation is competitive, based on (in some way) the differential ‘goodness
of fit’.
3. The continued acquisition and application of information is based on the premise

that future antigenic exposures are much like past exposures.
4. Adaptation continues for as long as relative improvements can be made under all of
the constraints of the strategy.
5. The emergent effects of the bottom-up exposure-wise interactions may be assessed in

the holistic composition and capabilities of the information acquired in the cellular
repertoire.
The implementation of such a strategy was shown in Section 3.3 to be commonly

realised as management of a repertoire of binary strings under assessment of an oracle
cost function resulting in duplication and mutation of low cost (high affinity) strings. The
generality of such a realisation may be constrained by the principles or axioms governing
cellular interactions with antigen from Section 3.2, as follows:
• Assessment: Affinity is an ordinal scoring of the ‘goodness of fit’ of a structure

provided by an oracle.
• Selection Competition: Competition is promoted between cells specialised for the

same antigen, and avoided between cells specialised for different antigen.
• Selection: A small and specific founding set of cells are selected to found a clonal
response (oligoclonal).
• Cloning: The number of clones created from an exposure is larger than the founding
set of cells selected for the clonal response.
78
• Mutation: The point-wise rate of mutation is higher than that of normal mitosis
although is fixed.
• Integration Competition: The cellular repertoire is large although finite, promoting

competition between cells for limited positions.
• Integration: Although many clones are created in a clonal response, few are retained
in the repertoire.
The listed computational principles may be considered an extension and elaboration

on the affinity maturation principle of de Castro and Timmis, and the clonal selection
principle of Cutello and Nicosia in Section 3.3.1. These principles are the interpreted
heuristics from reviewed immunological theory use in this work and are not intended
to provide laws for the realisation of cellular clonal selection algorithms. Specifically,
differential resource allocation can be realised via a variety of specialised mechanisms
including: assessment, selection, cloning, mutation, and integration1 . These principles
are used to assess and configure cellular clonal selection algorithms in this work, and are
augmented with knowledge from the study of the related GA’s and MHCA’s. Specifically
from Section 3.4.2: (1) the acceptance of structures with an improved or equal assessment
scoring to permit the acceptance of varied structures with neutral effect on assessment to
cross flat spots in the affinity landscape, and (2) the use of a 1
L fixed mutation rate as an
approximation of the optimal mutation rate of bit strings on linear binary cost functions.
4.2.2 Antigenic Exposure Paradigm

This section proposes an artificial conceptual model of a generalised natural antigenic
environment. This model is comprised of a set of external triggers that stimulate an
internal activation and response by an acquired immune system. This conceptualisation is
called an antigenic exposure, and provides a perspective on the adaptation-based learning
and memory qualities of the acquired immune system, and a standardised framework for
mapping problem domains onto the cellular clonal selection paradigm.
The immune system is embedded or situated in an antigenic environment, to which it
(in general) passively responds. Therefore, a system has qualities of an intelligent agent
(it is situated, intelligent, and acts autonomously), although does not act upon the envi-
ronment, unlike a reinforcement learning system. A consideration of antigen other than
superficial mapping of a given problem domain is largely ignored in the investigation of
Artificial Immune Systems in general, as the investigatory concerns lie predominantly with
the information processing qualities of the system. Antigen may be exogenous or endoge-
nous in origin and may be benign or actively adversarial. The environment encompasses
the scope of the information that may be acquired by the system, dictating what to learn
and when to learn it. Three antigenic exposure regimes and resultant models are pro-
posed, as follows: (1) single exposure, (2) multiple exposures, and (3) multiple antigen
1
As evidence, one may consider the variety of such mechanisms used for differential resource allocation
in the related field of evolutionary computation.
79
with multiple exposures, that highlight the spatial-temporal consistency concerns of clonal
selection at the cellular level.
Single Exposure Model
An antigenic exposure is an event that involves the arrival of antigens (information of ex-
ternal origin) to an immune system. The arriving antigen are identified (typically forced
as a requirement of the domain) by the system stimulating a proportionate immune re-
sponse. If correctly identified, the external stimulation results in an internal activation
of some lymphocytes in the immune systems repertoire. The response involves raising a
clone of lymphocytes to address the stimulation with regard to the magnitude and speci-
ficity of the response. The system is concerned with raising a good response quickly. The
principle of an antigen is that it may cause harm to the host in the form of tissue dam-
age promoting the immediacy of the response. The system must raise a clone that can
address the magnitude of the exposure, therefore the effectiveness of the response should
be proportional to the virulence of the antigen, and the response size (clonal expansion)
should be proportional to the number of arriving antigen (amplitude or dose).
Environment
Immune System
Antigenic arrival
Response to stimulation
(internal origin)
Stimulation of the system

(external origin)
Figure 4.1: Depiction of external stimulation resulting in internal activation.
This exposure virulence is a property of the antigen thus is not likely to change for
a given immune system unless the antigen itself changes. The virulence may define the
penalty for not identifying a antigen, as well as the amount of resultant specificity de-
sired or required in a clonal selection and expansion response. This concern (virulence-
specificity) is difficult to conceptualise as it defines the relative amount of refinement-based
learning required of the system for a given antigen. The concern of the size of an anti-
gen dose may be conceptualised as the amount or exposure amplitude of information an
antigen represents. The number of simultaneous antigen that arrive to the system in a
given exposure defines the amount of work the system may have to support. The num-
ber of antigenic molecules that arrive in an exposure may be more than the number of
lymphocytes in the repertoire. The result is that selection events may be more intense,
resulting in the delay of some interaction until after the initial cells have proliferated or
additional resources have been recruited. This exposure amplitude mismatch scenario has
two speculated implications: (1) Antigen may not be neutralised as quickly if it arrives
in unexpectedly large quantities, that would result in the antigen arrival being considered
a resources which must be degraded (neutralised over time), and (2) The increased se-
80
lective pressure from the size mismatch may result in the rapid synthesis of a response
large enough to address the exposure. This suggests dynamic and variable response size
capabilities inherent in the clonal selection and expansion response.
One may consider the effects of varied exposure virulence and amplitude on an immune
system. Both properties affect the amount of work required in an immune response,
although in different dimensions. Virulence may determine the desired specificity of a
response such that high damage-causing pathogens are neutralised effectively and low
damage-causing pathogens are neutralised with less specificity. The amplitude of the
exposure defines the number of cells and the size of the clonal response. Therefore, the
immune system must trade-off response strategies to the various anticipated virulence and
amplitude combinations. For a single exposure event, an immune system is only concerned
with quelling the antigen with a response. The scope of the concerns of an immune system
for a single exposure is the specificity of the lymphocytes to the antigen and synthesising
the cells in sufficient numbers. Initial antigen-identification (to trigger a response) may
be assumed. Further, all resources of the repertoire may be allocated to the response
for the exposure, as there are no future exposures. Iterations within a system under
these conditions are concerned with adaptation in the context of a fixed and known single
exposure event.
Multiple Exposure Model
A single exposure-response event may be abstracted to multiple antigen exposure-response

events. A principle property of a multiple exposure regime is the temporal dimension it
adds to such events. Exposure frequency facilitates conceptualising multiple exposures
as a series of atomic and quasi-independent single events, where the repetitiveness of
the arrival of a given antigen is defined as its frequency. The exposure frequency may
be measured as the pattern of exposure and no-exposure events in discrete time over an
interval. The pattern may be defined as a probabilistic function, random, uniform, or may
be a regular deterministic exposure function. The concept of a non-stimulation fostered
by no-exposure events suggests a downtime when the immune system is not stimulated
to respond. Such a non-activity event may also occur for an exposure event that is not
identified by the exposed system. This inactive time is referred to as antigenic exposure
downtime. The activity of a system may be considered trigger-based in which learning
and maintenance processes (such as cell aging and removal) only occur when the system
is exposed to pathogen. During non-triggered periods, the system may enter a period
of stasis awaiting subsequent triggers (the idle repertoire awaiting exposure model was
criticised in the Immune Network Theory). An alternative model is that the system
persists (always on), continuing to execute such maintenance processes in the absence
of stimulation. The two models highlight the potential for the decoupling the system
processes of repertoire maintenance and repertoire stimulation. Such a decoupled system
would require an effective memory system such that acquired immunity (information) was
not lost by the high-turnover of lymphocytes in the repertoire. The lack of adequate
memory for this system type in an extended period of non-stimulation may cause the
81
repertoire to devolve to a naı̈ve (random) state.
Variations in the exposure frequency control when the system can and must respond.
Variations in the exposure amplitude control the size (scope or how much) of the response
that a system must synthesise. Variations in exposure virulence and amplitude imply
variations in a systems response strategy in terms of response specificity and response
quantity. A similar relationship exists with multiple exposures with the addition of expo-
sure frequency and the systems memory of the response. There is an inversely proportional
relationship between immunological memory and exposure frequency (with constant vir-
ulence) such that when the exposure frequency is low, the memory must be long lived.
When the exposure frequency is high, memory is short lived, and in fact may be satisfied
with the quantity and specificity effects achieved from the single exposure without mem-
ory. One may consider the effect of combinations of high and low frequency exposures with
high and low amplitude pathogen arrival. From a system perspective, the frequency and
amplitude of exposures may be compressed (multiplicative relationship) into the amount
of work (response effort) required for the interval of time. This compression of exposure
frequency and amplitude may be referred to as the exposure regime of an antigen to a
system through time.
Figure 4.2: Example exposure regimes demonstrating some archetypical pathogen viru-
lence behaviours.
The relationship between a pathogen exposure regime and an immune systems re-
sponse effort is interesting. For a single exposure event, virulence defined the amount of
tissue damage an antigen may cause on a host, therefore it may provide a conception of
the amount of specificity refinement (effort) required of an immune system for the expo-
sure. An exposure regime provides a way of crisply defining antigenic virulence (distinct
from single exposure virulence) as a function of exposures through time. The amount of
specificity refinement by the system for the antigen (response effort) is proportional to the
exposure frequency and exposure amplitude (exposure regime). Antigenic virulence (and
exposure regimes) may be expressed as an exposure amplitude graph against time. From
this level of abstraction, one may give example exposure regimes as archetypical pathogen
virulence strategies with which to subject an immune system (see Figure 4.2). In addition
82
to static antigenic virulence, it is possible for virulence to vary through time. The variance
in virulence may be system-dependant such that virulence changes in response to the inter-
nal actions of the immune system such as in the case of an adversarial antigen (pathogen).
For a multiple exposure event, the system is primarily concerned with devising a response
to each exposure and ultimately using past exposures to anticipate the resources needed
for future exposures. This exposure model supersedes the single exposure event, as in
the case of a single exposure, all resources of the repertoire are allocated to the specific
concerns of the antigen (as there are no other antigens). The learning that the system
may achieve within the context of a single exposure may be limited, as a broader exposure
regime may define how often the exposures arrive to the system. The exposure regime
may impose a quantity and specificity capability constraints for a single exposure, meaning
that a finite amount of specificity learning and/or quantity learning may be achieved for
a single exposure event. The system must learn to anticipate (1) the required specificity
of future exposures, and (2) the required quantity of future exposures, both in the face of
non-exposure events.
Start Time
Exposure
No Is Exposure
Event?
Generate Yes
Response Expose
System
Yes Is antigen
remaining? Yes Is Antigen
Still Alive?
No No
End Exposure End Time
(a) Single Exposure Model. (b) Multiple Exposure Model.
Figure 4.3: Flow diagram of the single and multiple exposure models.
The exposure regime (frequency and amplitude) define the antigenic virulence and the
systems required specificity. Repeated application of blind clonal selection facilitates im-
provement of a repertoire specificity for the antigen, and for a system that enters stasis
during periods of non-stimulation (information acquired by the repertoire does not de-
grade), the density of information (clonal convergence and/or clonal densities) may define
the quantity or scope of information required for the regime. Therefore, a repeated sin-
gle selection on a relatively naı̈ve repertoire fosters a model that matches the virulence
of the antigenic exposure regime. For a repertoire with decoupled maintenance of cells
and triggered adaptation, an explicit memory mechanism is required that retains acquired
information proportional to the frequency of its exposure and complexity. For lower-
83
complexity information, a more efficient strategy may be to generalise or re-acquired as
need.
Multiple Antigen Model
The natural extension of multiple antigen exposures is to subject an immune system to

multiple concurrent exposure regimes. Each regime is expressed by a distinct antigenic
type or variety. Antigen types differ in their composition (surface features) such that the
immune system has to acquire different immunity characteristics for each. In addition
to the varied immunity characteristics, each antigen type has its own exposure regime,
thus in the context of an immune system, the regime elicits a distinct response effort.
The aggregation of multiple antigen types is an aggregation of multiple antigen exposure
regimes to a given immune system that may collectively be referred to as the exposure or
antigenic environment. In responding to one given antigen, the repertoire may acquire a
level of immunity (specificity) to another different and distinct pathogen type. This effect
is called cross-reactivity of the immune response and may be conceptualised as reuse and
generalisation of acquired information. One may consider an immune system to be exposed
to the multiple different exposure regimes shown in Figure 4.2. In this example, the system
is exposed to all four different antigen with varied amplitudes at the discrete time of t = 11.
This may be referred to a concurrent antigen exposure. Exposure concurrency requires
that the repertoire is (1) large enough (with regard to the quantity of lymphocytes), and
(2) diverse enough (with regard to lymphocyte specificity) to address multiple pathogen
types at the same time. It is expected that the efficiency of the system will be reduced in
such a situation, as resources are allocated proportional to the relative virulence of each
antigen type. This effect highlights the requirement of the system to be able to not only
respond to multiple antigen types at the same time, but to integrate the results of the
concurrent responses (expanded clones and shifted the repertoire specificities).
Unlike the previous two models that were able to allocate all the resources of the reper-
toire to a single exposure of antigen information, multiple antigens requires a proportional
allocation of resources to each antigen and its exposure regime. The proportionality of the
response and allocation of repertoire resources is defined by the relative antigenic viru-
lence to other antigen in its environment. The previous concerns of optimising specificity,
quantity, and the anticipation of these values may be reduced to the systems adapta-
tion of information densities toward exposure regimes. The density is representative of
the learned (aggregated from experience) cell quantity (clone size) and specificity in the
repertoire which facilitates anticipatory responses. Each cell has an affinity response sur-
face (affinity landscape) for each pathogen it has been (and may be) exposed to. The
clonal densities are acquired from the antigenic virulence which is defined as the ampli-
tude and severity of exposures (amount of learning opportunity offered) to the system
through time. This model provides the pinnacle in environments for clonal selection at
the cellular level and a complement for the quintessential clonal selection framework.
84
Exposure Models and Control
The environment and the system may be considered two interdependent and required com-
ponents of which neither is meaningful in isolation. In modelling these two components,
one may consider the variable (configurable) aspects of each of which may or may not be
within the control of a given simulation. An intuitive example is to consider the environ-
ment (and therefore all aspects of antigen exposure) as outside of the scope of control, and
all aspects of the system as inside of the scope of control. In this example the system must
do all it can to cope with the specific properties of the exposures it is subjected to by the
environment. The system must operate with finite resource, thus constraints are imposed
on the control over aspects the system. This example may be extended further to consider
that the evolution of antigen (pathogen) is influenced by the existence and evolution of
the host system and that this relationship is reciprocal (co-evolution). Therefore, the sys-
tem has influence (implicit influence via evolution) over aspects of antigen exposure over
longer periods of time, and the antigen has influence (implicit influence via evolution) over
aspects of the systems response over the same longer periods of time. This example may
be generalised such that aspects of control are variable across both the environment and
the system.
Control of
Environment
Open system
with general
applicability
Control of
System
Figure 4.4: Depiction of the trade-off between control of the environment and control of
the system.
In modelling environment and system and their interactions, control may be exhibited
over all aspects of both sides of the interaction. It is important to consider the influence
for control or lack there of (constraints on control) on a system and its environment as
it provides insight into the mapping of domain properties. A constraint may be consid-
ered the prior selection of a configuration parameter of a system or environment. The
fixing of a configuration parameter in the environment influences the suitability of other
unconstrained (controlled) parameters in the interaction, and the influence is likely to
be complementary (such that fixing an aspect of the environment strongly influences the
suitability of related parameters of the system and vice versa). This point highlights the
symmetry of constraint between systems and environments, and the need to specialise
a given system for the constraints of a given environment in which it is to be deployed
(mapped problem domain).
85
4.2.3 Colour-Space Domain
In his book detailing the Cognitive Theory of acquired immunity, Cohen used an example
of cellular degeneracy in the retina as an analogy for the degeneracy of cell-bound receptors
in the immune system [81]. Cohen described how individual cells in the retina have limited
capability, responding to constrained wavelengths of light, although from these specific
degenerate detectors higher-level information such as a coherent image can be perceived
through aggregation. The Self-Organizing Map (SOM) is a competitive and unsupervised
learning algorithm that is renowned for its feature extraction and dimensionality reduction
capabilities (see Section 3.4.4). In his doctoral work, Honkela used a colour domain (red-
green-blue or RGB), and a dataset of known colours in that domain [220] (page 14-18).
The learning and ordering of the domain of RGB colours based on the dataset was used
as an intuitive example of the topological preserving properties of the SOM algorithm.
Although Honkela may not have been the first to employ this example domain to the SOM,
it has become a canonical example domain.2 Inspired by the simplicity and suitability
of the colour domain to qualitatively (visually), and quantifiably (numerical measures)
demonstrate the properties of the SOM, this section defines a colour-space domain from
which a suite of test problem instances may be drawn to investigate clonal selection.
Inspired by Cohen’s analogy to the retina, colour-space is named for the relationship to
the shape-space formalism, and is supported by proven use of the same principle for a
related Competitive Learning paradigm.
Domain Definition
The domain does not provide a framework for investigating the so-called colour space
theory, which refers to the study of colour models used in display and printer systems. Nor
is the intention of the defined test domain for the study of colour theory, the investigation
of colour segregation, colour recognition, palette compression, or colour clustering (or any
other practical colour-based application domain). Given what the intentions are not, this
section outlines a series of general goals for designing a test domain for adaptive systems,
followed by the definition of a simple colour-space domain that meets these goals. The
ambition is to contrive a domain from which trivial problem instances can be drawn to
preliminary assess specific properties of adaptive models. These models may be engaged
in processes that equate to pattern recognition and optimisation, and may involved the
evaluation of broad characteristics such as learning, memory, and adaptation. See Table 4.1
for a list of the general design goals (requirements) for a problem domain suitable for this
purpose.
The colour-space domain formalism meets the list of goals proposed in Table 4.1. The
formalism is comprised of a series of terms to describe the characteristics of a colour
space domain. The characteristics identified include the environment: the cardinality, the
representation, and the mapping function (see Table 4.2 for their definition and summary).
The environment (E) is an n-dimensional hypercube (for visualisation purposes per-
2
There are many applet and tutorial web sites on the Internet that use this example to demonstrate
the SOM, the listing of which is not necessary.
86
G# Goal
G1 The domain must provide a numerical and/or combinatorial basis that may be manipu-
lated meaningfully by operators.
G2 The response surface must be correlated such that localised regions in the response surface
have a gradient.
G3 The domain must be suitable for optimisation tasks.
G4 The domain must be suitable for pattern recognition tasks.
G5 The domain must be parameterised such that the relative complexity may be meaning-
fully adjusted.
G6 The domain must be easily and meaningfully visualised both online (dynamically) and
off-line (end of run).
G7 The domain must be easy to understand, easy to implement, and easy to analyse.
G8 The properties of the domain (e.g. the response surface or visualisation) should facilitate
rather than mask model behaviours being investigated.
Table 4.1: List of goals for a colour-space domain.
haps limited to 1-3 dimensions), where each dimension represents a different colour axis.
The intention is that a distinct coordinate in a given environment (colour space) repre-
sent a distinct colour. For example, a monochrome (one-colour) environment would be
implemented as a one-dimensional space (line) perhaps a gradient of white to black. An
RGB colour space would be implemented as a three-dimensional (cube) environment. An
environment is a volume of feasible discrete points, and the cardinality (C) defines the
number of points in that volume. The granularity of each dimension in the environment
also determines the number of colours, thus cardinality may be thought of as the palette
of the environment. Further, the palette analogy may be exploited further in defining
standard cardinalities, such as those common palettes that are used in computer display
systems (for example, 8-bit colour (28 ) with a pallet of 256 colours). The coordinate-based
representation (R) may be used directly by an adaptive system. Alternatively a symbolic
or sub-symbolic representation may be used (such as bit-strings), which require a mapping
function (M ) to transform representation into a given environments coordinate system.
The task of an adaptive system is called an intention (I), examples may include the opti-
misation of a randomly initialised system to a set of pre-selected points, or the adaptation
of a system to a defined region within the environment.
One may measure the Euclidean distance between points in the colour space as a
quantitative measure of difference. Distances measures may not be limited to Euclidean
distance, for example Manhattan distance (and many others, see [261] pages 17-29) may
be used in the colour space. Distance measures may be augmented with coordinate radii
such that a point of interest in the space may represent a collective of points in its vicinity.
Such regions may be defined as hyper-cubes or spheres, and may have a distance-based
falloff such as linear, Gaussian or exponential. These distances and coordinate neighbour-
hoods are also qualitatively meaningful. Following a line (for example in a monochrome or
RGB) in colour space shows a transition through intermediate colour coordinates (colours).
These intermediate colours are quantifiably distinct (numerical coordinates), and qualita-
87
Term Name Summary
E Environment The search space (colour space) which defines the scope of feasible
coordinates defined by its dimensionality (1-n), a boundary (limits
of each dimension), and cardinality (C).
R Representation An encoding used and manipulated by adaptive models. It may
be binary, integer, real, and may or may not directly represent a
coordinate in the colour space (E).
M Mapping Function Converts a given representation (R) into a coordinate in the colour
space (E). If the representation used matches the coordinate sys-
tem of the environment, then the mapping function returns an
unchanged coordinate.
C Cardinality The number of discrete points in the domain, and ideally (given
display capabilities) the number of colours available for visualisa-
tion.
I Intention The goal of a system within the domain, a task or action it is to
perform or achieve. This is a cover-all term for something to do
in a defined colour-space.
Table 4.2: Summary of terms for the colour space domain formalism.
tively different (assuming the colour transition can be detected by the human eye). This
correlation is also useful for visualisation: agents representing coordinates in the colour
space may be implemented, which may permit qualitative measures and observations of
system behaviour.
Illustrative Examples
This section proposes an optimisation, pattern recognition, and classification examples

of Colour Space Problems that adaptive systems that may be employed or extended to
investigate adaptive models.
• Optimisation: Involving the pre-selection of a coordinate in the space that is with-

held from the adaptive system, and defining a cost function as a distance measure to
the given coordinate. For example if a space was defined as an 8-bit one-dimensional
domain (0-255), a goal coordinate of 127, and an Euclidean distance measure em-
ployed as a cost function, then the shape of the response surface would be a triangle
with the apex at the selected coordinate.
• Pattern Recognition: The optimisation domain may be generalised with a set of

goal coordinates (Colour Space Patterns) to which a given system is exposed. The
intention of the system for this problem is to optimise (optimally match) the set of
patterns, where the order of exposure of the patterns to the system is also withheld
from the system.
• Classification: A classification task may be defined as an extension of the pattern

recognition problem with the addition of categorical information. A geometry can
be defined within the colour space and class labels can be assigned to the resulting
concave regions. In addition to distance-based feedback, the system may be provided
88
with corrective categorical based feedback, relaxing the optimisation of the goal
coordinates to that of accuracy of predicting pattern classes.
Criticisms and Limitations
This section addresses the important consideration’s of the limitations of the formulated
problem domain, as follows:
1. Triviality: The domain and resultant problem definitions are primitive. The problem
instances are likely easily solved rapidly by standard deterministic techniques, and
even by naı̈ve approaches such as enumeration and random search.
2. Transferability: The results achieved on one or a set of derived problem instances

are very likely to be not transferable to problem domains of interest (difficult, real-
world). Conclusions regarding model performance are limited to the instances on
which they are tested, and perhaps related trivial domains.
3. Difficulty: Although superficiality they appear trivial, there is no innate sense of the
absolute or relative difficulty of the derived problem definitions.
4. Visibility: For high-cardinality domains (such as those at or above 24-bit or 32-bit

RGB) the human eye cannot detect colour differences between coordinates within
close proximity.
5. Novelty: The colour space domain and derived problem instances are not novel,
they are generalisations and simplifications of existent test domains and benchmark
problems.
It is prudent not to rebut these criticisms, but to acknowledge them as limitations

(designed or otherwise) of the domain and framework that it provides. The domain and
resultant problem instances are intended to be trivial (easy to solve) such that the attention
remain on the model under study. The performance of a given model is not intended to be
transferable, what is expected to be transferable are the general functional behaviours that
the domain assists in isolating and assessing. The relative difficulty of problem instances
may be determined by baseline strategies. Comparable relative and absolute difficulty may
be defined using the tools of probability, statistical mechanics, and additional (suitable)
mathematical apparatus. The visual distinctiveness of close-proximity coordinates in high-
cardinality spaces may be exaggerated where appropriate. For example, a dynamically
normalised colour scale may be allocated to coordinates of interest. Finally, novelty is
not claimed, rather the domain is a consistent reformulation of classical optimisation and
pattern recognition benchmark domains.
4.3 Realised Cellular Paradigm

This section provides a realisation of the concerns of the abstract Cellular Clonal Selection
Paradigm presented in Section 4.2. This includes general definitions of a standardised
89
problem domain in the Antigenic Exposure Problem as a specialisation in colour space,
and the Cellular Clonal Selection Algorithm that provides a basis for investigation into
elaborated cell interaction schemes. A series of cell-based measures are presented for
assessing cellular algorithms on instances of Antigen Exposure Problems that provide
quantitative indicators regarding the concerns of information acquisition and application.
Collectively, this section provides a basis for the implementation and exploratory empirical
investigation into the Cellular Clonal Selection Paradigm, as well as a bridge between the
biology (Section 3.2) existing approaches (Section 3.3), and the abstract clonal selection
and antigenic exposure models (Section 4.2).
4.3.1 Antigenic Exposures

This section considers the realisation of the specialisations of the exposure paradigm out-
lined in Section 4.2.2 as a basis for empirical investigation and ultimate application of
cellular models to problem domains. This realisation includes the definition of general
antigenic exposure problem with exposure regimes, and a specialisation to the colour
space domain.
Antigenic Exposure Problem
For the purposes of discussing exposure, a cellular clonal selection algorithm may be
reduced to its information content, distinct from the specific clonal selection processes
operating upon that information. Specifically, a cellular algorithm may be considered
a population or repertoire (T , that stands for ‘tissue’ which will be made apparent in
proceeding chapter) of discrete cells (C), as follows: T = {C1 , C2 , C3 . . . , Cn }. Therefore
an Antigenic Exposure Problem (AEP) is defined as the discrete interaction with a T ,
with a set (I, which stands for ‘infection’) of antigen (A) at the same level of abstraction,
as follows I = {A1 , A2 , A3 . . . , An }. An antigen is comprised of sub-antigenic information
referred to as determinants (D), such that A = {D1 , D2 , D3 . . . , Dn }.
Algorithm 4.1: Antigenic Exposure Problem (AEP).

Input: T, Ndeterminants Nantigen
Output: Trs
1 I ← 0;
2 for i←0 to Nantigen do
3 Ai ← CreateRandomAntigen(Ndeterminants );
4 I ← Ai ;
5 end
6 Trs ← 0;
8 Trs ← Exposure(I, T);
9 end
10 return Trs ;
Algorithm 4.1 defines a simple antigenic exposure problem, with symmetrical informa-
tion content for interaction with a given cellular system, where Ndeterminants and Nantigen
90
define the number of determinants per antigen and total antigen within the scope of the
AEP. The algorithm clearly shows the AEP’s in control of the interactions with a given T ,
where discrete interactions are mediated via an Exposure(I, T ) operation. The algorithm
definition also highlights that from an exposure perspective, the AEP is only concerned
with a systems aggregate responses to exposures. The tissue result set denoted as Trs may
be considered to represent a cellular algorithm’s (T ) capability to address a given antigen
exposure problem (I).
Cellular Exposure Regimes
The defined antigenic exposure problem delegates the specifics of algorithm-problem in-
teraction to the Exposure(I, T ) operation, the behaviour of which may be referred to as
a Cellular Exposure Regime (CER). For example, the information content of the problem
may be exposed to a system (1) consistently in the same order, (2) in a randomised order,
or (3) in some biased manner resulting in an asymmetric perspective of such information
content. Referring to such interactions as an exposure regimes also highlights the impor-
tant point that the cellular algorithm is considered a passive information processing system
that responds to the active exposure of antigenic stimuli. Therefore, depending on the spe-
cific problem domain, the concerns of exposure, re-exposure, and multiple exposures (from
Section 4.2.2) may or may not be within the scope of control of the system. Algorithm 4.2
provides an example of a regular and consistent exposure regime Exposure(I, T ) operation
where a system is exposed and must respond to the scope of the AEP each I exposure
called an epoch. The example also delegates the responsibility of specific cell interaction
(selection and response) to the cellular clonal selection algorithm in the Exposure(A, T )
operation. This operation encapsulates the scope of concerns of algorithms in the cellular
paradigm.
Algorithm 4.2: Exposure Function (CER) for the Antigenic Exposure Problem.
Input: I, T
Output: Trs
1 Trs ←0;
2 foreach Ai ∈ I do
3 Ci ← Exposure(Ai , T);
4 Trs ← Ci ;
5 end
6 return Trs ;
The cellular exposure regime represents the uncertainty in the information processing
faced by a cellular algorithm, including but not limited to (1) the number of antigen
(Nantigen ), (2) the number of determinants for a given antigen (Ndeterminants ), the diversity
of the information in the antigenic environment (I), and the ordering of the discrete
information exposures defined in the the exposure regime itself (CER). Finally, discrete
cellular interactions with antigen and determinants may not yield specific information,
rather a general indication. For example from an abstract point of view an Exposure(A, C)
or Exposure(D, C) may result in a relative indication of fit or suitability (affinity or
91
avidity) rather than absolute cost. This is motivated by the pre-committed, iterative, and
trial-based (selectionist) strategy reduced from the clonal selection theory.
Antigen Colour Space Problem
The antigenic exposure problem may be specialised in the colour space domain as a pattern
recognition problem called the Antigen Colour Space Problem (ACSP). The problem is
defined as a set of Colour Space Patterns (CSP) each of which is defined as a colour
comprised of three 64-bit colour components. The set of patterns are generated as a set
of random binary strings (A& = {0, 1}L , where L = (3 × 64)) and decoded using the
!
Gray Code method for each colour component (Equation 4.1, where denotes addition
modulo two) to a three dimensional colour vector A = [0, 1]3 . Cellular solutions (C)
to each antigen colour space pattern (A) are 192-bit binary strings that decoded to 3-
dimensional colour vectors using Gray Code. The affinity (avidity) function for a decoded
cell, the Exposure(A, C) is defined as the Euclidean distance between the two vectors
(Equation 4.2).
   
64−1 64−j
1 $ %
GrayCode(A&) = 64   A&(i − 1)64 + k  2j  (4.1)
2 − 1 j=0 k=1
(
) n
)$
EuclideanDistance(A, C) = * (Ai − Ci )2 (4.2)
i=1
This mapping of the colour space problem onto the antigenic exposure problem high-
lights the important point of the minimal cases of the problem, specifically: (1) minimum
determinant, and (2) minimum antigen. From an abstract perspective, a cell matches onto
one determinant, from an antigen of multiple determinants. Therefore, from a system
perspective a one-to-one relationship exists between a given cell and its cognate antigen
mediated via the specific sub-feature (antigenic determinant). A Minimum Determinant
Antigenic Exposure Problem (MDAEP) may be defined where each antigen is comprised
of a single determinant (Ndeterminants = 1, ∀A ∈ I). This case forces such a one-to-one re-
lationship at the cost of the exclusion of the cross-reactivity of cells. In the same manner,
the number of antigen in the set may be minimised (Nantigen = 1) to define a Minimum
Antigenic Exposure Problem (MAEP) where a given system T is assessed based on its
capability with a single antigen with a single determinant. This minimal case (single
or multiple exposures of a single determinant), provides a minimal context in which to
consider fundamental behaviours of cellular algorithms and their extensions.
4.3.2 Cellular Clonal Selection

This section considers the realisation of specialisations of quintessential clonal selection
outlined in Section 4.2.1 as a basis for empirical investigation and ultimate application of
cellular models to problem domains. This realisation includes the definition of a general
cellular clonal selection algorithm as a homologue to CLONALG, and a replacement based
extension.
92
Cellular Clonal Selection Algorithm
A Cellular Clonal Selection Algorithm (CCSA) is defined as the application of the compu-
tational properties of the clonal selection theory constrained by the concerns of the theory
at the cellular-level. These constraints include but are not limited to (1) the interaction
with antigen via their determinants, (2) the embodiment of information in a self-contained
repertoire of discrete cells, and (3) the cellular response to active antigenic exposures. The
extent of the taxonomy of clonal selection algorithms reviewed in Section 3.3 may be con-
sidered to reside in the cellular level (within the scope of the Cellular Clonal Selection
Paradigm). This section considers an abstraction of cellular clonal selection algorithms
in the context the concerns considered in quintessential Clonal Selection in Section 4.2.1.
The Antigen Exposure Problem defined a cellular system in terms of its information con-
tent, where the problem is the active concern and the cellular system is passive until
exposed with stimuli to which it reflexively responds. The scope of the concerns of cellular
clonal selection algorithms are the management of information within the repertoire, and
in particular the management of information under the discrete exposures of an AEP.
Algorithm 4.3: Initialisation Function for the Cellular Clonal Selection.

Input: Ncells
Output: T
1 T ← 0;
2 for i←0 to Ncells do
3 Ci ← CreateCell();
4 T ← Ci ;
5 end
6 return T;
Algorithm 4.3 defines an initialisation operation for a the Cellular Clonal Selection
Algorithm, where the CreateCell() may be specialised to create random cells with a
representation appropriate for the specific problem domain (192-bit binary strings in the
case of the ACSP). Algorithm 4.4 defines the Exposure(A, T ) operation for the CCSA that
provides a generalisation of the CLONALG, BCA, AIRS, and IA family of the reviewed
clonal selection taxonomy. Specifically, the algorithm responds to antigen exposures by
assessing the extent of the repertoire and selecting a subset (Nselected ) to comprise a
response. Each cell in the selected set creates a set of clones (Nclones ) with mutations
(Pmutation ) that are assessed against the antigen and compete with the progenitor of the
selected set for a position in the repertoire. This generalisation which strongly resembles
CLONALG, implicitly manages the memory cell set and explicitly returns a best matching
cell (GetBestM atchingCell(T )) for each exposure as the most suitable response within
the scope of the repertoire. Importantly, the CCSA provides the flexibility to assume
a variety of configurations, specialisations (such as the CSA taxonomy), as well as the
extensions considered in this chapter.
The CCSA may be reduced to a minimal configuration to provide a complement to
the Minimal Antigenic Exposure Problem called the Minimal Cellular Clonal Selection
Algorithm (MCCSA). This minimal configuration under the MAEP provides a realisation
93
Algorithm 4.4: Exposure Function for the Cellular Clonal Selection.
Input: A, T, Nselected , Nclones , Pmutation
Output: Trs
1 Trs ← 0;
2 foreach Ci ∈ T do
3 Expose(A, Ci );
4 end
5 Tselected ← Select(T, Nselected );
6 foreach Ci ∈ Tselected do
7 Tclones ← 0;
8 for i←0 to Nclones do
9 C&i ← Clone(Ci );
10 C&i ← Mutate(C&i , Pmutation );
11 Tclones ← C&i ;
12 end
13 Expose(A, Tclones );
14 T &clones ← Select(Tclones , Nselected );
15 for i←0 to Nselected do
16 Cselected ← Tselected i ;
17 Cclone ← T &clones i ;
18 if Cclone .score ≤ Cselected .score then
19 T.Remove(Cselected );
20 T.Add(Cclone );
21 end
22 end
23 end
24 Trs ← GetBestMatchingCell(T);
25 return Trs
of of the ES(1 + 1) algorithm and connection with mutation based hill climbers made in
Section 3.4. For example, ES(1 + 1) would be defined by the configuration: Ncells = 1,
Nselected = 1, Nclones = 1, and Pmutation = L.
1
Replacement Cellular Clonal Selection Algorithm
An expected effect with the Cellular Clonal Selection Algorithm is that it results in the
iterative adaptation of specialised and independent cells. This is expected (1) because it
was observed in CLONALG with regard to a niching-like effect, and (2) because CCSA
and CLONAL resemble a parallel mutation based hill climber. The Replacement Cellu-
lar Clonal Selection Algorithm (RCCSA) defined in Algorithm 4.5 is an extension of the
CCSA where all clones created for an exposures are aggregated into a clonal set, which
then competes with the cells in the repertoire on a per-cell basis for a limited position in
the repertoire. This ‘clonal set-repertoire competition’ is intended to provide more oppor-
tunity for per-exposure resource allocation, by allowing clones to compete and potentially
displace repertoire members that may or may not be the clonal progenitor (progenitor
of clones), clonal siblings (same progenitor), or exposure clonal siblings (same exposure).
This competition is promoted firstly through a cell-to-cell distance assessment that simu-
94
lates the competition the cells face during selection by antigen where similar cells compete
for the same antigenic patterns, and in the affinity comparison that selects those cells that
respond better to the specific antigen to which the system has been exposed. This compe-
tition mechanism decouples intra-repertoire competition for survival in the integration of
clones from the antigenic-based selection that resulted in the creation of the clones. Such
decoupling may be exploited in the integration of cells with varied origins.
Algorithm 4.5: Exposure Function for Replacement Cellular Clonal Selection.

Input: A, T, Nselected , Nclones , Pmutation
Output: Trs
1 Trs ←0;
3 Expose(A, Ci );
4 end
5 Tselected ← Select(T , Nselected );
6 Tclones ←0;
7 foreach Ci ∈ Tselected do
8 for i←0 to Nclones do
9 C&i ← Clone(Ci );
10 C&i ← Mutate(C&i , Pmutation );
11 Tclones ← C&i ;
12 end
13 end
14 Expose(A, Tclones );
15 for Cclone to Nclones do
16 Cselected ← 0;
18 if Distance(Cclone , Ci ) < Distance(Cclone , Cselected ) then
19 Cselected ← Ci ;
20 end
21 end
22 if Cclone .score ≤ Cselected .score then
23 T.Remove(Cselected );
24 T.Add(Cclone );
25 end
26 end
27 Trs ← GetBestMatchingCell(T);
28 return Trs ;
4.3.3 Empirical Assessment

This section defines a series of general empirical measures that provide instantaneous infor-
mation regarding a given CCSA on colour space specialisations of the Antigenic Exposure
Problem. An important consideration of these measures in the context of investigation via
exploratory experimentation is not their absolute value, but rather their relative change
in value with changes to the systems being investigated. Two general classes of measures
are considered, (1) error measures that assess the capability of a system in the context of
95
a given problem instance, and (2) diversity measures that consider the general state of the
information content within a given system.
Average Cell Error
The error of a cell may be assessed as its affinity for a given antigen, and may be assessed
using the problem domains specific Exposure(A, C) operation, in this case Euclidean Dis-
tance (defined in Equation 4.2). Although not used directly, Cell Error (CE) (defined in
Equation 4.3) provides an error-centric buillding building block for addition error assess-
ments.
CellError(A, C) = EuclideanDistance(A, C) (4.3)
An important application of Cell Error is its calculation for each cell in in the result
set (Trs ) returned from a cellular algorithms Exposure(I, T ) operation. The Average Cell
Error (ACE) defined in Equation 4.4 is a problem specific (ACSP) measure for assessing
the general capability of a cellular algorithm on an antigenic exposure problem.
Trsn
1 $
AverageCellError(I, Trs ) = CellError(Ai , Ci ) (4.4)
In i=1
Average Cell Diversity
A diversity measure of the cellular repertoire provides some indication of the variation
in the discrete information stored within the repertoire. Low diversity reflects a homoge-
neous repertoire, whereas high diversity reflects a heterogeneous set of cells. Given the
selected bit string basis of the CCSA, diversity between two cells may be measured as
the hamming distance between the cells (see Equation 4.6, where Hamming Distance is
defined in Equation 4.5, and ⊕ denotes exclusive-or (XOR).
L
$
HammingDistance(X, Y ) = (Xi ⊕ Yi ) (4.5)
i=1
CellDiversity(C1, C2) = HammingDistance(C1, C2) (4.6)
Cell Diversity (CD) can be calculated for each cell in the repertoire against all other
cells in the repertoire to provide an indication the average number of bit differences of
a given cell to the rest of the cells in the repertories, called the Average Cell Diversity
(ACD) defined in Equation 4.7.
 
Tn Tn
1 $ 1 $
AverageCellDiversity(T ) =  CellDiversity(Ci , Cj ) (4.7)
Tn i=1 Tn j=1
96
4.3.4 Trends and Behaviours
This section considers the expected emergent effects and behavioural trends of cellular-
based clonal selection algorithms as defined in the specific realisations.
Primitive Behaviour Trends
The foundational expectation is that the realised Cellular and Replacement Clonal Selec-
tion Algorithm behave much like CLONALG and AIRS. Specifically, that the algorithm
will acquire information via selection and adaptation. This section outlines the specific
foundational behaviour trends of the algorithms in the context of the pressures and prin-
ciples of the clonal selection strategy discussed in Section 4.2.1.
1. Repertoire Size: The repertoire size must be sufficiently large to ensure that se-
lection of cells and integration of clones do not result in competition between cells
specialised for different antigen. Such competition by an insufficiently sized reper-
toire is expected to result in unstable behaviour.
2. Selection Size: The number of clones selected per antigen exposure defines the
amount of resources allocated to each antigen under fixed clonal integration. An
increase in resource allocation via an increase in selection is expected to result in
increased specialisation via the promotion of ‘concurrent redundant perspectives’ of
a given antigen.
3. Cloning Size: The number of clones created per selected cell reflects the amount of
computational effort expended on re-sampling modifications on a ‘known-to-be-good’
structure. Therefore an increase in cloning and acceptance of limited high-affinity
clones is expected to result in an increased specialisation toward exposed antigen.
4. Replacement Constraints: Replacement constrained by a clonal progenitor is ex-

pected to result in a constrained and specialised hill climbing-like behaviour (MHCA).
The relaxation of this constraint such that clonal progeny compete directly with the
repertoire is expected to enhance the so-called ‘concurrent redundant perspectives’
to ensure that the maintained multiple perspectives are the most competitive, likely
resulting in improved repertoire specialisation.
Contextual Adaptation Expectations
This section outlines the specific expectations regarding contextual adaptation both with
regard to the localisation of activated and responding cells.
1. Localised Activation: Strong and non-overlapping (specialised for the same antigen)
competition by selection is expected to sufficiently localise polyclonal activation and
resultant adaptation of the clonal selection strategy (oligoclonal activation).
2. Localised Response: A top-down centralised aggregation of degenerate cells is ex-

pected to sufficiently localise a polyclonal response of selected cells under clonal
selection.
97
4.3.5 Paradigm Agenda
This section outlines the research agenda regarding the cellular clonal selection paradigm in
the remainder of this chapter. Specifically, the agenda is constrained to the investigation
of the expectations made regarding primitive behavioural trends of the cellular clonal
selection algorithms and the expectations regarding contextual adaptation of cells.
1. Investigate the expectations regarding the foundational clonal selection behaviour.
2. Investigate the expectations regarding contextual adaptation.
3. Investigate additional mechanism to influence the localisation of adaptation in the

clonal selection strategy.
4.4 Cellular Clonal Selection

This section considers three empricial studies of the cellular clonal selection algorithm.
Specifically, Section 4.4.1 considers the selection and cloning sizes in the Cellular Clonal Se-
lection Algorithm (CCSA), Section 4.4.2 considers repertoire resource allocation in CCSA
and the Replacement Cellular Clonal Selection Algorithm, and finally Section 4.4.3 consid-
ers contextual adaptation of symbolic and sub-symbolic degenerate cells under the clonal
selection strategy.
4.4.1 Cellular Empirical Study

Aim
The aim of this empirical study was to investigate the CCSA as a viable realisation of
cellular clonal selection and to assess the expected behaviour of this foundational algorithm
with regard to the number of cells selected and the number of clones created under varied
antigenic environments. Toward this end, the study had the following goals:
1. Assess the repertoire composition and capability of CCSA with small and large
selection and clonal sets.
2. Assess such behaviour under small and large antigenic environments.
Method
Problems The colour space specialisation of the Antigenic Exposure Problem (AEP)
was used called the Antigen Colour Space Problem (ACSP). Two variations of the problem
were employed ACSP-1 with a single CSP (Nantigen = 1) also referred to as the Minimal
Antigen Colour Space Problem, and ACSP-10 with 10 CSP (Nantigen = 10)3 . Both prob-
3
It is important to note that 10 patterns does not represent the expected extent of the approaches
capability, rather relative increase over the minimal problem case of one pattern. The scalability of the
pattern recognition capability of the investigated approaches is outsides the scope of this thesis, although
the approach is expected to be effected by the curse of dimensionality faced with similar pattern recognition
systems (see Section 8.4 for further discussion).
98
lem instances used the minimal determinant configuration where the scope of the determi-
nant was defined by the extent of each Colour Space Pattern (Ndeterminants = 1). Colour
Space Patterns were randomly generated at the start of each algorithm-problem run. In
the case of ACSP-10, CSP were randomly generated and accepted with the constraint
that pattern binary strings must have a minimum of a 64-bit Hamming Distance from all
other generated patterns for the run. The simple consistent Cellular Exposure Regime
(CER) defined in Algorithm 4.2 was used to manage the interaction of the problem with
the cellular algorithm.
Algorithms The study considered four different configurations of the Cellular Clonal
Selection Algorithm (defined in Algorithm 4.4). The CCSA configurations for the ACSP-
1 and ACSP-10 are defined in Table 4.3 and Table 4.4 respectively. The mutation rate
(Pmutation ) was fixed at 1
L = 1
192 ≈ 0.005 for all algorithm configurations. In addition
to the CCSA parameters, the tables also list the derived values for the number of clones
created per antigen exposure (Clones (A)), the number of positions in the repertoire under
competition per antigen exposure (Positions), and the total number of clones created per
infection exposure or epoch (Clones (I)). For all configurations, the number of cells was
set to ensure that selection and cloning did not result in explicit competition for resources
between cells specialised for different antigen.
CCSA Ncells Nselected Nclones Clones (A) Positions Clones (I)

1+1 1 1 1 1 1 10
1+N 1 1 10 10 1 100
N+1 10 10 1 10 10 100
N+N 10 2 5 10 2 100
Table 4.3: Summary of the assessed configuration of CCSA for ACSP-1.
CCSA Ncells Nselected Nclones Clones (A) Positions Clones (I)

1+1 1 1 1 1 1 10
1+N 1 1 10 10 1 100
N+1 100 10 1 10 10 100
N+N 100 2 5 10 2 100
Table 4.4: Summary of the assessed configuration of CCSA for ACSP-10.
Experiment Each algorithm used the Maximum Epochs Stop Condition (MESC) de-
fined in Equation 4.8 with M axEpochs = 1000, where an epoch was defined as the com-
plete exposure of an I to the T . The two cellular specific measures defined in Section 4.3.3
were collected from the state of the system after the triggering of the stop condition. These
measures included the Average Cell Error (ACE) and the Average Cell Diversity (ACD).
Each algorithm and problem received a new and different random number seed each run.
Algorithm-Problem combinations were repeated 30 times.
StopCondition(Epochi ) = (Epochi ≥ M axEpochs) (4.8)
99
Results
Table 4.5 provides a summary of results for each algorithm-problem combination including
the mean (x̄) and standard deviation (σ) of collected measure values. The non-parametric
Mann-Whitney U statistical test was calculated pair-wise for all algorithms in each problem
group. Results are summarised as statistically significant if the null hypothesis that any
two given populations in the group are drawn from the same distribution (H0 : µ0 = µ1 ).
Specifically, H0 is rejected if the calculated p-value < α = 0.05, meaning the results are
statistically significant at the 5% level. Figures 4.5a and 4.5b show the ACD and ACE for
all algorithms on ACSP-1 respectively, and figures 4.5c and 4.5d show the ACD and ACE
for all algorithms on ACSP-10 respectively Figure 4.6 provides an example plot of the Trs
returned from the CCSAN+N algorithm on the ACSP-1 and ACSP-10 problems at after
the triggering of the stop condition (same configurations as experiment, random seed of 5
and 1 for the problem and algorithm respectively).
Problem System ACD ACE

ACSP CCSA x̄ σ x̄ σ
ACSP-1 CCSA1-1 0 0 0.165 0.074
ACSP-1 CCSA1-N 0 0 0.004 0.007
ACSP-1 CCSAN-1 85.267 1.104 0.041 0.027
ACSP-1 CCSAN-N 86.233 1.055 0.014 0.027
Significant Truea True
ACSP-10 CCSA1-1 86.363 0.808 0.161 0.037
ACSP-10 CCSA1-N 86.053 1.016 0.073 0.031
ACSP-10 CCSAN-1 95.006 0.103 0.053 0.013
ACSP-10 CCSAN-N 95.035 0.094 0.01 0.006
Significant True True
a
False for CCSA1-1 and CCSA1-N
Table 4.5: Summary of results for CCSA on ACSP-1 and ACSP-10.
Analysis
This section provides an analysis of the results reported in the previous section in the
context of the goals of the empirical study. Specifically, the analysis is broken down into
the selection and clonal set size trends for small and large antigenic environments.
Selection and Clonal Trends (ACSP-1) This section considers the selection and
clonal trends of the CCSA configurations on ACSP-1. The increase in the number of
selected cells from 1 to N resulted in a large increase in diversity as expected given the
required shift in increase in the number of cells in the repertoire to support the increased
selection size. The increase in the number of cells selected resulted in a decrease in error
with a single clone and an increase when multiple clones were created. The increase in the
number of clones created for selected cells from 1 to N resulted in very minor differences
in the repertoire diversity, although did result in a large decrease in the average cell error.
The results confirmed the expectation that in the optimisation of a repertoire for a single
100
(a) Average Cell Diversity (ACD) on ACSP-1. (b) Average Cell Error (ACE) on ACSP-1.
(c) Average Cell Diversity (ACD) on ACSP-10. (d) Average Cell Error (ACE) on ACSP-10.
Figure 4.5: Box-and-whisker plot’s from the Cellular Empirical Study.
(a) CCSA(N+N) on ACSP-1. (b) CCSA(N+N) on ACSP-10.
Figure 4.6: Example plots of Trs from the CCSA(N+N) at the end of the run on left section
on both plots, for both ACSP-1 and ACSP-10, where the problem solution is represented
on the right of each plot, and colour components within each CSP.
101
antigen, the creation of more clones results in an improved system capability given that
each clone represents a new trial and potential improvement over an already relatively
good structure.
Selection and Clonal Trends (ACSP-10) This section considers the selection and
clonal trends of the CCSA configurations on ACSP-10. The increase in the number of
selected cells resulted in a decrease in the system error with a small repertoire, although
in this case the trend held for both small and large repertoire sizes. The increase in the
number of clones resulted in only very minor changes to the average cell diversity, although
resulted in a consistent decrease in the the average cell error. The increase in the size of
the antigenic environment did not disrupt the general trend of increase number of clones
resulting in increase system capability. Unlike the single antigen case, the increase in the
number of selected cells also consistently resulted in a decrease in average cell error.
Conclusions
This section summarises the findings of the empirical study into the Cellular Clonal Se-
lection Algorithm in terms of the primitives that were the focus of the study and the
expectations that motivated the study.
1. The Cellular Clonal Selection Algorithm successfully addressed 1 and 10 antigen

problems, demonstrating itself as a viable realisation of a foundational clonal selec-
tion adaptive strategy under the circumstances considered.
2. The increase in the number of selected cells (allocated resources) results in improved
system capability with small cloning on the single antigen problem, although is
consistently beneficial on the larger antigenic problem.
3. The increase in the number of clones (trials) results in an improved system capability
across the tested antigenic environments.
4.4.2 Replacement Empirical Study

Aim
The aim of this empirical study was to investigate the properties of the CCSA and the
RCCSA as a viable realisation of cellular clonal selection that promote improved utili-
sation of repertoire resources. Improved utilisation will effect the composition and may
effect the capabilities of a repertoire, where the promotion and maintenance of redundant
perspectives is expected to provide benefits in terms of fault tolerance. Toward this end,
the study had the following goals:
1. Assess varied clonal integration mechanisms with CCSA that promote concurrent
redundant perspectives of each antigen.
2. Assess varied clonal integration mechanisms with RCCSA that promote concurrent
perspectives and compare to CCSA.
102
Method
Problems This study used the ACSP-10 problem used for the CCSA empirical study
in Section 4.4.1.
Algorithms This section considers varied configurations of the CCSA and the RCCSA.
The configuration parameters of each algorithm type are presented in Table 4.6. The
chosen configurations of the repertoire, number of selected cells and number of clones fa-
cilitate the allocation of a maximum of six repertoire position per antigen. The CCSA
configurations included the CCSA(N+N) from CCSA empirical study in Section 4.4.1 as
well as two extensions. CCSA(N+N)-G aggregates all clones into a single group which
collectively compete with the Nselected selected cells from the repertoire for their position.
CCSA(N+N)-GS also aggregates all clones into a single clonal group, each of which com-
petes with the same number of the best matching cells as there are in the clonal set drawn
from the repertoire. Two variations of the RCCSA algorithm (defined in Algorithm 4.5)
were assessed, both of which used Hamming Distance (Equation 4.5) between cells in the
replacement mechanism. The RCCSA-H configuration allowed clones to select the most
similar cells in the repertoire to compete with for a position, whereas the RCCSA-H-
ES configuration restricted selection to those cells in the repertoire that were not clonal
siblings (hamming similarity excluding siblings).
CSA Ncells Nselected Nclones Clones (A) Positions Clones (I)

N+N 60 2 3 6 2 60
N+N-G 60 2 3 6 2 60
N+N-GS 60 2 3 6 6 60
RCCSA-H 60 2 3 6 6 60
RCCSA-H-ES 60 2 3 6 6 60
Table 4.6: Summary of the assessed configuration for the CCSA-RCCSA empirical study.
Experiment This study used the same experimental configuration including stop condi-
tions, measures, as were used for the CCSA empirical study in Section 4.4.1. An additional
measure was used to provide an assessment of the number of cells (C) in the repertoire
(T ) specialised to each antigen (A) in the scope of the problem (I). This measure is called
the Average Best Matching Cells Per Antigen (ABMCPA) defined in Equation 4.9 that
assessed and located the set of best matching cells for each antigen (Algorithm 4.6) and
averaged the number of cells returned for each antigen.
In
1 $
AverageBM CP erAntigen(I, T ) = ABM CS(Ai , T )n (4.9)
In i=1
Results
103
Algorithm 4.6: Best Matching Cells (BMC’s) for a given Antigen.
Input: T, A
Output: TBM C
1 TBM C ←0;
2 for Ci ∈ T do
3 Exposure(A, Ci );
4 if TBM C n > 0 then
5 C& ← TBM C i ;
6 if Ci .affinity < C&.affinity then
7 TBM C ←0;
8 TBM C ← Ci ;
9 else if Ci .affinity ≡ C&.affinity then
10 TBM C ← Ci ;
11 end
12 else
13 TBM C ← Ci ;
14 end
15 end
16 return TBM C ;
Mann-Whitney U statistical test was calculated pair-wise for all algorithms. Figures 4.7a,
4.7b, and 4.7c show the ACD, ACE, and ABMCPA on all algorithms respectively. Fig-
ure 4.8 provides example plots of the state of the repertoire (left of each plot) compared
to the problem domain (right of each plot) from each of the five algorithms (same config-
uration, random seeds of 1 and 5 for the algorithm and problem respectively).
Problem System ACD ACE ABMCPA

ACSP CSA x̄ σ x̄ σ x̄ σ
ACSP-10 CCSA(N+N) 94.356 0.171 0.027 0.012 1 0
ACSP-10 CCSA(N+N)-G 93.995 0.24 0.023 0.014 1.82 0.124
ACSP-10 CCSA(N+N)-GS 88.456 0.937 0.062 0.015 4.907 0.528
ACSP-10 RCCSA-H 94.37 0.18 0.03 0.013 1 0
ACSP-10 RCCSA-H-ES 88.083 0.79 0.028 0.013 2.653 0.385
Significant Truea True Trueb
a
False for CCSA(N+N) and RCCSA-H
b
False for CCSA(N+N) and RCCSA-H
Table 4.7: Summary of results from the CCSA-RCCSA empirical study on ACSP-10.
Analysis
context of the goals of the empirical study.
CCSA Trends With regard to average cell diversity, the aggregation and thus com-
petition between the entire clonal set with the selected set resulted in a slight drop in
diversity, although the relaxing of the competition from progenitor cells to the entire
104
(c) Average BMC Per Antigen (ABMCPA) on ACSP-10.
Figure 4.7: Box-and-whisker plot’s from the Replacement Empirical Study.
105
(a) CCSA(N+N). (b) (N+N)-G.
(c) (N+N)-GS. (d) RCCSA-H.
(e) RCCSA-H-ES.
Figure 4.8: Plots of the ordered state of the repertoire after the triggering of the stop
condition (left of each plot) for the assessed CCSA and RCCSA algorithms on the ACSP-
10, where the right of each plot provides the problem optima.
106
repertoire resulted in a further larger decrease in diversity by ≈ 5.5 bits. Interestingly
the CCSA(N+N)-G achieved a slightly lower error than CCSA(N+N) where the further
relaxation of competition in CCSA(N+N)-GS resulted in a large relative increase in the
ACE. Correlating with the decrease in ACD, the relaxation in the integration of the clones
resulted in an increase in the number of best matching cells (BMC’s) per antigen, rising to
an average of ≈ 5 in the case of CCSA(N+N)-GS. The results demonstrate that multiple
concurrent and redundant perspectives of antigen can be achieved by relaxing the replace-
ment competition in CCSA which decreases repertoire diversity. The results also highlight
that competition between the clonal set results in an improved result in terms of ACE
and ABMCPA over CCSA(N+N), although the relaxation from progenitor set to whole
repertoire dramatically increases ABMCPA at expense of increasing the systems general
capability. The specialisation of each antigen’s cellular imprint in the repertoire provided
by CCSA(N+N)-GS is clearly demonstrated in Figure 4.8c compared to CCSA(N+N)-G
in Figure 4.8b.
RCCSA Trends The exclusion of clonal siblings from replacement resulted in a signif-
icant decrease in diversity, the same seen between CCSA(N+N)-G and CCSA(N+N)-GS.
The exclusion of progeny also resulted in a very slight decrease in ACE over RCCSA-
H. Interestingly, the repertoire-wide competition between the aggregated clonal set in
RCCSA-H resulted in a single average BMC per antigen, whereas, the exclusion of progeny
during replacement promoted the competition between the concurrent clonal set such
that ABMCPA increased to ≈ 2.5. Also interesting was that the ACE and ABMCPA
results for RCCSA-H were not significantly different from CCSA(N+N), demonstrat-
ing that unrestricted similarity-affinity based replacement into the repertoire behaves
much like CCSA(N+N) confirming the expectation that motivated the exclusion of clonal
progeny: that they are more similar. Also interesting is that the increase in ABMCPA
with RCCSA-H-ES also resulted in a relative increase in ACE compared to CCSA(N+N)
and CCSA(N+N)-G, suggesting that for the mechanisms used, increasing the specialisa-
tion of the cellular footprint for each antigen comes at the expense of general repertoire
competence.
Conclusions
This section summarises the findings of the empirical study into the CCSA and RCCSA
in terms of the primitives that were the focus of the study and the expectations regarding
proportional resource allocation.
1. Relaxing the constraints for competition in the integration clones into the repertoire
results in increased number of concurrent redundant perspectives of an antigen in
the repertoire for CCSA and RCCSA.
2. Strong inter-clonal competition in CCSA results in improved performance with a

slight increase in size in the number of redundant perspectives.
107
3. Repertoire-wide integration restrained by only similarity-and-affinity in RCCSA ap-
proximates CCSA(N+N) in behaviour.
4. Repertoire-wide integration restrained by affinity in CCSA or by similarity-and-

affinity-exclusion in RCCSA results in a large increase in the size and quality of the
number of redundant perspectives at the cost of general repertoire capability under
both algorithms.
4.4.3 Degenerate Empirical Study

There is a lack of a one-to-one correspondence between antigen and receptors in the
acquired immune system. A given antigen may have the capacity to trigger a very larger
number of receptors, therefore resulting in a polyclonal activation, although a response is
typically oligoclonal. This may be accounted for by the competition between clones for
selection by the antigen. In addition, a given receptor may respond to a large number of
antigens, thus resulting in a polyclonal response, something that cannot be accounted for
by the clonal selection theory. In the Cognitive Theory of Immunology, Cohen refer to
this polyclonal response of a given receptor as cellular degeneracy, where those receptors
that are without a context are cross-specific, including auto-specific [82]. A polyclonal
response is the antithesis of clonal dominance (a feature that underlies the clonal selection
theory). The solution as proposed in cognitive theory is that specificity is an emergent
phenomenon [80, 205]. Unlike the clonal selection theory, where specificity is a property of
antigen-receptor interaction, emergent specificity is a down-stream effect that occurs after
the initial interaction. A collection of varied and communicating cell types respond to
the antigen in context. This is the so-called meta-response of cognitive theory, called the
co-response or corespondence. The degeneracy of cell signalling is proposed as the basis
of plasticity both in the brain and in the immune system, and is the feature exploited by
antibiotics and pharmaceuticals.
Different
Antigen receptors
Poly clonal A ctivation
Receptor
Different
antigens
Polyclonal Recognition
Figure 4.9: Depiction of a polyclonal activation and a polyclonal response.
Clonal selection accounts for the restriction of a potential polyclonal activation of the
repertoire to an oligoclonal activation by dominant clones out competing the less well
fitted clones (referred to as clonal dominance). The classical theory does not account
108
for lack of one-to-one correspondence between cells and antigen (the cross-reactivity of
receptors referred to as polyclonal response). Cognitive immune theory accounts for poly-
clonal response between antigen and receptors by suggesting (1) that cells in isolation are
degenerate and cross-specific, (2) context defines the use for degenerate cells resulting in
emergent specificity.
Aim
Section 4.3.4 considered the expectations of clonal selection on a degenerate repertoire,

specifically with regard to the pressures required to facilitate polyclonal activation and
response. The aim of this empirical study is to investigate these expectations in the
context of cellular clonal selection. Toward this end, the study had the following goals:
1. Investigate a clonal selection with a degenerate representation.
2. Assess selective pressures and explicit aggregation to constrain polyclonal activation

and response.
Method
Problems The definition of the Colour Space Domain in Section 4.2.3 highlighted a
‘retina and colour perception’ analogy made by Cohen to describe cellular degeneracy and
emergent specificity. The problems in this section are motived by that analogy. This study
used the ACSP-10 problem used for the CCSA empirical study in Section 4.4.1, although
evaluated on a per colour component basis. To differentiate this mode of evaluation from
the holistic ACSP, it is referred to as the Determinant Colour Space Problem (DCSP-
10). A degenerate colour representation was used called component degeneracy where
each colour was divided into the its Red, Green, and Blue components. Components are
denoted D, where a A = {D1 , D2 , D3 } for each colour space pattern. Degenerate compo-
nent cells were assessed using Equation 4.10, where each degenerate component cell was
defined by a 64-bit bitstring and decoded to a real value using Gray Code (Equation 4.1).
Components were considered a symbolic degenerate representation as each component has
direct meaning via their linear aggregation in the the context of a colour space pattern.
ComponentDistance(D, C) = |D − C| (4.10)
Algorithms The study considers a Degenerate Cellular Clonal Selection Algorithm (DCCSA)
as an extension of the RCCSA (defined in Algorithm 4.5) that uses a degenerate repre-
sentation. RCCSA was selected as a base algorithm because (1) it achieves adequate
results in terms of repertoire capability, and (2) because it achieves multiple high-affinity
concurrent perspectives (proportional repertoire composition) on antigen to which it is
exposed. The RCCSA was configured with the following configuration: Ncells = 100,
with Nselected = 3, and Nclones = 1 per determinant exposure, resulting in a resource
allocation of three cells per component, or nine cells per antigen. A Hamming-based sim-
ilarity assessment was used in replacement, with clonal sibling exclusion. Two variations
109
of the degenerate cellular component selection were considered as follows: (1) explicit
per-component selection called DCCSA-E, and (2) pre-committed component cells called
DCCSA-P. Both approaches aggregated the best degenerate cell for each component into a
holistic solution to address each given antigen in the ACSP. Explicit per-component selec-
tion maintained a repertoire of uncommitted components that were assessed and selected
in the context of each component of each colour space pattern. This allowed degenerate
cell reuse across the components of a single antigen and across the components of multiple
antigen. Pre-committed component cells involved the management of one sub-repertoire
for each component, where the size of each repertoire was a fraction of the total number
of cells (. Ncomponents
1
× Ncells /).
Experiment This study used the same experimental configuration including stop condi-
tions, measures, as were used for the CCSA empirical study in Section 4.4.1. An additional
measure was introduced for the degenerate component algorithms that assessed the ex-
tent of the cross-reactivity of the repertoire called the Average Polyclonal Response Error
(APRE), defined in Equation 4.11. APRE provides an indication of the average repertoire
error per component.
 + ,
In A Tn
1 $ $ in
1 $
AP RE(I, T ) =  ComponentDistance(Ai j , Ck )  (4.11)
In × Ain i=1 j=1 Tn k=1
Results
Mann-Whitney U statistical test was calculated pair-wise for all algorithms.
Problem System ACD ACE APRE

AEP CSA x̄ σ x̄ σ x̄ σ
DCSP-10 DCCSA-E 31.638 0.047 0.003 0.001 0.333 0.018
DCSP-10 DCCSA-P 31.627 0.059 0.008 0.003 0.334 0.019
Significant False True False
Table 4.8: Summary of results from the DCCSA empirical study on DCSP-10.
Analysis
context of the goals of the empirical study, specifically (1) restrictive selection, and (2)
explicit aggregation.
Restrictive Selection The results showed little significant difference on the selected
measures between the explicit and pre-committed selection mechanisms, other than a
slightly improved cell error for the explicit approach. The strong similarity in results, in
110
particular the Average Polyclonal Response Error demonstrate that the degenerate com-
ponent clonal selection was generally unaffected when the repertoire was split into sub-
repertoires per component compared to an integrated cross-component competitive reper-
toire. The minimum Hamming distance between generated colour space patterns likely
resulted in little cross-colour space pattern reuse given the per-pattern/per-component
specialisation of the repertoire, and the lack of difference between segregated and inte-
grated component repertoires. An important consideration is that CCSA and RCCAS
implicitly rely on restrictive selection to bound polyclonal activation of cells. This is
clear if we consider a principle of cellular clonal selection that the scope of the repertoire
available allowing Ncells interactions (assessments and activations). Without constrained
selection, the scope of repertoire would respond in its entirety each antigen exposure.
Explicit Aggregation The clonal selection strategy by-design continues regardless of

the integration of the product of exposure events. With respect to the problem domain,
aggregation of the product of exposures is the critical concern of a degenerate represen-
tation. The chosen degenerate component representation was chosen because of its easy
linear aggregation of the best matching components, always resulting in feasible aggregated
responses at the per-D, per-A, and per-I levels. One may consider a sub-symbolic repre-
sentation for the ACSP that does not easily aggregate into holistic colour space patterns
solutions, such as sub-bitstrings. This is an important example because it highlights that
such representation does not hinder the clonal selection strategy. Degenerate sub-strings
may be assessed based on hamming distances from exposed A’s or D’s providing enough
information for selection, cloning and integration. The explicit aggregation of sub-strings
is less natural than the component case, requiring perhaps a per-bit position frequency
assessment and a deterministic or probabilist generation of a viable holistic solution at the
A scale. The example highlights that explicit aggregation may be suitable for the linear
aggregation of components, although becomes more difficult with sub-symbolic represen-
tations one which clonal selection can still operate effectively. Explicit aggregation also
highlights that this problem-centric concern is scale independent, where the components of
a colour, and the colours of a colour set are the linear aggregation of selected information
which may or may not be the case for more elaborate applications.
Conclusions
This section summarises the findings of the empirical study into the DCCSA in terms of
the primitives that were the focus of the study and the expectations regarding proportional
resource allocation.
1. Degenerate Components
(a) Clonal Selection on components provides a viable realisation of cellular degener-

acy under the circumstances considered where strong selection and aggregation
can constrain polyclonal activation.
111
(b) A RCCSA results in little difference in the partitioning or aggregation of the
repertoire and thus competition of degenerate cells, suggesting clonal indepen-
dence promoted via replacement is sufficient for non-overlapping clonal special-
isation (assuming a large enough repertoire).
(c) Explicit aggregation provides a context in which to bound polyclonal response
to strongly activated (selected) structures.
2. Selection and Aggregation
(a) Strong selection is required in all cellular clonal selection algorithms to man-
age the implicit polyclonal activation that occurs as a result of each antigens
interaction with the entire repertoire.
(b) Aggregation is independent of the principle clonal selection operations, requir-
ing explicit integration into the process (for those applications that operate on
sub-solution structures).
4.5 Spatial Repertoire

This section considers a spatial context as a method for forming relationships between
units undergoing clonal selection and expansion. A spatial clonal selection paradigm is
investigated which outlines a fixed lattice structured repertoire on which clonal selection
operators are applied.
4.5.1 Spatial Clonal Selection

The processes of selection and maturation of lymphocytes occurs in the spatially dis-
tributed confines of the host organism lymphatic system [19]. Further, the spatial or-
ganisation may be required to provide context to guide emergent specificity [81]. The
competition in the clonal selection principle is between cells for resources in the reper-
toire. Competition can be facilitated through differential selection of an activated set
which results in differential allocation of resources for clones from the activated set. In
envisaging the repertoire as a spatial structure, an additional level of competition is in-
troduced called spatial competition. This competition puts pressure on cells in the same
spatial neighbourhood to compete with each other. This pressure may be used for either
the activation of cells during selection or replacement by clones of activated cells. A simple
one or two-dimensional lattice structure may be used in which receptors occupy grid posi-
tions of the lattice, and the ends of the structure wrap around (ring or toroid), removing
edge effects. The structure may be an equally arbitrary number of dimensions, although
low dimensionality facilitates visualisation. Such a spatial repertoire structure provides
a manifestation of the space complexity limits imposed on repertoire size, and imposes
relationships between arbitrary neighbouring receptors for a given antigenic stimulus.
112
Antigen
Cells
Toroidal lattice
Figure 4.10: Depiction of a one-dimensional toroidal lattice.
Spatial Selection
The spatial structure may be exploited in the selection of the activated set for a given
antigen. In the selection scheme the repertoire is evaluated, scanned for the highest
affinity cell, which is returned to address the needs of the antigen, and an activated
set is selected from the high-affinity cells. Therefore, the conventional activated set may
be conceptualised as the selection of cells from across the spatial structure. One may
partition antigenic signals and match them to regions (logical partitions) of the spatial
structure, such that regions are directed in a top-down manner to be responsible for
specific antigenic signals. The input space can be partitioned by a measure relevant to
the representation of the antigen or the cells in the repertoire. This partitioning scheme
would be implemented such that activated sets of receptors may only be drawn from
the allocated region of the spatial structure, therefore putting pressure on the region
to produce and develop receptors suitable to the allocated portion of the input space. A
concern with this approach is that information effective for representational partitions may
be developed outside of allocated regions and thus would not be exploited (for example the
cross-reactive response for generalisation). Further, without inter-region competition, the
partitions may be considered isolated repertoires such that the partitioning prevents the
take-over or resizing of spatial regions with proportion to signal frequency and complexity
(an effect maintained by differential resource allocation across the whole repertoire).
Spatial Replacement
The Replacement CCSA investigated in Section 4.4.2 may be elaborated to exploit a spatial
structure such that replacement competition is spatially localised or constrained to the
locations in the repertoire that responded strongest to an antigenic stimulus. Competition
for resources may be constrained to the spatially local neighbourhood of high affinity
cells. The principle parts of a spatial replacement mechanism are as follows: a structured
repertoire or lattice data structure in which grid positions represent cells, a neighbourhood
function that identifies the scope of competition (see Figure 4.11), and a competition
function that defines how resources (lattice positions) are allocated to clones of selected
cells. Spatial competition introduces spatial meaning between positions in the lattice, such
that visualisation of the lattice can highlight relationships and facilitate the extraction of
qualitative information from the repertoire.
The spatially restricted ‘back-end’ selection (neighbourhood function) is expected to
113
Antigen
Activated
Repertoire
Neighbourhood Replacement
Clones
Figure 4.11: Depiction of clones competing in the neighbourhood of their progenitor.
result in the assignment of antigenic-responsibility to regions of the spatial repertoire struc-

ture4 . Therefore, the neighbourhood function constrain’s resource allocation resulting in
a localised clonal dominance effect, where those cells that win upfront selection will pro-
duce clones that dominate sub-structures in the lattice. In effect, the spatial replacement
strategy facilitates a mapping of the antigenic space onto the lower-order topology of the
spatial repertoire. The clonal selection axiom of ‘similar cells taking responsibility for sim-
ilar antigen’ is bonded to the lattice structure, facilitating the self-organisation of receptors
to input signals. Ultimately, the expected organisation of similar cells into regions results
in a broader competition between regions for activation, cloning, and the resulting ongoing
retention of the region for spatial resources. The selection of the replacement function can
play an important role in the maintenance behaviour of cell clusters. Clones are expected
to have the same or similar affinity for the antigen as the activated cell given relatively low
mutation rates. A clone that displaces a neighbouring cell is likely to be an activated cell
in the following exposure. A similarity or affinity based replacement function is expected
to maintain a cell cluster tied to a spatial locality and a random replacement function is
expected to shift the centroid of the cluster around such that the fringe of the cell group
may interact with neighbouring groups. In the case of the a stable spatial locality, upfront
dominant cells are expected to put pressure on their neighbourhood to conform to the
antigenic signals to which they dominant. Dominance leads to neighbourhood takeover
which results in the emergent effect of self-organised responsibility.
Explicit and Implicit Organisation
Exploitation of the spatial structure as an upfront pressure results in the partitioning of

input signals to regions of the spatial structure. This has the benefit of explicitly limiting
the scope of interest of input signals space for regions of the lattice, with the problems
associated with making assumptions regarding the partitioning the signal space. All re-
sources are employed for a specific task, although not all the allocated resources may
be required for their allocations. The exploitation of the spatial structure as a back-end
pressure results in the implicit partitioning of the input space and self-organisation of
receptor clones to take responsibility for the automatic partitions. This has the benefit of
4
‘Back-end’ refers to competition at the end of an iteration of the algorithm cycle.
114
proportionate allocation (specialisation) of resources for the automatically identified com-
plexities of the input space with the computational and delay costs for the self-organising
process. Resource allocation is determined automatically, although not all resources may
be utilised effectively. An interesting observation is that one method will lead to the other,
such that the explicit exploitation of the spatial competition of both ends is not required.
The application of upfront spatial partitioning will force regions to adapt to input signals,
thus result in replacements occurring in the same region, reinforcing the partitioning. The
application of back-end partitioning will self-organise the responsibility for antigenic sig-
nals via replacement, which will reinforce future input signals being directed to responsible
areas of the spatial structure.
A B A B
A B A B A
Figure 4.12: Depiction of the explicit versus implicit partitioning of input signals and the
effects on the repertoire.
4.5.2 Spatial Repertoire Empirical Study

Aim
The aim of this investigation is to investigate cellular clonal selection constrained by a

spatial repertoire structure. Toward this end, the study had the following goals:
1. Investigate the effects on the repertoire composition and capability in adopting a

spatial repertoire structures.
2. Assess the effects on the repertoire in exploiting the spatial structure to localise
clonal integration via replacement.
Method
Algorithms This study considers the Replacement Cellular Clonal Selection Algorithm
(RCCSA), and two variations of the Spatial Cellular Clonal Selection Algorithm (SCCSA).
The RCCSA defined in Algorithm 4.5 was configured with Ncells = 100, Nselected = 2, and
Nclones = 5 to promote an allocation of ten cells per antigen in ACSP-10. Unlike RCCSA,
SCCSA does not aggregate and integrate clones as a group into the repertoire, instead
clones are replaced individually, with per-progenitor clonal sibling replacement exclusion.
Two variations of the SCCSA algorithm were investigated: (1) a variation that organised
the repertoire into a lattice although was constrained by the spatial organisation called
holistic replacement (SCCSA-HR), and (2) a variation that exploited the spatial repertoire
115
and constrained the integration of clones to the spatial neighbourhood of the progenitor cell
and with exclusion of clonal siblings (like RCCSA and SCCSA-HR) called neighbourhood
replacement (SCCSA-NR). The neighbourhood was defined as the square of the nine cells
surrounding and including a given position in the toroidal lattice. Both SCCSA used the
same configuration as RCCSA, although the lattice dimensions used the square root of
Ncells resulting in a 10-by-10 2-dimensional repertoire.
in Section 4.4.1.
Experiment This study used the same experimental configuration including stop con-
ditions, measures, as were used for the RCCSA empirical study in Section 4.4.2, including
the Average Best Matching Cells Per Antigen (ABMCPA). An additional measure was
used that calculated the average diversity of the 9-cell neighbourhood for each position in
the lattice, called the Average Cell Neighbourhood Diversity (ACND). The diversity was
calculated using the Average Cell Diversity (Equation 4.7) method for each neighbour-
hood, and averaged across all 100 neighbourhoods.
Results
4.13b, 4.13c, and 4.13d show the ACD, ACE, ABMCPA, and ACND on all algorithms
respectively. Figure 4.14 provides example plots of the state of the repertoire from the two
spatial-based algorithms (same configuration, random seeds of 1 and 5 for the algorithm
and problem respectively).
Problem System ACD ACE ABMCPA ACND

ACSP CCSA x̄ σ x̄ σ x̄ σ x̄ σ
ACSP-10 RCCSA 89.128 0.732 0.002 0.002 3.94 0.803 N/A N/A
ACSP-10 SCCSA-HR 92.876 0.405 0.012 0.012 2.83 0.352 83.257 0.508
ACSP-10 SCCSA-NR 87.592 0.968 0.002 0.001 3.417 0.424 65.119 1.514
Significant True Truea True True
a
False for RCCSA and SCCSA-NR
Table 4.9: Summary of results for SCCSA on AEP 10.
Analysis
116
(c) Average BMC Per Antigen on ACSP-10. (d) Average Cell Neighbourhood Diversity on ACSP-
10.
Figure 4.13: Box-and-whisker plot’s from the Spatial Repertoire Empirical Study.
(a) SCCSA-HR on ACSP-10. (b) SCCSA-NR on ACSP-10.
Figure 4.14: Example plots of the spatial repertoire from two variations of SCCSA at the
end of the run on ACSP-10 with BMC represented with circles.
117
Spatial Structure Trends This section considers the effects of storing the repertoire
in a spatial structure by comparing the results of RCCSA to SCCSA with holistic re-
placement. The use of the spatial structure in SCCSA-HR resulted in a small increase
in repertoire diversity and in particular response error where ACE also showed a large
increase in the variance of final ACE scores. The increase in diversity and error corre-
lated with a decrease in the average number of BMC’s. The spatial structure was used in
SCCSA-HR to hold the repertoire although was not exploited in any way. The difference
in the repertoire composition and capability between the approach and RCCSA demon-
strate the effect of the per-progenitor clonal set replacement used in SCCSA compared to
the aggregated clonal set replacement in RCCSA.
Localised Replacement Trends This section considers the exploitation of the spa-
tial structure in the neighbourhood replacement used in SCCSA compared to holistic
replacement. The SCCSA-NR resulted in a the increased organisation of the repertoire
as expected by localising cells for the same antigen into areas on the spatial repertoire
structures. This was demonstrated by the decrease in repertoire diversity compared to
both SCCSA-HR and RCCSA, and a small increase in ABMCPA compared to SCCSA-
NR. Interesting the exploitation of spatial neighbourhood competition resulted in a ACE
not significantly different from RCCSA, suggesting that such competition results in the
same level of per-antigen specialisation within the repertoire as RCCSA. Importantly the
increased organisation was demonstrated by the significant decrease in average repertoire
diversity compared to holistic replacement, which was depicted in the example plot in
Figure 4.14 showing that this decrease was caused as a result of the clear grouping of
similar cells (similar colours) in the spatial repertoire.
Conclusions
This section summarises the findings of the empirical study into the Spatial Cellular Clonal
Selection Algorithm in terms of the primitives that were the focus of the study and the
1. Neighbourhood replacement provides intra-clone competition that results in similar

general behaviours as consolidating the clone for replacement in RCCSA.
2. Neighbourhood replacement results in the spatial localisation of clones on the lattice,

the so called spatial responsibility effect as expected.
4.6 Response Mediation

This section elaborates on the clonal selection by introducing a second repertoire of in-
termediate (mediator) cells to provide an adaptive context for localising the differential
resource allocation of clonal selection, providing a model of decoupled feature detection
and system response.
118
4.6.1 Mediated Clonal Selection
Two-signal theories have been proposed in immunology as verification signals permitting
the proliferation and differentiation of lymphocytes. Historically developed for the ac-
tivation of B-lymphocytes (in the context of self-nonself discrimination), the two-signal
approach has also been extended to the various classes of T-cells. The dominant type of B-
cell activation is dependant on Helper T-cells that identify MHC molecules on the surface
of activated antigen-presenting B-cells. The detection of MHC provides a secondary veri-
fication signal to the B-cells, allowing them to proliferate and differentiate. This process
is called the co-stimulation of B-cells by T-cells or T-cell dependant activation of B-cells.
Likewise, T-cell activation requires an antigen specific signal via an antigen-presenting cell
and a co-stimulating ‘antigen non-specific’ verification signal. The theory of the two-signal
activation of T-cells is called associative recognition theory [57].
Mediated Clonal Selection or so-called inter-repertoire recognition decomposes the
clonal selection process such that a mediation process is responsible for selecting those
cells from the activated set that may proliferate. The mediation process is controlled by
a sister repertoire of cells that also perform a clonal selection process using the activated
cells of the first repertoire as input signals. The Helper T-cell metaphor is adopted such
that the repertoire of cells (first repertoire) represents B-lymphocytes, which require a
second verification signal from the helper T-lymphocytes (second repertoire) before pro-
liferating and differentiating. After providing the secondary signal to the activated set of
B-lymphocytes, the helper T-lymphocytes proceed with the cloning and maturation of the
activated T-cells. Therefore, the T-cell repertoire is an application of the clonal selection
algorithm that accepts the activated set of another clonal selection algorithm as a pseudo-
‘antigenic set’. Two concerns that must be reconciled in integrating the two repertoires are
as follows: (1) The T-cells must select those B-cells from the B-cells activated set that may
proliferate (2) The B-cell activated set provides multiple antigens simultaneously to the
T-cell repertoire, to which it is possible to generate a T-cell activated set for each member
of the B-cell activated set. The remainder of this section considers the implications of the
second clonal selection governed repertoire from a variety of different perspectives.
Mapping Function
The proliferation verification signal for the activated B-cell’s may be antigen-dependent
or antigen-independent. If antigen-dependent, the information provided by the verifica-
tion signal may be considered a generalisation of the input signals provided to the B-cell
repertoire. If the verification signal is antigen-independent, then the information pro-
vided is a generalisation of the cells themselves in the B-cell repertoire. This is a subtle
but potentially important difference that effects the representation of information by the
system.
An antigen-dependent mapping from B-cell to T-cell requires that some regularities
of the input signal are preserved. The transformation may be of the whole antigen itself,
or just the part of the antigen recognised by the B-cell receptor. The T-cell repertoire
adapts to a consistent (B-cell invariant) generalisation of the antigenic input patterns. An
119
antigen-dependent mapping between the repertoires provides a top-down organisation of
the relationship between the repertoires, where simply matching onto the second reper-
toire provides the verification signal. A potential concern with antigen dependence is that
the mapping of the input patterns may be achieved in one repertoire, making the second
repertoire redundant. An antigen-independent mapping disregards the regularities of the
input signal, such that the T-cell is performing pattern recognition of B-cell receptors,
rather than a pattern recognition of antigen (whether symbolic or sub-symbolic). The
result is a T-cell repertoire that is developed independent of antigen, and dependent on
the B-cell repertoire (which in turn develops in response to antigen). A potential concern
with an antigen-independent mapping is that it is arbitrary such that T-cells are com-
pletely dependent on specific B-cell lineages, and regularity between B-cell patterns has
no meaning other than ancestry (likely to make the mapping harder).
Formation of High-Order Structures
Interestingly, the relationship between an antigen and cells, as well as cells between reper-
toires may be considered in the context of the formation and maintenance of higher-order
structures. For example in the case of a single repertoire, one antigen may map onto one
cell providing a specialised mapping. This may be elaborated to include the extremes
cardinality of relationship. For example, a one-to-many relationship between antigen and
cells maybe considered a decomposition of feature extraction, particularly if the mapping
is partial or sub-symbolic. Additionally, many antigen may be mapped onto many cells
either directly (generalisation of one-to-one) or partially (cross reactivity), or onto a single
cell, compressing the signal.
The relationship between repertoires adds a second level of complexity to the hierarchi-
cal structure. The same relationships apply although in this case between cells, with the
importance difference that the relationships can be manipulated with mapping schemes,
activated B-cell set size, and activated T-cell set size. For example, it may be desirable
to coerce the B-cell repertoire to decompose antigen into regular features (many), and the
T-cell receptor to generalise those features toward a specific meaning (few). Alternatively,
the B-cell repertoire may be coerced to generalise (compress) based on common antigenic
features (few), and the T-cell repertoire to decompose the B-cell information content into
a variety of different meanings (many).
T-Cells
B-Cells
T-Cells
B-Cells
(a) Many-to-One. (b) One-to-Many.
Figure 4.15: Depiction of inter-repertoire structures based on the mapping of cells.
120
Rather than treating each activated B-cell independently, the activated B-cells may
interact with the T-cell population concurrently, such that multiple B-cells may activate
a single T-cell. The T-cell repertoire may be configured such that a T-cell is required to
match more than one B-cell before it may become activated. This provides a mechanism
in which the T-cell repertoire facilitates the aggregation of multiple antigenic features
into a single T-cell. The second repertoire facilitates the addition of a second layer of
meaning. In the many-to-one case a T-cell may represent a high-order concept as an
aggregate of multiple lower-level concepts (B-cells). The meaning is a single or small
collection of concepts that are dependent on the presence of specific features (specific
context). For example, Many-to-One provides a consensus forming operation in which
many lower-order concepts (such as extracted features) are aggregated together to form
a high-order concept. The One-to-Many relationship provides a descriptive operation in
which a high-order concept is decomposed into a number of lower-order concepts (such
as descriptive features). In both of these examples, the meaning implies a mapping that
manages to preserve some regularities of the input signal, such that higher-order or lower-
order concepts may be described.
Supervised Concept Formation
The incorporation of additional feedback into the algorithm allows an external process
to assign meaning and supervise the mapping and formation of concepts. For the single
repertoire algorithm, T-cell mediation provides an immunologically plausible basis for
feedback (verification) to supervise the selection of activated receptor patterns in response
to antigenic exposures. For the dual repertoire algorithm, tissue damage provides an
immunologically plausible basis for feedback to supervise the selection of concepts formed
in the T-cell repertoire, which in turn feeds back to the B-cell repertoire via the mediation
process. The proposed integration of feedback results in two flows of information through
the system. The top-down flow of antigenic signals results in B-cell receptors competing for
activation, and T-cell receptors competing for activation of the B-cell receptor activated
set. Those B-cells that receive the secondary signal, proliferate. The bottom-up flow of
tissue-damage input results in T-cells competing for encouragement that the concepts that
they represent are useful. Those T-cells that receive a positive (non-negative, and perhaps
non-neutral) feedback, proliferate. Therefore, the application of feedback (if available) is
to the activated set of T-cells. In the absence of feedback, there is simply the absence of
corrective behaviour: the punishment, or reward of the activated T-cell concept.
Bottom-up supervision of concept formation provides a natural agenda for investigating
mediated clonal selection beyond the basic principles of mappings and structure formation
and maintenance (for example, see Figure 4.16). The integration of a feedback mechanism
provide a natural relationship with the study of Reinforcement Learning (Section 3.4.5).
121
Input signal Input sig nal
B-cells B-cells
T-cells T-cells
Incorrect Correct
Incorrect Correct
Feedback Feedback
(a) Many-to-Few. (b) Few-to-Many.
Figure 4.16: Depiction of the integration of feedback for supervised structure formation.
4.6.2 Mediated Empirical Study

Aim
The aim of this investigation was to asses the cellular clonal selection algorithm constrained
by inter-cellular repertoire interactions. Toward this end, the study had the following goals:
1. Assess the behaviour of different inter-repertoire antigen-dependant mapping schemes.
2. Investigate the effects of varied cellular cardinality in the relationships exploited via
inter-repertoire mapping.
Method
in Section 4.4.1.
Algorithms Algorithm 4.7 defines the Mediated Cellular Clonal Selection Algorithm
(ECCSA), where Exposure(A, T ) is a modified version of the Exposure operation for RC-
CAS (defined in Algorithm 4.5) that returns the selected set rather than the best matching
cell. The RCCSA uses clonal sibling exclusion and a Hamming similarity function during
replacement.
Algorithm 4.7: Exposure Function for Mediated Cellular Clonal Selection.

Input: A, Tbcells , Ttcells , Nbc−selection , Nbc−clones , Ntc−selection , Ntc−clones , Pmutation
Output: Trs
1 Trs ←0;
2 Tbc−rs = Exposure(A, Tbcells , Nbc−selection , Nbc−clones , Pmutation );
3 Ttc−rs = Exposure(Tbc−rs , Ttcells , Ntc−selection , Ntc−clones , Pmutation );
4 Trs ← SelectBestMatchingCell(Ttc−rs );
5 return Trs ;
The modified RCCSA Exposure operation also permits the mapping of a selected set
of cells (Trs ) to be treated as an antigen. This is achieved through use of an exposure
122
operation that provides a mapping function that aggregates affinity for a given cell against
a stimulating set of cells. The first part of this study considers two different examples of the
mapping function: (1) Euclidean Mapping, and (2) Hamming Mapping. The affinity scores
are aggregated within each cell of the repertoire against the stimulating set to provide a
general indication of a cells capability against a stimulating set. In this part of the study,
both the B-cell and T-cell repertoires used identical configurations as follows: Ncells = 100,
Nselected = 1, Nclones = 10 (10% of each repertoire per antigen). The second part of the
study considered varied relationships between the two repertoires in terms of the size of the
selected B-cell set which stimulates the T-cell repertoire, and the size of of the activated
and responding T-cell set. A series of one-to-one and many-to-many relationships were
assessed with the intermediates, where the number of clones and this repertoire assigned
per antigen was kept constant at six (10% of each repertoire per antigen) with regard to
the selection size. The configurations are summarised in Table 4.10. A Euclidean mapping
was used between the repertoires in this second part of the study.
Relationship B-Cells T-Cells

Name Nb−cells Nbc−selected Nbc−clones Nt−cells Ntc−selected Ntc−clones
1-to-1 60 1 6 60 1 6
1-to-N 60 1 6 60 6 1
N-to-1 60 6 1 60 1 6
N-to-N 60 6 1 60 6 1
Table 4.10: Summary of the configuration of the ECCSA with a series of inter-repertoire
selection relationships.
Experiment This study used the same general experimental configuration including
stop conditions as were used for the CCSA empirical study in Section 4.4.1. A series of new
measures were defined for the experiment. The Average Cell Diversity measure (defined in
Equation 4.7) was calculated for each of the B-cell and T-cell repertoires referred to as the
Average B-Cell Diversity (ABCD), and the Average T-Cell Diversity (ATCD) respectively.
This treatment was performed to the Average Cell Error measure (defined in Equation 4.4)
referred to as the Average B-Cell Error (ABCE) and the Average T-Cell Error (ATCE)
respectively, where the measure was taken as the best response from each repertoire against
the antigen. The Average BMC per Antigen (defined in Equation 4.9) was adapted for
the two re-repertoire configuration referred to as the Average Best Matching B-Cells Per
Antigen (ABMBCPA) and the Average Best Matching T-Cells Per Antigen (ABMTCPA)
respectively. The system’s response was taken as the best matching cell from the set of
selected (activated) T-cells for each antigen exposure, called the Response Error (RE). A
new measure was defined that calculated the T-cell error after mapping, averaged across
each cell in the T-cell repertoire called the Average T-Cell Mapping Error (ATCME).
This measure provided a general indication of the mapping error between the repertoire
in the units of the specific mapping scheme used. The measure was also taken against
just the set of T-cells activated and selected by the activated B-cells called the Average
T-Cell Selected Set Mapping Error (ATCSSME). Figure 4.17 provides a depiction of the
123
repertoires, the antigen, the selected sub-sets and the system response as well as how all
the collected measures relate to these principle attributes of the ECCSA.
Antigen
RE
ABCE
ABMBCPA
ABCD
Selected
B-Cells T-Cells
ATCME
ABMTCPA ATCD
ATCE ATCSSME
Selected T-Cells
T-Cells
Response
Figure 4.17: Depiction of the principle concerns of the MCCSA and the relationships in
terms of the selected measures.
Results
Table 4.11 and Table 4.12 provide a summary of results for each algorithm-problem com-
bination including the mean (x̄) and standard deviation (σ) of collected measure values.
The non-parametric Mann-Whitney U statistical test was calculated pair-wise for all al-
gorithms. Figures 4.19a, 4.19b, 4.19c, 4.19d, and 4.19e show the ABCD, ABCE, ATCD,
ATCE, and RE respectively. Figure 4.20 provide example plots for all four configurations
on the second relationship part of the study. Results from all example plots are taken
from the end of the run with algorithm and problem configurations matching those used
during experimentation, and a random number generator seed of 1 and 5 for the algorithm
and problem respectively.
Analysis
124
Measure Euclidean Hamming Significant
x̄ σ x̄ σ
ABCD 89.399 0.822 89.399 0.822 False
ATCD 89.234 0.718 87.999 0.94 True
ABCE 0.003 0.002 0.003 0.002 False
ATCE 0.009 0.006 0.098 0.029 True
RE 0.009 0.006 0.617 0.123 True
Table 4.11: Summary of results for ECCSA with two different mapping schemes on ACSP-
10.
(a) MCCSA-Euclidean on ACSP-10. (b) MCCSA-Hamming on ACSP-10.
Figure 4.18: Example plots of the B- (top of each plot) and T-Cell (bottom of each plot)
repertoires from the ECCSA at the end of the run on ACSP-10 for two different inter-
repertoire mapping types.
Measure (1-1) (1-N) (N-1) (N-N) Sig.

x̄ σ x̄ σ x̄ σ x̄ σ
ABCD 88.333 0.812 87.967 0.721 94.419 0.113 94.408 0.147 True
ATCD 88.005 0.755 94.298 0.146 88.103 0.924 94.309 0.105 True
ABCE 0.027 0.024 0.028 0.012 0.067 0.014 0.067 0.015 True
ATCE 0.052 0.021 0.076 0.015 0.102 0.019 0.093 0.013 True
ATCME 0.604 0.065 0.648 0.04 3.523 0.307 3.845 0.188 True
ATCSSME 0.042 0.02 0.158 0.023 0.844 0.129 1.161 0.121 True
ABMBCPA 2.713 0.33 3.13 0.426 1 0 1 0 Truea
ABMTCPA 3.25 0.318 1 0 2.597 0.549 1 0 Trueb
RE 0.059 0.031 0.083 0.016 0.112 0.022 0.122 0.019 Truec
a
False for ECCSA(N-1) and ECCSA(N-N)
b
False for ECCSA(1-N) and ECCSA(N-N)
c
False for ECCSA(N-1) and ECCSA(N-N)
Table 4.12: Summary of results for ECCSA with four different inter-repertoire relationship
schemes on ACSP-10.
125
(a) Average B-Cell Diversity (ABCD) on ACSP-10. (b) Average B-Cell Error (ABCE) on ACSP-10.
(c) Average T-Cell Diversity (ATCD) on ACSP-10. (d) Average T-Cell Error (ATCE) on ACSP-10.
(e) Response Error on ACSP-10.
Figure 4.19: Box-and-whisker plot’s for the relationship configuration with ECCSA.
126
(a) MCCSA(1-1). (b) MCCSA(1-N).
(c) MCCSA(N-1). (d) MCCSA(N-N).
Figure 4.20: Example plots of the B- (top of each plot) and T-Cell (bottom of each
plot) repertoires from the MCCSA at the end of the run on ACSP-10 for all four inter-
relationship types.
127
Mapping Trends This section considers the effects of using the Euclidean and Ham-
ming antigen-dependant inter-repertoire schemes. The B-cell diversity and error show no
significant difference between the mapping schemes. This is an expected result, as the
mapping scheme and the chosen ECCSA does not effect the B-cell repertoire. The T-
cells showed a small decrease in diversity with the Hamming mapping, and more critically
showed a factor of 10 increase in Average T-cell Error compared to the Euclidean mapping.
This difference in error was reflected more strongly in the Response Error that showed a an
increase in error of ≈ 70. The example plots in Figure 4.18 clearly show well organised B-
cell repertoires under both mapping schemes as expected, and the clear organisation and
disorganisation in the T-cell repertoires between the Euclidean and Hamming schemes
respectively. These results demonstrate that under the circumstances investigated, that
Hamming distance is not a sufficient approximation for Euclidean distance, not provid-
ing enough information to promote the same structures in the T-cell repertoire as were
selected in the B-cell repertoire. These results demonstrate the fragility of proxy-based
response to the antigen-centric mapping function.
Inter-Repertoire Relationship Trends This section considers the mapping relation-

ships between the number of selected (stimulating) cells in each repertoire. The diversity
and error of the B-cell repertoire increased with the number of selected B-cells. This same
general behaviour was observed with the T-cell diversity and the number of selected T-
cells. The ATCE increased with the number of selected T-cells with 1 triggering B-cell,
although showed a small decrease with the increase in selected T-cells with N selected
T-cells. The average B-cell to T-cell mapping error showed an increase in the number
of selected T-cells with both 1 and N selected B-cells, and effect also observed with the
mapping error of the selected T-cell set. Together these observations suggest that the
selection of a single cell and integration of its clones results in a more specialised (lower
diversity and error) repertoire and mapping. Conversely, the result suggests that the se-
lection of a large activated set and integration of a few clones from each results in a more
diverse cellular perspective on the cellular and/or antigenic trigger. This observation
lends support to the oligoclonal axiom of the clonal selection strategy outlined in Sec-
tion 4.2.1. More specifically, in the context of the replacement-based resource allocation
scheme used in the assessed algorithms, selection and integration of few cells with many
clones results in a more specific and less diverse repertoire. This finding in the context
of the bi-repertoire model with antigenic and cellular triggers lends support to the same
finding in Section 4.4.2. This trend is observed poignantly in the Response Error from
the relationship results where the 1-1 configuration achieved the lowest response-by-proxy
error, increasing with the number of selected T-cells and selected B-cells. The plots in
Figure 4.20 depict this in the organisation of the B-cell and T-cell repertoires in those
configurations where a single cell is selected by antigen or cellular trigger resulting in the
integration of a set of clonal siblings (1-1, 1-N , and N -1).
128
Conclusions
This section summarises the findings of the empirical study into the Mediated Cellular
Clonal Selection Algorithm in terms of the primitives that were the focus of the study and
the expectations that motivated the study.
1. Mappings
(a) An antigen-centric mapping scheme results in the effective blind promotion

(specialisation) of the T-cell repertoire and proxied response.
(b) The quality of the mapping with regard to proxied responses in mediated clonal
selection is as good as the amount of antigen-centric information mapped be-
tween the repertoires.
(c) The remapping of the antigen-assessment Euclidean distance resulted in an
effective proxy response, whereas the approximation of the assessment in the
Hamming distance resulted in a markedly unsuccessful proxied response.
2. Relationships
(a) The number of activated B-cells and thus stimulus for the T-cell repertoire
was an important factor in response capability, resulting in a specialisation of
feature detectors for remapping.
(b) The one-to-one and one-to-many relationship configurations between the reper-
toires resulted in the better proxied responses, with more than a factor of two
decrease in capability with an increase in the number of selected B-cells.
(c) One-to-Many selection and clonal integration for a given repertoire results in
improved repertoire capability (lower error), increase repertories organisation
(lower diversity), and an improved mapping given the specialisation it provides
to the adaptive process.
4.7 Cell-Cell Recognition

This section elaborates on clonal selection by introducing a pattern recognition interpre-
tation of the Immune Network Theory and its integration into the cellular clonal selection
framework.
4.7.1 Network Clonal Selection

The Idiotypic Network Theory proposed by Jerne indicates that receptors (free or surface
bound antibody) are selected by other receptors in addition to antigen (Section 2.4.2). As
such, one may define two additional types of receptor-to-receptor interaction in additional
to conventional interaction with antigen: (1) The activation of a receptor by the idiotype
of another receptor that results in the creation of more anti-receptor receptors, (2) The
activation of a receptor that is specialised toward an antigen by another receptor that
results in the creation of more receptors for the antigen without the presence of the antigen
129
and anti-idiotype receptors for the triggering receptor. The theory suggests that the
aggregation of these low-level behaviours results in a network of receptors that may interact
with antigen and each other providing both an antigen-recognition system and the self-
regulation of immune response.
Recognition and Relationships
The two direct receptor relationships include: Receptor-Antigen: the traditional relation-
ship where receptors for the antigen (target) are created with minor variations toward
improving recognition, and Receptor-Receptor : the the idiotypic relationship where recep-
tors for the triggering receptor (target) are created (anti-receptor receptors) with minor
variations toward improving recognition. Both of these cases provide examples of direct
relationships, that of a receptor and an antigen or a receptor and another receptor. In the
first case, receptors have an implicit relationship with other receptors that also match for
the same antigen in that they compete with each other for selection by the antigen. The
second case is a lot more interesting, given the recurrent relationships that may result.
These relationships are considered in the context of how a given receptor came to be,
assuming a single matching source is responsible for receptor maturation. For example,
a receptor may match to another receptor that has been matured for an antigen. There-
fore, there is a direct relationship between the first receptor and the second, as well as a
proportional relationship between the first receptor and the second receptors antigen (for
example of both relationships see Figure 4.21).
Antigen Receptor Antigen
Matches for
antigen
Receptor Matches for Also matches
Receptor receptor for antigen
Matches for
receptor Receptor
(a) Proportional Relationship. (b) Implicit Relationship.
Figure 4.21: Examples of more complex relationships that may be formed given receptor-
receptor recognition.
Matching Function
The clonal selection algorithm results in a repertoire that directly (relationship-wise) mod-
els features of the input space. When receptors themselves are treated as input signals a
variety of representations may be used in the matching. The network theory proposes that
receptors match onto a part of the receptor that is distinct for receptors of that lineage,
although different from the receptors combining region (part of the other receptor that
does the binding). A natural implementation scheme is to assign a random or remapping
of the primary bit string to each naı̈ve receptor that, like the primary string, is inherited
130
and maturated by progeny cells. This results in the co-evolution (co-adaptation) of the
primary string for activation and the secondary string for secondary effects (for example,
see Figure 4.22).
Matching
Antigens and
other receptors
Primary String
Secondary String
Other receptors
Being Matched
Figure 4.22: Depiction of the two-string representation scheme.
A natural representation (using the same string for both) will likely result in a positive
feedback system, where the exposed antigen will be amplified by repeated simulated ex-
posures. Any regularity of the antigenic pattern that is provided on the secondary strings
(such as sub-strings or reordering) will likely have this amplification effect, proportional
to the fidelity of the mapping of the regularity. This amplification of input signal may be
useful to reinforce the acquisition of knowledge. A natural representation will continue
to promote the same instigating signal for as long as activated members are promoted to
antigen-like status.
Higher-Order Structures
A steady-state application of the approach emphasises some important properties regard-

ing the formation of relationships as networks (graphs) of activation and exposure. In this
interpretation, an antigen matches onto and activates a single receptor. The activated
receptor is copied from the repertoire and is provided as an input signal in the next cycle.
In this next cycle, both an antigen and the previously activated receptor are offered as
input signals to the repertoire. A queue-based exposure scheme is used such that the more-
recent activated receptors (instigation of relationship) are admitted, and older receptors
are discarded. Therefore, the previously activated receptor is discarded, and the newly
antigenic-activated cell is promoted to the next cycle as an input signal. The behaviour
of this scheme may be described with an example (see Figure 4.23 and Table 4.13).
If a natural mapping is used for string2, then the scheme provides a simple signal-
reinforcement scheme that provides a linear amplification (t − 1) reinforcement of signals
past. If a remapping scheme is used for string2 (such as a random string), then the scheme
provides new information in subsequent exposures that may result in the formation of a
connection between two different input signals. The simplest example is the selection of
R1a by A2. In a natural mapping, if A2 selected R1a then A1 and A2 would be the
same input signal, and R1 and R1a are likely the same receptor or clonal siblings (same
131
t-1
Antigen
Receptors
Repertoire
Figure 4.23: Example of a simple two-receptor promotion scheme.
Cycle Exposed Activated

1 A1 R1
2 A2, R1 R2, R1a
3 A3, R2 R3, R2a
4 A4, R3 R4, R3a
t At , Rt−1 Rt , Rt−1 a
Table 4.13: Exposure-activation scheme for a minimal implementation.
ancestor). In a remapping scheme (such as random), it is possible for the remapped string
to match for a receptor that is also matched for by an antigen (see Figure 4.24).
R1a R1a
R1 R2 R1
A1 A2 A1 A2
Natural Mapping Rema pping
Figure 4.24: Depiction of the example difference between a natural and remapping scheme.
The remapping example provides an example of a relationship that may form between
two different antigen (A1 and A2), facilitated by the remapping of receptors using the
two-string scheme. If A2 is withdrawn, the relationship is fostered by A1, which activates
R1, which in turn activates R1a. If A2 is withdrawn for all time, then the relationship will
deteriorate due to genetic drift5 , thus A2 provides a correcting influence to the relationship.
If A1 is withdrawn then the other half of the relationship is reinforced (R1a), and again if
A1 is not returned, then R1a may drift such that R1 does not match to it any longer. The
example relationship is unidirectional, in that the primary string of R1a is a generalisation
of A2 and R1’s secondary string. A natural extension is to consider the implications of
5
The relationship is fostered by activation (usefulness), although here genetic drift refers to the adap-
tation of receptors (their progeny) in response to activation.
132
R1a’s secondary string. If R1a matches for R1 then both R1 and R1a provides surrogates
for A1 and A2 reinforcing each other and the repertoires relationship between A1 and A2.
It requires that R1’s primary string is a generalisation of A1 and R1’s secondary string.
This network may be depicted as follows (R1a is renamed R2).
A1 R1(ps) R2(ps) A2
R1(ss) R2(ss)
Figure 4.25: Depiction of the relationship between remapped antigen and receptors (ps is
primary and ss is secondary string).
A larger queue allows more intra-antigen relationships (larger relationship structures

in general), which in turn may prolong the activation of structures, but does not facili-
tate perpetual (antigen-independent) structure formation. Therefore, Structure Durability
is defined when the exposure set size defines the sustainability of a structure in active
memory (essentially defining short-term memory), in the absence of renewed promotion.
Finally, the examples do not take into account the proliferation of selected receptors,
therefore the density concerns were subsumed with activation counts that delineated re-
ceptor persistence. Additionally, in the structures proposed the secondary strings become
surrogates for antigen strings, thus ‘take the form’ of antigen signals. This suggests that
a mapping, such as substring’s or reordering of the primary string may be easier for the
system to retrofit for such a purpose rather than a random-based mapping (assuming
similarity rather than complementarity of the mapping between receptors and antigen).
4.7.2 Recurrent Empirical Study

Aim
The aim of this empirical study is to investigate some of the primiting behaviours of the
cellular clonal selection strategy with intra-repertoire (network) interactions, specifically
the recurrent network model. Toward this end, the study had the following motivating
goals:
1. Investigate the effects on the composition and capability of the repertoire under
recurrent cellular exposures.
2. Assess a variety of cell promotion pressures in the recurrent network model.
Method
in Section 4.4.1.
133
Algorithms Algorithm 4.8 defines the recurrent variation of the Network Cellular Clonal
Selection Algorithm. Like ECCSA, two exposures occur each epoch, although unlike
ECCSA both exposures occur to the same repertoire of cells. The second exposure involves
the exposure of the Best Matching Cell (BMC) from the last antigen exposure (BM Ct−1 )
to the repertoire. A cellular exposure involves the assessment of the repertoire against the
representation in the cell, specifically using the Euclidean affinity function between the
representation as is used between an antigen and a cell (defined in Equation 4.2).
Algorithm 4.8: Recurrent Network Cellular Clonal Selection.

Input: A, T, BM Ct−1 , Nselection , Nclones , Pmutation
Output: Trs
1 Trs ← Exposure(A, T, Nselection , Nclones , Pmutation );
2 Exposure(BM Ct−1 , T, Nselection , Nclones , Pmutation );
3 return Trs ;
A Natural Mapping variation of the approach was assessed to provide a baseline

of performance referred to as RE-NCCSA-NM. The algorithm responded to each expo-
sure using the replacement mechanisms of RCCSA, and used the following configuration:
Ncells = 100, Nselected = 1, and Nclones = 5, assigning 5% of the repertoire per exposed
stimulus. This approach was called natural mapping, because indeed no mapping was
used, where BM Ct−1 were exposed directly to the repertoire, likely causing their clonal
siblings in the repertoire to respond. A remapping approach was used where each cell was
provided with two data strings: a primary string which responded to exposed stimuli, and
a secondary string which may be used as an antigenic stimulus (the two-string method
described in Section 4.7.1). Table 4.14 defines a set of four different replacement-based
promotion schemes used with the mapping variation of the algorithm. The schemes are
replacement based in that the configurations define the specific pressures applied during
the competition for limited position in the repertoire during the integration of clones. Sim-
ilarity refers to the specific representation of cells (primary or secondary) used to match
clones with cells in the repertoire, and Assessment refers to the affinity scoring (against
the primary or secondary string) used during the tournament for the position in the reper-
toire between a clone and its most similar counterpart in the repertoire. The promotion
of the antigen in the secondary representation is expected to promote similar cells in the
recurrent exposure. The promotion of the cellular stimulus in the primary representation
in the cellular exposure is expected to improve the mapping in the recurrent exposure.
R-NCCSA Antigen Exposure Cellular Exposure

Name Assessment Similarity Assessment Similarity
PP Primary Primary Primary Primary
PS Primary Primary Secondary Secondary
SS Secondary Secondary Secondary Secondary
SP Secondary Secondary Primary Primary
Table 4.14: Summary of the various replacement configurations for the mapped Recurrent
Exposure Network Cellular Clonal Selection Algorithm (RE-NCCSA).
134
stop conditions as were used for the CCSA empirical study in Section 4.4.1. Two addi-
tional diversity measures were introduced to assess the composition of the repertoire with
regard to secondary mappings. The Average Cell Mapped Diversity (ACMD) assesses
the average Hamming distance of a given cell to all the other cells in the repertoire with
regard to the secondary representation. The Average Network Cell Diversity (ANCD)
provided a similar diversity measure although treats both representations of a given cell
as one providing an indication of the composition of the repertoire independent of repre-
sentations. Both diversity measures use the same mechanism as ACD in Equation 4.7 on
their representations respectively. Finally, a measure was introduced to assess the state
of the mappings between the repertoire and recurrent cellular exposures called the Aver-
age Mapped Best Matching Cell Error (AMBMCE), that averaged the Euclidean distance
between recurrent cells secondary representation and the BMC in the repertoire for each
epoch using the same mechanism as ACE in Equation 4.4.
Results
Table 4.15 provide a summary of results for each algorithm-problem combination including
4.26b, 4.26c show the ACE, ACD, and AMBMCE respectively. Figure 4.27 provide ex-
ample plots of the response mapping and final repertoire. Results from all example plots
are taken from the end of the run with algorithm and problem configurations matching
those used during experimentation, and a random number generator seed of 1 and 5 for
the algorithm and problem respectively.
System ACD ACE ACMD ANCD AMBMCE

NCCSA x̄ σ x̄ σ x̄ σ x̄ σ x̄ σ
NM 93.218 0.248 0.019 0.011 N/A N/A N/A N/A 0 0
PP 91.974 0.413 0.09 0.025 91.777 0.46 183.751 0.669 0.184 0.026
PS 91.912 0.31 0.03 0.013 92.032 0.453 183.944 0.597 0.173 0.025
SS 91.805 0.429 0.211 0.04 91.783 0.419 183.588 0.609 0.147 0.034
SP 91.744 0.421 0.234 0.039 91.745 0.403 183.489 0.542 0.162 0.03
Sign. True True Truea Trueb Truec
a
False for RNCCSA-PP and RNCCSA-SS, RNCCSA-PP and RNCCSA-SP, RNCCSA-SS and
RNCCSA-SP
b
False for RNCCSA-PP and RNCCSA-PS, RNCCSA-PP and RNCCSA-SS, RNCCSA-PP and
RNCCSA-SP, RNCCSA-SS and RNCCSA-SP
c
False for RNCCSA-PP and RNCCSA-PS, RNCCSA-PS and RNCCSA-SP, RNCCSA-SS and
RNCCSA-SP
Table 4.15: Summary of results for Recurrent Network Cellular Clonal Selection Algorithm
(R-NCCAS) on ACSP-10
135
(a) Average Cell Error on ACSP-10. (b) Average Cell Diversity on ACSP-10.
(c) Average Mapped BMC Error on ACSP-10.
Figure 4.26: Box-and-whisker plot’s from the Recurrent NCCSA Empirical Study.
(a) Response Mapping plot for R-NCCSA-PS, (b) Repertoire plot for R-NCCSA-PS, shows pri-
shows A, BM Ct1 a, BM Ct−1 , BM Ct1 b from left mary and secondary representations and antigen
to right, for an epoch. from left to right.
Figure 4.27: Example plots from of the response mappings and final repertoire from R-
NCCSA-PS on ACSP-10.
136
Analysis
Promotion Trends This section considers the results of the recurrent network algo-
rithm in the context of the trends for the different promotion schemes. The natural
mapping achieved the best results with regard to ACE as expected, showing a AMBMCE
of zero suggesting that cells mapped onto themselves or clonal siblings with no error on
average by the end of the run, providing an ideal (in the mapping sense) case for compar-
ison. The various promotion schemes for the mapped recurrent algorithm showed little
comparative difference between the three diversity scores: ACD, ACMD, and ANCD. The
ACD scores showed a slight decrease with the remapping approaches compared to the nat-
ural mapping. The similarity between ACD and ACMD scores suggests that organisation
of the information in the primary and secondary representations (pseudo-repertoires) was
generally equivalent across the promotion schemes. The telling results were provided in
the error measures. The two promotion schemes that promoted the primary representation
achieved a much lower ACE as expected than those that promoted the secondary represen-
tation against the antigen. Interestingly the configuration that promoted the secondary
representation against the cellular exposure (PS) performed better than the promotion
of the primary string on both exposures (PP). The error results for the recurrent cellu-
lar exposures were poor compared to error scores for viable ACSP (< 0.1), where the
promotion of the secondary string against antigen exposures resulted in relatively lower
AMBMCE than the promotion of the primary string. A general observation of the state
of the repertoire and exposure mappings during the execution of general runs reviled that
consistent intra-repertoire mappings were unstable across the various promotion schemes
(for example see Figure 4.27). This is reflected in the poor mapping error scores achieved.
Curtailing Instability The complexity of the approach makes analysis difficult al-
though the observed behaviour may be explained by the instability of the general approach.
This instability is likely the result of the competition between similar representations for
antigen and for cells. For example in the case of PP, the promotion of primary strings
in the cellular exposure may be considered the promotion of random secondary strings,
with no explicit pressure of adding meaning to the secondary strings of antigen-selected
BMC’s. This is addressed in PS by the cyclic promotion of secondary strings which are
expected to provide pattern-matching loops which counter some of the instability effective
ACE. The results for secondary string promotion during antigen exposures suggest that
the promotion of BMC’s for antigen in the repertoire via proxy (t+1 cellular exposure),
in particular R-NCCSA-SP, is insufficient for the repertoire to address the ACSP. This
suggests that the de-coupled promotion of antigen-BMC within the repertoire may result
in conflicting competition between primary and secondary representations.
137
Conclusions
This section summarises the findings of the empirical study into the Recurrent Network
Cellular Clonal Selection Algorithm in terms of the primitives that were the focus of the
study and the expectations that motivated the study.
1. Primitives
(a) The Recurrent NCCSA provides a viable approach for investigating the subtle
effects of intra-repertoire remapping of exposure signals and the subtleties of
competition in such a mapping.
2. Behaviours
(a) The recurrent exposure of cells via a natural mapping provides a reinforcement
of antigenic signals
(b) The recurrent exposure of remapped cells requires careful consideration of the
competition between representations to avoid conflicts in such competition.
4.7.3 Dual Exposure Empirical Study

Aim
The aim of this study is to investigate the antigen-driven formation of intra-repertoire

structures as outlined in Section 4.7.1, and Figures 4.25 and 4.25. Specifically, this em-
pirical study is concerned with further investigating the subtle application of replacement
pressure as was considered in the recurrent model, although toward the formation and
promotion of antigen-dependant rather than cellular-dependant structures. Toward this
end, this study has the following goals:
1. Assess the composition and capability of the repertoire under a dual-exposure model.
2. Investigate the formation of antigen-dependant high-order structures via the appli-

cation of intra-repertoire promotion schemes.
Method
in Section 4.4.1.
Algorithms Dual exposure refers to the way in which the repertoire interacts with
the antigen, specifically two at a time (requiring the ACSP to be comprised of an even
number of CSP). Dual exposure may be considered a variation on the linear Cellular
Exposure Regime considered in Algorithm 4.2 where the first antigen in the pair is always
exposed to the repertoires primary string representation, and the second antigen to the
secondary string representation (see Algorithm 4.9). Such exposure governs the selection
of the BMC returned as a response from the repertoire to satisfy the requirement of the
exposure represented.
138
Algorithm 4.9: Dual Exposure Network Cellular Clonal Selection.
Input: A1 , A2 , T, Nselection , Nclones , Pmutation
Output: Trs
1 Trs ← Exposure(A1 , T, Nselection , Nclones , Pmutation );
2 Trs ← Exposure(A2 , T, Nselection , Nclones , Pmutation );
3 return Trs ;
As with the recurrent approach (defined in Algorithm 4.8) a series of configuration

schemes are considered that provide subtle adjustments to the replacement competition
of aggregated clonal sets against the repertoire. Unlike the recurrent approach, these
configurations are concerned with the effects of varying or decoupling the context for
similarity and assessment used during replacement (see Table 4.16). In this case, Similarity
refers to the string representation (primary or secondary) used to locate the most similar
repertoire member for a clone to compete with, and Assessment refers to which antigen the
representation is assessed against (primary antigen or secondary antigen). As mentioned,
the first exposure is always assessed against the primary representation, and the second
against the secondary representation.
DE-NCCSA First Exposure Second Exposure

Name Assessment Similarity Assessment Similarity
PP-SS Primary Primary Secondary Secondary
PS-SP Primary Secondary Secondary Primary
SS-PP Secondary Secondary Primary Primary
SP-PS Secondary Primary Primary Secondary
Table 4.16: Summary of the various replacement configurations for the Dual Exposure
Network Cellular Clonal Selection Algorithm (DE-NCCSA).
stop conditions as were used for the R-NCCSA empirical study in Section 4.7.2.
Results
Table 4.17 provide a summary of results for each algorithm-problem combination including
Mann-Whitney U statistical test was calculated pair-wise for all algorithms. Figure 4.28
shows the Average Cell Error (ACE) for the four different dual exposure configurations.
Figure 4.29 provides example plots of the four schemes at the end of a run on ACSP-2.
The algorithms and problem use a random number generator seed of 1 and 4 respectively.
The configuration of each algorithm was reduced to Ncells = 20 to ensure a proportional
distribution of cells in the repertoire. The smaller ACSP-2 was chosen to clearly high-
light the formation and/or lack of formation of bi-antigen structures in the primary and
secondary representations.
139
System ACD ACE ACMD ANCD
NCCSA x̄ σ x̄ σ x̄ σ x̄ σ
PPSS 89.213 0.647 0.02 0.02 89.069 1.001 178.282 1.267
PSSP 89.231 0.678 0.018 0.016 89.382 0.766 178.613 1.09
SPPS 89.289 0.705 0.211 0.05 89.551 0.751 178.839 1.233
SSPP 89.301 0.876 0.214 0.041 89.581 0.835 178.882 1.382
Sign. False True Falsea False
a
True for DE-NCCSA-PPSS and DE-NCCSA-SSPP
Table 4.17: Summary of results for the Dual Exposure Network Cellular Clonal Selection
Algorithm (DE-NCCSA) on ACSP-10.
Figure 4.28: Box-and-whisker plot of the Average Cell Error for the DE-NCCSA schemes
on ACSP-10.
140
(a) DE-NCCSA-PPSS. (b) DE-NCCSA-PSSP.
(c) DE-NCCSA-SPPS. (d) DE-NCCSA-SSPP.
Figure 4.29: Example repertoire plots of the four Dual Exposure NCCSA configurations
on ACSP-2 showing primary and secondary representations and antigen left to right.
141
Analysis
Promotion Trends This section considers the promotion trends for the four different
configuration schemes of the dual exposure algorithm, The results show little difference
with regard to the collected diversity measures. Specifically there was no significant differ-
ence in the ACD and ANCD, and generally no difference in ACMD. As with the empirical
study into recurrent network the effect of the promotion schemes was on the repertoire ca-
pability. For both schemes where the exposed antigen is promoted via assessment (PPSS
and PSSP), the ACE demonstrated viable results against the ACSP-10. This suggests
that regardless of whether the cells are competing with repertoire members for the same
primary or secondary antigen, replacement tournaments must be decided based on the
representation used to satisfy the antigen. Interestingly, this effect can be enhanced by
competing with repertoire members with similar representations that are not the rep-
resentation used for selection and assessment. Specifically, PSSP resulted in a slightly
(although significantly lower) ACE than PPSS. The example repertoire plots provided in
Figure 4.29 highlight the behaviour of this class of NCCSA. specifically, the plots suggest
that the promotion via assessment against selecting antigen results in the specialisation
toward the respective antigen, although decoupled structures. In both cases where re-
placement competition occurred against the complementary antigen (SPPS and SSPP),
the plots strongly suggest that such structured were formed in the repertoire (top of each
plot for SPPS and SSPP). The ACE results recorded suggest that such structures were of
a much lower capability than the promotion of selecting antigen, which can be explained
by the reversal of the specialisation of the representations observed clearly in the example
plots (primary string to second antigen and secondary string to first antigen as opposed
to the reverse case expected by the exposures). Interesting this was not reflected in the
ANCD. This suggests that the cross-promotion can form such structures, although the
selective pressure of the antigen alone was insufficient to promote improved speciality for
exposed antigen.
Conclusions
This section summarises the findings of the empirical study into the Dual Exposure Net-
work Cellular Clonal Selection Algorithm in terms of the primitives that were the focus of
the study and the expectations that motivated the study.
1. Primitives
(a) The dual exposure model provides a viable complementary approach to recur-
rent for investigating promotion pressures in NCCAS.
2. Behaviours
142
(a) Antigen-dependant cross-exposure structures can be formed within the reper-
toire by biasing the integration of clones toward the complementary parts of
the structure.
(b) Such structures cannot be formed by only biasing the similarity pairing alone,
requiring a biasing in the assessment used for competitive tournaments in sim-
ilarity pairs.
4.8 Chapter Summary

4.8.1 Paradigm Review
The Cellular Clonal Selection Paradigm is a re-definition of the current state of clonal
selection algorithms. This rephrasing is differentiated from the current state of the field in
the following ways: (1) the separation of the concerns into a system (cellular algorithm)
and an environment (antigenic environment), (2) the explicit definition of the computa-
tional properties of clonal selection as an adaptive knowledge acquisition strategy, and
(3) the restriction of the concerns of the clonal selection systems and environments to the
cellular-level. The cellular level is defined as the maintenance and operation of clonal se-
lection on a repertoire of discrete cells, not limited to the antigenic, molecular, and cellular
interactions that may occur in such a repertoire.
4.8.2 Paradigm Trends and Findings

The following summarises the important principles and findings from the definition and
investigations into the Cellular Clonal Selection Paradigm:
Primitives
1. Cellular : Increasing the proportion of the repertoire dedicated to an antigen and/or

the number of clonal trials, increases the specialisation and therefore the capability
of the repertoire for an antigenic environment.
2. Replacement: Repertoire-wide integration of an aggregated clonal set using affinity

tournaments between similar cells, with sibling exclusion provides a controlled spe-
cialisation of a footprint of concurrent redundant perspectives in the repertoire for
each antigen.
3. Degenerate: The strong selection of activated cells is exploited by cellular clonal

selection to constrain the inherent polyclonal activation of the repertoire to each
antigen exposure. The explicit aggregation of repertoire responses is an artefact
of the exposure paradigm constraining the polyclonal activation of the degenerate
repertoire, an important principle not limited to DCCSA.
Extensions
143
1. Spatial : Constraining the integration of clones to the neighbourhood of the antigen
selected cells provides a localising spatial pressure for areas of a spatial repertoire to
take responsibility for specific antigenic signals.
2. Mediated : The use of cell casts to mediate responses to antigen relies heavily on
an antigenic mapping between the casts and an oligoclonal (small founding set)
relationship between the feature detector and concept formation tiers.
3. Network : Intra-repertoire cell interaction requires careful management of the pres-

sures that govern the differential allocation of resources. The cross promotion can
facilitate the formation of antigen-dependant structures across multiple exposures
although at the expense of specificity.
4.8.3 Integration
The Cellular Clonal Selection Paradigm defined and investigated in this chapter provides
a bedrock in understanding of clonal selection as an adaptive strategy in the context of
an antigenic exposure paradigm specialised in the colour space domain. The features and
follow-on information processing characteristics set the investigated approaches apart, the
relative merits of which are considered in the context of application problem domains in
Section 8.3 and Section 8.4. The chapters that follow build upon this bedrock providing
a Tissue (Chapter 5) and Host (Chapter 6) constrained perspective on clonal selection as
an adaptive knowledge acquisition strategy, that collectively with the Cellular perspective
provide an integrated hierarchical framework for clonal selection algorithms (Chapter 7).
144
Chapter 5
Tissue Clonal Selection

This chapter exploits the principles and findings of the cellular paradigm by taking them
for granted, instead investigating clonal selection where the cellular paradigm is a prim-
itive component in a broader Tissue Clonal Selection Paradigm. Section 5.2 reviews the
acquired immune system from the perspective of the tissues that provide the structure and
manage the function of the lymphocytes central to the clonal selection theory. Section 5.3
considers an abstraction of the reviewed physiology and tissue-based immunology and the
basis of the tissue paradigm. Most importantly in this abstraction are the spatial and tem-
poral considerations of the discrete exposure of repertoires of cells to information, and the
tissue architectures that provide base patterns for the design of inspired algorithms and
implementations. The paradigm is realised as the investigation of Tissue Clonal Selection
Algorithms in the context of an Infection Antigenic Exposure Problem, that provides the
basis for empirical investigation in the remainder of the chapter. Tissue-based immunology
strategies for organising information are defined and empirically investigated in an colour
space specialisation of the infection problem, resulting in a series of algorithms: the Min-
imal, Recirculation, Homing, and Inflammation Tissue Clonal Selection Algorithms. The
investigation of these algorithms confirms expectations regarding the natural organisation
of information based on the regularity and consistency of exposure, as well as important
findings as to how the defined strategies may be exploited when such regularities and
consistencies are not known a priori.
5.2 Physiology of Lymphocyte Migration

The white blood cells and immunoglobulin within a host that comprise the recognition
and response component of the acquired immune system are mobile. There are pools of
lymphocytes that recirculate the blood stream and the lymphatic system. More interest-
ingly there are lymphocytes that selectively home to tissues close to where the cells were
created and there are pools lymphocytes that are recruited to sites of infection and in-
flammation. The lymphoid tissue of the lymphatic system provides the structural scaffold
for the mobility of these cells and related molecules, although interestingly the migratory
145
behaviour is controlled at the finest level: through localised molecules and receptors on
the cells themselves. The migration of cells involved in the immune system is highly com-
plex and is still not completely understood. This section reviews some superficial features
of lymphocyte migration in the immune system with the interest of extracting principles
for use in clonal selection-based Artificial Immune Systems. Lymphocyte migration is
reviewed in the context of the human and related mammalian immune systems, providing
perspectives of cell movements from the lymphatic system, movement types, cell types,
and cell classes.
5.2.1 Lymphatic System

The lymphatic system is a complex collection of lymphoid organs integrally related to the
functioning of the immune system and responsible for the transport and filtering. Lymph
is a clear bodily fluid that surrounds all tissue that forms when proteins and cells leak
out of blood carrying venules and capillaries. The lymph carrying capillaries flow uni-
directionally draining the lymph back to lymphoid tissues that make up the lymphatic
system. Also carried in this lymph are antigens that may have entered the organism. The
lymphatic system acts as a secondary circulation system (to the cardiovascular circula-
tory system) transporting lymphocytes between lymphoid organs, and carrying antigen
to lymphoid organs for interaction with lymphocytes. There are two types of lymphoid
organs (1) the primary or central tissues which are the sites of lymphocyte formation (the
thymus and bone marrow), and (2) the secondary or periphery lymphoid organs where im-
mune response take place (the spleen, lymph nodes and gut-associated lymphoid tissues).
One may consider the rest of the bodies’ tissues as tertiary lymphoid tissues which nor-
mally only contain few lymphoid cells, although on infection and inflammation, manage
to recruit many immune cells.
Central Lymphoid Tissues
Primary lymphoid organs consist of the thymus and the bone marrow, and are respon-
sible for differentiating stem cells into pre-immune cells (pre-B-cells and pre-T-cells) in
a process that is believed to be independent of antigenic stimulation. The bone marrow
provides a micro-environment for the production of blood cells. B-cells migrate to sec-
ondary lymphoid organs, whereas T-cells migrate to the thymus. The thymus provides an
environment for the further antigen-free maturation of T-lymphocytes. Cells proliferate
and differentiate in a process involving negative and positive selection where the majority
of produced T-cells never leave the thymus. Those cells that do survive (less than 5%
per day) may join the recirculating lymphocyte pool, or migrate to secondary lymphoid
tissues. There are two types of of primary lymphoid tissues, as follows:
• Bone Marrow : Tissue located inside large bones responsible for the production
of many different blood cell types not limited to lymphocytes (white blood cells).
Provides an environment for differentiation of naı̈ve B-cells.
146
• Thymus: An organ located behind the sternum, providing an environment for the
differentiation of naı̈ve lymphocytes from bone marrow. These lymphocytes differ-
entiate into T-lymphocytes, which involves a negative selection maturation process.
Peripheral Lymphoid Tissues
The secondary lymphoid tissue is responsible for collecting antigens at points in the or-
ganism where they may enter, and facilitate their exposure to recirculating lymphocytes.
For example, the lymph nodes are primarily responsible for filtering the lymph for anti-
gens, the spleen is responsible for filtering the blood for antigens, and the tonsils for the
respiratory system. Goodnow puts this concisely suggesting that the “. . . central function
of secondary lymphoid tissues is filtering: collecting blood-borne, lymph-borne, or mucus
membrane antigens and holding these so that they can be surveyed by immune cells before
being destroyed ” [181] (page 6). The secondary lymphoid organs are as follows:
• Tonsils: A collection of lymphoid tissue on the side of the throat responsible for pro-
viding lymphocyte access to and protection of the respiratory system from antigen.
• Lymph Nodes: Small gland like structures found throughout the body that filter the
lymph for foreign antigen material, which are then presented to lymphocytes and
other immune system cells.
• Peyer’s Patches: A collection of lymphoid tissue found in the lowest section of

the small intestine. There are numerous instances of these patches in the intestine
and they act like lymph nodes, and provide centres for lymphocytes protecting the
gastrointestinal tract.
• Spleen: An organ in the upper abdomen, responsible for the destruction of old red
blood cells. It is the only lymphoid organ that crosses the blood stream, and thus
provides a site where lymphocytes can interact with antigens carried in the blood.
• Lymphatic vessels: Vessels throughout the body that carry lymph. They are respon-
sible for transporting lymph from tissues to the blood (vascular circulatory system)
and to the lymph organs.
In addition, the secondary lymphoid tissue provides a suitable micro-environment for

the development and maturation of an immune response, such as the formation of Germinal
Centres (GC’s) for B-cell clonal expansion and affinity maturation. Germinal Centres are
little understood dynamically generated regions in lymphoid tissue where activated B-
lymphocytes clonally expand, undergo hypermutation, and ultimately differentiate into
plasma and long-lived memory B-lymphocytes [37, 387]. GC’s are founded by a small
number of activated B-cells (oligoclonal), which are in turn co-stimulated by helper T-
cells [275]. It is believed that there are repeated rounds of selection, expansion, and
mutation of B-cells within the GC. The majority of the produced B-cells may be plasma
cells and have a lower affinity than the founding cells, although a few higher affinity clones
are preferentially selected to survive as plasma or memory cells [277, 292]. Those cells
147
that are worse die due to the lack of positive selection. GC-like structures may also occur
outside of lymph nodes, such as sites of infection and allergic inflammation [182].
The geography of the lymph nodes and other secondary lymphoid tissues plays an
important structural role for lymphocyte selection and response, for example Zinkernagel,
et al. suggests that the “. . . secondary lymphoid organs present antigen optimally and
enhance the chances of specific antigen encounter and specific cell interactions.” and that
“. . . these requirements render chance encounters of antigen by lymphocytes and activation
elsewhere extremely inefficient and of no biological relevance.” [446] (page 201). It is
likely that the organisation and function of the secondary lymphoid organs has been
optimised for recirculating lymphocytes to efficiently detect and remove pathogens from
the organism, for example Fu and Chaplin suggest that “. . . secondary lymphoid organs
are thought to be organised into structures that optimize cellular interactions that support
the efficient removal of unwanted pathogens” [156] (page 400). Finally, in addition to
providing a scaffold for lymphocyte recirculation, the secondary lymphoid tissues possess
a high regenerative capacity. When circulation routes are disrupted or severed, the tissue
is able to adaptively regenerate a connection between the effected nodes [309].
Summary
There are no lymphocytes or lymphocyte recirculation without the lymphatic system.

The tissues are responsible for the development and management of the cells throughout
their life cycle. The following summarises the lymphatic system’s governing role over
lymphocytes:
1. Formation: Provide an optimal micro-environment for the production and matura-

tion of lymphocytes.
2. Presentation: Collect diverse populations of lymphocytes into organ systems that

drain antigens from entry points into the organism.
3. Regulation: Regulate interactions of different classes of lymphocytes in drainage

organ systems, such as the arrangement of B-cells and T-cells.
4. Dissemination: Disburse effector elements of the immune response throughout the

organism, and disseminate and amplify the immune response systematically through-
out the lymphatic system.
For more information regarding the anatomy and physiology of the human lymphoid
system see any current anatomy and physiology text on the subject, such as [280]. Other
references used include lymphoid organ summaries by de Castro and Timmis [109] (page
71), Anderson [19], and Swartz [377]. See Andrian and Mempel for a review of lymph
node anatomy and physiology in the context of cell migration [411].
5.2.2 Lymphocyte Mobility

Cell movement is the basis of immunology. It is involved in inflammation, differentiation,
adhesion, recruitment, and required for cellular and molecular interactions. The recir-
148
culation of lymphocytes was confirmed experimentally almost 50 years ago by Gowans
and colleagues [183, 184], although the study of the topic is often neglected in favour of
cell interactions [442, 322]. Interestingly much of the modelling work with populations of
lymphocytes does not take into account the spatial heterogeneity of the lymphatic system.
The adhesion properties of cells plays a critical controlling role in the movement of
lymphocytes (see Anderson for a treatment [19]). Adhesion is used by cells to crawl
through tissues, and is used by activated cells to stick to the molecule that activated them.
Adhesive receptors (so-called ‘homing receptors’) and chemicals are little understood and
are a topic for intense study. The adhesive characteristics of a cell are believed to be
selected for and differentiated along with cell antigenic receptor characteristics. These
adhesive interactions provide bottom up control at the finest level [324], believed to be
the basis for homing, recruitment and may ultimately control the extent and scope of the
immune response [416]. A procedure or series of adhesive-based decisions have to be made
for a lymphocyte to be recruited into tissues, in particular from recirculation in the blood
into lymphoid tissues. This process is called the multi-step extravasation or molecular
regulation procedure, which has the following four steps: (1) primary cell adhesion, (2)
rapid cell activation, (3) activation dependant arrest, and (4) diapedesis (the movement
into surrounding tissue).
Movement Types
Physically, there are two ways lymphocyte cells can move: crawling on tissues and flowing
between tissues. In crawling, cells use chemical receptors to adhere to their surroundings
and use this method to slowly migrate through tissues, a place where they spend most
of their time given the slow pace of movement. The second mode is movement in fluid
space such as in blood or lymph where cells are capable of moving a lot faster, covering
great distances within the organism. The process of lymphocyte migration was originally
considered to be random, although it is now known that this is not necessarily the case.
Lymphocytes may be preferentially recirculated, and are able to home in and target specific
tissues. In addition, memory lymphocytes show different migration behaviour to naı̈ve
lymphocytes. For example, memory T-cells preferentially migrate to non-lymphoid tissues.
If a memory cell was created in a lymph node near the skin, the cell will preferentially
migrate into skin tissue near the lymph node, if created in a lymph node near the gut,
it will migrate to neighbouring gut tissue. Naı̈ve T-lymphocyte cells are activated by
dendritic cells and educated as to the chemical properties of the site of infection. These
prepared T-cells recirculate around the host organism in search of their cognate antigen,
and upon detecting the chemical properties of the site of infection, home into the tissue.
The directed trafficking behaviour is called T-lymphocyte homing [67, 66, 324, 416], and
the information which controls the where T-cell traffic is expressed as receptors on the
surface of the cell for the chemical properties of the site of infection [139, 343].
The immune system maintains a pool of recirculating lymphocytes that cycle around
the blood and the lymphatic system. The pace of recirculation is high, the number of
lymphocytes entering the blood from the lymph each day is 10 times the size of the
149
recirculating lymphocyte pool [19]. The number of lymphocytes recirculating at one time
may be anywhere from 1%-2% of all lymphocytes in the body in young adult animals
[442, 371]. Lymphocytes may only stay in circulation in the blood for about 30 minutes.
Trepel provided a seminal, although outdated extrapolation of the number and distribution
of lymphocytes in man [398]. Such numbers may be useful for a general guide, as follows:
2.2% in the circulating blood, 41.3% in the lymph node and tonsils, 15.2% in the spleen,
4.3% in the gut-associated lymphoid tissues, 10.9% in the thymus, 10.9% in the bone
marrow, and about 15% in other tissues. One may summarise the primary types of
lymphocyte migration, as follows:
• Trafficking: The directed (non-random) movement of cells from tissues, blood, or

lymph. May refer to a cells trafficking route as it homes to a specific region in
the body. For example: pre-B-cells and pre-T-cells migrate to secondary lymphoid
tissues to further differentiate and mature.
• Recirculation: The movement of lymphocyte cells around the body from lymphoid
tissue, to the blood, to the lymph, and back to lymphoid tissue to repeat the pro-
cess (circulation or rolling lymphocytes). Naı̈ve and memory cells are the main
recirculating lymphocyte types.
• Recruitment: Accumulation (sequestration) of cells such as a site of infection or tis-

sue damage, such accumulation may occur through chemotaxis (directed movement
in response to a chemical gradient).
• Homing: The directed (preferential tendency) of lymphocytes activated in a particu-

lar region of the body, to return to that part of the body (localisation). May refer to
the arrival of lymphocyte to lymphoid or non-lymphoid tissue from the blood stream.
Also called tissue-selective trafficking. Those cells with a memory of where they were
differentiated may localise back to these regions after a period of recirculation.
• Stationary: These are cells that selectively arrest their movement, or do not move
from the location of their differentiation. For example in the recruitment of cells,
the differentiation of cells such as plasma B-cells created in germinal centres that
release large amounts of antibody, and sentinel effector cells.
Lymphocyte Types
There are both circulating and non-recirculating populations of B- and T-cells. Although
the behaviours of both cell types are tightly interrelated, both have differing migration
behaviours during their development and life cycle. The general migratory life cycle of
T-lymphocytes is as follows: pre-T-cells migrate from bone marrow to the thymus for
maturation, surviving T-cells may migrate to tissues and becomes sentinel cells. Other
cells may recirculate and seek activation in secondary lymphoid tissue. Memory T-cells
preferentially migrate to non-lymphoid tissues. For B-lymphocytes: pre-B cells migrate
from the bone marrow to secondary lymphoid tissues. Some cells will be activated by
150
antigen in the secondary tissue and differentiate into plasma and memory cells. Other
cells will be activated and migrate to the spleen before differentiating.
There is commonality in the development of both B- and T-cells, specifically in terms
of the generalised classes of lymphocyte they may differentiate into. These include the
classes: naı̈ve, effector, and memory. Naı̈ve cells are untested lymphocytes that seek a
potential cognate antigen. Effector and memory cells result from the union of naı̈ve cells
and antigen. Effector cells, such as Helper T-cells and plasma B-cells remain in lymphoid
tissue. Memory cells make up the majority of the recirculating pool and continue to
circulate between the bloodstream and the lymphatic system, at a higher rate than naı̈ve
cells. One may define the migratory behaviours of the generalised lymphocyte classes or
casts, as follows:
• Naı̈ve Lymphocytes: Recirculate between blood and lymphoid tissue, primarily in-
volved in responding to antigen presented in lymph nodes, differentiating into ef-
fector and memory cells. Relatively homogeneous in their recirculating behaviour.
Naı̈ve cells compete with each other for activation and contribution into the memory
recirculating pool.
• Effector Lymphocytes: Typically do not recirculate, stay at the site of differentiation,

such as plasma B-cells, which differentiate from naı̈ve B-cells in lymph nodes or the
spleen.
• Memory Lymphocytes: Recirculate around blood and lymphoid tissue, but also ex-
travasate to other non-lymphoid tissues. They are heterogeneous in their recircu-
lation behaviour, with restricted and selective recirculation circuits. They home to
areas where they are most likely to encounter, or re-encounter their cognate antigen.
The number of memory cells is maintained within a moderate range during adult
life.
Cell Mobility as a Strategy
The anatomy and physiology of the immune system may be thought of a defence strategy
for the organism, and cell mobility is an integral part of that strategy. Picker and Butcher
propose such an evolved strategy where the solution to a complex antigenic environment
is “. . . to compartmentalize the principle functions of the lymphoid system into discrete
organs and tissues in the body, and to connect and unify these organs through the operation
of an elegant system of targeted lymphocyte trafficking and recirculation.” [324] (page 562).
Some resources of the immune system are fixed in position and distributed throughout
the body such as lymphoid tissues, and so called sentinel T-cells. Draining lymph into
lymphoid tissue provides a way to localise antigen. Keeping the majority of the immune
responses in movement (the detectors and effectors) provides a patrolling approach to
rapidly deploy resources to wherever they are needed. The immune system has to produce
many cells to detect unknown antigen, and must collect antigen in such a way that rare
detector cells are given a chance to detect them. It must provide a micro-environment
for controlled proliferation and differentiation, and disperse effectors to where they are
151
needed. Further, Rosen, et al. consider a recirculating population as an efficient approach
in the context of alternative designs where “Recirculation contributes to the efficiency of
peripheral immune responses by maximising stochastic probabilities of a productive meeting
between antigen and its cognate T-cell receptor ” [339] (page 161). This tissue-recirculation
adaptive strategy may be considered to have the following general properties:
• Movement: Different cells and cell types have differentiated migration potentials (so-
called ‘differential migration’). The lymphatic system provides a pathway, which al-
lows lymphocytes to recirculate between blood, peripheral tissues, and lymph nodes.
• Balance: The so-called stirring or mixing effect of the recirculating lymphocyte pool
facilitates the survival of the fittest, that is the survival of the most appropriate
clones in the repertoire. The segregation in the repertoire prevents competition
between unrelated lymphocyte subsets. Tissues selectively facilitate the segregation
of lymphocytes to where they are needed. Such a balance between recirculation and
homing may be critical to lymphocyte homoeostasis [66].
• Coverage: A repertoire of cells is distributed and rotated in an attempt to expose rare

(specialised) cells to as many opportunities for activation as are available. This may
be considered the maximisation of coverage, where the rolling population seeks to
maximise the probability of specialised cells being exposed to their cognate antigen.
Large-scale migration allows a wide repertoire of specific lymphocytes to exist at
a very low frequency, yet still function efficiently. An example of this is T-cells.
Perhaps one in ten thousand T-cells released from the thymus will have a receptor
that can detect a given antigen. This means that statistically at least ten thousand
inappropriate cells must encounter an antigen before the right one does [19].
• Surveillance: Lymphocytes may be thought of as patrolling the organism, seek-

ing cognate antigen or recruitment. They may also thought of as monitoring the
organism for change to self for self-antigen such as tumours in what is known as
immunosurveillance.
• Amplification: Recirculation permits recruitment, the dynamic reallocation (local-

isation) of specialised resources to locations where they are needed. Further, the
system is capable of further specialising the recruited resource and disseminating
that information.
• Dissemination: The pathway provided by the lymphatic system provides a highway

to rapidly distribute information such as the immune effector cells, antibodies, and
long lived lymphocyte cells that retain a refined memory of an antigenic exposure
for a future rapid response.
• Alternative: It provides a distributed mobile defence strategy and is different to

other subsystems of the body such as the nervous system, which is distributed but
structurally fixed. Further, if the repertoire did not recirculate then each lymph
node would be an isolated (island) repertoire. Thus, each such isolated repertoire
152
would either have to be as large as the entire recirculating pool, or a smaller pool
resulting in less antigenic coverage.
Problems and Changes to Migration
The rate and nature of lymphocyte migration does not remain constant, and in fact, it
may be desirable in certain circumstance to manipulate the recirculation behaviour of
lymphocytes. Alternatively, the recirculating behaviour of lymphocytes that provides a
dynamic defence may also facilitate significant problems for the host organism. During
an immune response, the lymph nodes may rapidly recruit cells from blood, temporally
blocking them from leaving the tissue. This in addition to clonal expansion causes the
nodes to swell in size and become sore. This swelling of the glands is a common sign of ill-
ness. It highlights the point that when an antigenic stimulus is detected, it has a profound
effect on the entire system, affecting production and circulation behaviour of lympho-
cytes. There are some diseases that disrupt and decrease the migration and homoeostasis
of lymphocytes [20], thus it may be desirable for the system to temporally increase the
rate of cell migration. Some approaches may include increasing blood flow, increasing the
expression of adhesive chemicals, and manipulating the tissues where immune responses
take place. Alternatively, there are diseases where it is desirable to decrease the recircu-
lation of lymphocytes. Examples include when tissues are transplanted and autoimmune
diseases. In both cases the immune cells seek to destroy healthy tissue, thus strategies
focus on affecting the recruitment of cells to these areas.
The lymphatic system may facilitate the spread of disease in the body. For example,
cancers and tumour cells are able to spread quickly throughout the lymph nodes of a
host. Bacteria can use the lymphatic system to disseminate throughout the body. An
example mentioned by Andrian and Mempel is the bacterium Yersinia pestis one of the
most devastating bacterial pathogen in human history (the cause of the bubonic plague)
[411]. It manages to get in to tissue through flea bites and moves to the lymphatic system
where it proliferates, quickly overwhelming lymph nodes and spreading throughout the
body. Additionally widely cited examples include the human immunodeficiency virus
(HIV) that can lead to the condition of acquired immunodeficiency syndrome (AIDS),
Mycobacterium that causes diseases including tuberculosis and leprosy, and the bacterium
Bacillus anthracis that causes the disease Anthrax.
5.3 Abstract Tissue Paradigm

Section 5.2 demonstrated that the tissues that house lymphocytes govern their life cycle,
and that together they form a high-level strategy for partitioning and managing the sys-
tems interaction with antigen. This section abstracts and elaborates on this tissue-centric
adaptive strategy and proposes a Tissue Clonal Selection Paradigm that takes the clonal
selection of the Cellular Paradigm for granted, and focuses on the inter-tissue interaction
and information sharing provided by bottom-up cell migration mechanisms.
153
5.3.1 Model Components
Architecture Components
Discretising the repertoire refers to transitioning from an abstracted clonal selection sin-
gle and self-contained repertoire, to multiple of such repertoires. The direct approach
proposed in this work is to focus on the lymphoid tissue as the nodes in a distributed
repertoire and the vascular system (the veins and such that transport the lymphocytes
between the lymphatic tissues) as the edges of the network. This graph-based network for-
malism provides an architecture that separates the transport of lymphocytes on the edges
of a network structure from the cellular-based information processing (antigen-lymphocyte
reactions) that occur in the nodes of that network. Refer to Figure 5.1 for a conceptual-
isation of this relationship. A ‘tissue’ is a generic node in the network topology model,
and may be specialised to various types of tissue with differentiated functional roles. A
tissue contains a repertoire of lymphocytes and facilitates antigen-lymphocyte interaction.
Therefore, a tissue node is a reformation of what might be considered the repertoire in the
Cellular Paradigm. Tissue nodes may be functionally specialised after the general tissues
classes in the lymphatic system, such as primarily, secondary, tertiary lymphoid, and the
germinal centre (Section 5.2.1).
Cellular Components
Lymphocyte cells are discrete information packets that collectively in a repertoire or reper-
toires are an internally generated model of the external domain. Antigen may also be
considered information packets and/or direct interaction with the domain, themselves be-
ing of external origin with regard to the lymphocyte model. The cellular components are
the interest of the Cellular Paradigm, although this section exploits the rise in the level
of abstraction provided by tissue to re-define cells from an amorphous information packet
into functional classes outlined in Section 5.2.2. The classes define different information
process behaviour, particularly with regard to the mobility of the cells within the network
topology architecture.
Movement Mechanisms
Movement operators define the fine-grained control of lymphocyte migration. Without

movement, the discretised repertoire may be considered multiple independent instances of
some variant of the Cellular Paradigm. Movement may be considered as a decision process
at a per-lymphocyte level involving the interaction of the lymphocyte (lymphocyte class),
and its locality (tissue type and location such as a germinal centre). More broadly, the ag-
gregation of lymphocyte-wise movement between tissue may be considered an inter-tissue
communication process of information sharing. A lymphocyte may possess a movement
intention given its class, and so may a node, given its tissue type. Further movement
operators may not be limited to lymphocytes, but encompass all so-called cellular-level
components such as antigen. Movement influences both the (1) repertoire homoeostasis
(both discrete repertoire and the aggregate of discrete repertoires), and (2) the dissemina-
154
tion of localised acquired knowledge. Both of these points may be considered to influence
a trade-off between recirculation and localisation of lymphocytes that movement operators
must address. A node may have finite capacity for the number of cellular components.
Further movement may be quantified in terms of the number of cellular components over
time (a rate). This rate may be aggregate or fine-grained by component class. Seeking
points of stability or equilibria (homoeostasis) may involve direct and dynamic manip-
ulation of these movement rates1 . Homoeostasis is perturbed by movement operators
themselves and can be controlled through the management of aggregate movement rates.
Another influence is the dynamic creation of lymphocytes (potentially on a large scale)
in germinal centres and the recruitment behaviour at sites of infection and inflammation.
Such functional behaviours have a potential to overwhelm the homoeostasis capabilities
of the model.

Discrete Repertoire Exposures
The Antigenic Exposure Paradigm (Section 4.2.2) was presented as a series of three ex-
posure types which increase in complexity: (1) single exposure, (2) multiple exposure,
and (3) multiple pathogens. The primary different between the tissue model type and
a single repertoire model, is that the system has multiple points of exposure which may
be all or a subset of the repertoires of the a given tissue model. Figure 5.1 provides a
conceptualisation of a simple chain of discrete repertoires situated in an antigenic envi-
ronment with an asymmetric antigenic exposure pattern. A concern raised with regard
to reconciling the exposure model with multiple points of exposure is referred to as the
Tissue Exposure Regime (TER). A given systems exposure pattern defines the selection of
tissue repertoires of the model to expose with antigen in the event of an exposure event.
For example, a natural inclination may be to use a uniform repertoire exposure pattern
for a single exposure (see Figure 5.2a). In the case of multiple sequential exposures, a uni-
form exposure pattern may not be desirable. Antigen may be considered to arrive to the
system in different locations, thus the antigen-to-repertoire patterns may be irregular (see
Figure 5.2b). In the case of multiple antigens with multiple exposures, not only may the
exposure patterns be irregular, but also the timing of antigen-type exposures. Exposures
may occur concurrently to the system across different repertoires.
In order to provide symmetry in the problem domain with the raise in the level of ab-
straction in the system, a new abstraction of the antigenic exposure paradigm is required.
The exposure paradigm at the cellular level was concerned with an antigen of determi-
nants, thus a suitable metaphor for an exposure of a tissue is an Infection of Antigen.
An infection is defined as the invasion and multiplication of a pathogenic micro-organism
that may cause injury the the tissues. This definition provides a suitable metaphor for
the point-wise exposure of tissue for two reasons (1) invasion and replication suggests a
1
An alternative strategy for direct influence of homoeostasis is that of cellular component deletion
(cellular removal). Implemented models are expected to combine deletion and movement in an effort to
seek homoeostasis, such as stable repertoire sizes.
155
Figure 5.1: Multiple tissue model and the situated spatial properties that may affect
antigenic exposure.
Point of exposure Points of Exposure
(a) Example Symmetric Tissue Exposure. (b) Example Asymmetric Tissue Exposure.
Figure 5.2: Example of two different Tissue Exposure Regime with controlled selection of
discrete repertoires on exposure events.
single antigenic type (pattern) perhaps with variation, and (2) tissue injury promotes ur-
gency which may be interpreted as efficient and effective response, whether it is acquired
in the tissue or imported from surrounding tissues. Although the abstraction is pathogen-
centric given both the exogenous origin and selected metaphor, one may just as easily use
a endogenous-centric metaphor such as tumour occurrence and growth. In both cases the
abstraction properties of discrete (point-wise) information mediation hold.
Spatial-Temporal Exposures
The specifics of the Tissue Exposure Regime may or may not be in the control of tissue
model, which is dependant on the application domain and model configuration. One may
consider a possibility space of all exposure-to-repertoire interactions and consider ques-
tions of lymphocyte movement types in the context of desirable system behaviours under
different regimes. Irrespective of the exposure properties, the concern of the system is to
address what repertoires are exposed at what times. These concerns may be generalised to
that of spatial and temporal consistency of exposure patterns. Temporal consistency refers
to the systems anticipation of the amount of memory required. Spatial consistency refers
to systems anticipation and the positioning of lymphocytes across the repertoires. A lack
of consistency in either dimension removes the ability of a system to anticipate specifically,
requiring general (all-repertoire, all-time) anticipation. It is important to highlight that
the temporal memory applies system-wide without spatial memory (in the face of spatial
inconsistency), and spatial memory applies for all time without temporal memory (in the
156
face of temporal inconsistency). Therefore, the default case is remember everything every-
where for all time in the face of spatial and temporal inconsistency. This is an important
implication of the antigenic exposure paradigm as it defines the role of inter-repertoire cell
movement as the exploitation of patterns in the combination of spatial and temporal con-
sistency of system exposures such that the response memory may be generated and shifted
(adapted) in response to changes in the spatial or temporal properties of the exposure.
The Temporal Memory Effect refers to the systems capability to retain an impression
of an exposure event and recall that impression to provide an improved (compared to
random) response under similar circumstances in the future. The effect is intuitive for a
single repertoire model where it may be facilitated by memory cells or a change in cell
compositional densities. The exploitation of the cells becomes a probabilistic function of
the antigen encountering a memory in the repertoire. Figure 5.3 provides a depiction of
the temporal memory effect in a single repertoire, where cell attrition via repertoire ‘turn
over’ results in the fading of acquired information over time.
Time
Figure 5.3: Depiction of the temporal memory effect in a single repertoire model over
time, where the arrows at the top represent information flow into the system and the
arrow along the bottom represents time.
The multiple repertoire model further requires the dissemination of memory across
multiple repertoires. The memory represents a system-wide spatial anticipation of future
exposures to the same or similar antigen. The following summarise the properties of
temporal memory in the single repertoire model:
• Short Term: A response results in a large short-term memory in the form of ef-
fector cells (short lifespan), which are useful initially at the point of exposure, and
disseminated throughout the system.
• Long Term: A response results in a small long-term memory in the form of memory
cells (long lifespan), which also disseminate throughout the system.
• Memory Size: The larger the antigen dose, the larger the short-term and long-term
memories.
• Memory Stability: With the increases in the frequency of exposures (over the sys-
tems lifetime), the larger and more stable the long-term memory. The memory
size is expected to reach a point of stability where antigen-to-memory cell match is
practically assured.
157
This Spatial Self-Organisation Effect is an emergent property of point-wise exposure
of repertoires in the face of lymphocyte recirculation. The exposure and stimulation of
a repertoire results in a response to the pathogen. The exposure is spatial, therefore the
response is spatial. The following summarises the properties of spatial self-organisation in
the discrete repertoire model:
• Interaction: (response maturation) The clonal selection principle raises a maturated

response to the antigen that results in effectors and memory cells. Some of the raised
clone disseminates early, although the majority remains during the stimulation of the
repertoire. This is the formation of a germinal centre and constraints are imposed
on outbound lymphocytes.
• Recruitment: (response amplification) The repertoire tissue becomes inflamed re-

sulting in the increased intake of recirculating lymphocytes, a response amplification
effect via lymphocyte sequestration. The inbound cells may be naı̈ve cells, effector
cells, or memory cells and may or may not have improved specificity for the pathogen
than the cells already in the repertoire.
• Homing: (preferential recirculation) Memory cells created in a specific repertoire,

preferentially recirculate back to that repertoire, in an effort of spatial anticipation
of the same antigen in the same place.
Figure 5.4: Depiction of the repertoire specialisation that results in the Spatial Self-
Organisation effect, where the arrows from above represent information flow into the
system and the arrows below represent dissipation of acquired information.
The result of the general behaviours of the tissues and the cells is a specific spatial
response strategy that attempts to get the best from the system at the point of exposure
in a bottom-up manner. The spatial organisation may extend through time, such that
in addition to the patrolling effectors and memory cells that disseminate throughout the
system, effectors may remain behind and provide a short-term memory and spatial antici-
pation of response. The following summarises the effects of spatial self-organisation in the
discrete repertoire model:
• Localisation: Specific response at the time and the location of the exposure. Natural
spatial-temporal response.
158
• Surveillance: Long-term and short-term memory in recirculation anticipating the
exposure system wide. The premise that the system may be exposed to the same
pathogen again the future.
• Sentinels: Short-term memory, stationary at the point of exposure, anticipating

exposure at the same location. The premise that a location that has been exposed
may be exposed again.
Spatial anticipation is facilitated in the short-term by sentinel effectors and in the long-
term by memory cell homing patterns. In addition to the spatial self-organisation effect,
the system mixes (perhaps disorganisation) the acquired immunity across the repertoires
of the system. The principles of spatially mixing response effectors and memory provide a
system-wide uniform (or consistent) response to a previously encountered pathogen. This
ensures that immunity learned at one exposure point of the system, may be exploited
elsewhere (anywhere) in the system.
Figure 5.5: Depiction of the unbiased dissemination of information that results in the
Spatial Consistency of Response, where the arrow above represents information flow into
the system and the arrow below represents information availability throughout the system.
This effect called the Spatial Consistency of Response is achieved through the in-
formation dispersal properties of lymphocyte recirculation that ensure that a fraction
of the larger lymphocyte pool is in motion between the repertoires (immunosurveillance).
The effect promotes system-anticipation rather than the more specific spatial-anticipation,
and provides a distributed knowledge or memory that may rapidly self-organise to varied
repertoire exposure patterns. The consistency provided is not strictly uniform, rather it
is quasi-uniform, driven by the bottom-up stochastic decision properties that govern lym-
phocyte recirculation. From a lymphocyte perspective, movement in the face of unknown
exposure patterns provides the system-wide goal of antigen-selection maximisation. That
is, lymphocytes exist to be selected by antigen, thus the behaviours of movement between,
and residence within repertories seeks to maximise the change of this event occurring. As
already described, this is already facilitated by the homing behaviour of memory lympho-
cytes, and the patrolling recirculation behaviour of effectors and naı̈ve lymphocytes. Some
concerns of this behaviour are (1) the quantity of the immunity acquisition, and (2) the
rate of dispersal of the memory. The following summarises the concerns with achieving
consistency of response:
159
• Acquisition Quantity: The system response may be proportional to the exposure,
although the memory (both short-term and long-term) must be sufficient to provide
coverage to all repertoires of the system. This memory is additive, and is expected
to expand in proportion to additional exposures by the same pathogen.
• Acquisition Dispersal : The memory of an exposure originates at the point of ex-

posure, and the delay of dispersal is a function of the (1) quantity of the memory,
(2) the size (number of repertoires) and connectivity of the system, and (3) the
rate of recirculation. Dispersal of acquired immunity must be such that coverage
(probabilistic) of all repertoires is obtained within a reasonable amount of time.
The dispersal or mixing effect in the consistency of response facilitates the blind (re-
circulation) and specific (recruitment) sharing of acquired immunity between repertoires.
Sharing may refer to the exploitation of specific immunity at a exposure location different
from that at which it was acquired. Further, sharing may also refer to the cross-reactive
effects achieved through less specific acquired immunity. This effect includes the abil-
ity of the system to generalise from varied exposure types such that new variations and
combinations may also be exploited.
Figure 5.6: Depiction of information sharing and a cross-reactive response, where the
arrows above represent the flow of information into the system as well as an implicit
anticipation for the application of acquired information.
The integration of these features is a system that is expected to be plastic in its

response capability, although is still capable of generating, maintaining, and self-organising
a specific response to external stimulation. This behaviour may represent a trade-off in
efficiency and efficacy of response between (1) system-wide consistency of response, and
(2) segregation of response into specialised spatial compartments. The system may be
configured to achieve one or the other of these effects, although achieving both effectively
and concurrently represents the challenge for the system in unknown environments.
5.3.3 Tissue Architectures

Exploiting Tissue Responsibilities
This section describes a lymphoid tissue model in which the tissue types (nodes) constrain
the cell-based information processing that may occur to the cells housed within. Tissues
are considered the units of adaptation that may receive and transmit information. Cells
160
System Repertoire
Exposure Anticipation Exposure Anticipation
Uniform mix Recirculation with Segregation

preferential trafficking
Figure 5.7: Depiction of the trade-off between the general anticipation of exposure and
organised anticipation of exposure.
are the substrate for tissues knowledge representation, communication, and adaptation.
This section describes a number of tissue-tissue interactions.
Cell Formation Naı̈ve lymphocytes are created in primarily lymphoid tissue. A naı̈ve
lymphocyte is generated in a pseudo-random manner. It represents an untested guess of
a unit of information that may be useful to the system. A naı̈ve cell may be created in
one primarily lymphoid tissue and migrate to another for specialisation, similar to the
migration of pre-lymphocyte cells from the bone marrow to the thymus to become naı̈ve
T-cells. Those cells that remain in the bone marrow become B-cells. The primary tissue
nodes may condition the formed naı̈ve cells. For example, the cells may be subjected to a
negative and positive selection process such as that subjected to T-cells in the thymus. The
result of such a process is the assessment and ultimate judgement of the feasibility of the
proposed information guesses that the cells represent. In addition to the migration of naı̈ve
cells between primary lymphoid tissue for maturation, naı̈ve cells may uni-directionally
migrate from primary to secondary lymphoid tissue. This represents an upgrade in status
for the cell from unprepared to prepared naı̈ve cells, ready for information processing in
the secondary tissue. As an interaction between primary and secondary lymphoid tissue
it represents a transmission of feasible quasi-random information that may be useful to
the system.
Primary
Tissue
Secondary
Tissue
Figure 5.8: Depiction of the relationship between primary and secondary lymphoid tissues,
where the arrows show the sharing of information between primary tissues and the outward
push of information from primary to secondary tissues..
Cell Maturation Secondary tissues houses many naı̈ve lymphocytes as well as some
memory cells and effectors if available. The tissue accepts a steady stream of naı̈ve lym-
phocytes from primary tissue and integrates them into the cell population. Secondary
161
tissue is responsible for exposing resident cells to antigen and facilitating any immune
response that occurs as a result. Exposure is selection of cognate receptors for a given
antigen, and response is the clonal selection process or similar as outlined in previous
models. The result of the response are memory cells which have demonstrated their use-
fulness, and effector cells for addressing the immediate concerns of the antigen. Effector
cells may or may not be given an impression of the tissue origin of the antigen (site of
infection), as considered in the case of T-cell homing. Memory cells and non-homing ef-
fector cells recirculate between secondary tissues (lymph nodes for example), using the
vascular system as the transport mechanism. The recirculation of memory and effector
cells between secondary lymphoid tissue represents a transmission of acquired information
regarding the environment. Portions of acquired information (individual cells) are shared
between secondary tissues such that the information may be available for application or
further maturation. Acquired information in the form of effector cells is transmitted from
the location of maturation in the secondary tissue, to the location of application in the
tertiary tissue.
Secondary
Tissue
Tertiary
Tissue
Figure 5.9: Depiction of the relationship between secondary tissues and secondary and
tertiary tissues, where the arrows show the sharing of information between secondary
tissues and the outward push of information from secondary to tertiary tissues.
Cell Application Tertiary tissues accept and house many effector cells. The tissues
accept a steady stream of effectors from locally connected secondary tissues, which are
integrated into the population. Tertiary tissues are responsible for being receptive to
pathogenic infection, thus may be considered receptive fields for information from the
environment. The effector cells are employed at the sites of infection as needed, addressing
the major function of their life cycle: application. Samples of the antigen may be taken
from the site of infection in the tertiary tissue and transported into secondary tissue by
courier cells, taking in addition information as to where the antigen was collected (to
facilitate homing). Samples of antigen and some effector cells may be drained back into
locally connected secondary tissue. This draining effect is facilitated by the lymph fluid
that permeates tertiary tissues. The secondary tissue collections and filters the lymph for
antigen, which are presented to resident lymphocytes. Pathogen exposure to tertiary tissue
represent samples of the environment collected by the system. Different tertiary tissues
may collect samples in different ways (for example skin, food, respiratory). The couriering
162
and streaming of samples of pathogen back to secondary tissue provides a transmission of
sensory information from the environment to areas of the system that can be adapted to
address it.
Tertiary
Tissue
Secondary
Tissue
Figure 5.10: Depiction of the relationship between tertiary and secondary tissues and
tertiary tissues and the environment, where the arrows depict the flow of information
from the environment to the tertiary tissues and ultimately to the secondary tissues.
Integration
The aggregation of all three tissue types describes an integrated model with layered in-
formation processing. From the inside out, the information availability increases in mat-
uration (usefulness). The tertiary tissue contains no other cells than those known to be
immediately useful, the primary tissue contains no other cells than those whose usefulness
is unknown. The secondary tissues act as the mediator between these two extremes, pro-
viding the testing and maturation processes for information acquisition. The outward flow
of prior knowledge through the layers of the system, whereas information about the en-
vironment flows inward. Principle pathogen exposures occur in the tertiary tissue where
it may be addressed by resident effectors. If sufficiently large or novel, samples of the
pathogen find their way into secondary tissue (courier or stream) for presentation to the
learning mechanisms of the system. Therefore, lymphocytes migrate from the centre of the
system outward, increasing in proficiency as they approach the surface, whereas pathogen
are housed in the outer layer and seep into the second tier.
Antigens that penetrate the system and manage to take-up residence in the tertiary
tissue are drawn to the lymph nodes of the secondary tissue by fluid called lymph. Lymph
drains into the lymph nodes, which in turn filter the fluid collecting antigens, which
are presented to local and recirculating lymphocytes. From an information processing
perspective, a lymph node is provided an information stream from the regionally located
tertiary tissue. A proportionally large region of tertiary tissue feeds pathogen to a single
lymph node, in effect proving multiple feeding information channels. The information
is collected is specialised regions within the node. B-cells are exposed to the presented
information, and if sufficiently activated may migrate to another specialised region on
the node and form a germinal centre in which bouts of selection, proliferation, receptor
maturation and ultimately cell differentiation occur. The product of this learning process
are effector cells called plasma cells that produce antibodies for the triggering antigen
163
Pathogen
Primary
Secondary
Tertiary
Figure 5.11: Depiction of the layered tissue architecture and information flows between
the layers, the arrows show the flow of information from the environment into the system
as well as the outward push of new detectors from the centre of the system.
in large numbers, and mature long-lived memory cells. The antibodies enter the blood
stream and seek a serendipitous interaction with their cognate antigen. The memory cells
recirculate between the secondary tissue and sometimes move into the tertiary lymphoid
tissue, with the same mission. There is no homing effect for the B-cells. Their effectors
flood the system, and their memory cells patrol. Importantly, the effectors and the memory
cells are created and released in the vicinity of where the pathogen was encountered. Thus,
they are more likely to encounter their cognate antigen sooner rather than later.
Tertiary tissue
Pathogen
Lymph
Lymph Node
Figure 5.12: Depiction of the information streaming providing by lymph collecting in

lymph nodes.
Another type of cell, called a dendritic cells, patrol tertiary tissue for pathogen. Once
detected, these cells collect the material and migrate into the secondary lymphoid tissue
(such as lymph nodes). In the secondary tissue, the dendritic cells present the pathogenic
material to T-cells that respond, proliferate, and differentiate. The product of this learning
process is mature T-cells in the form of effectors and likely long-lived memory cells. The
effector T-cells are imprinted by the dendritic cell as to the location in the tertiary tissue
of the pathogen exposure. The effectors recirculate the system and penetrate the tertiary
tissue when the specific chemical signature is detected. Information is streamed from the
tertiary tissue to the secondary tissue although the streaming is facilitated by courier
cells. These cells are aware of where they collected their material and thus are capable
of imprinting that information onto the produced effector cells. The result is an effect
164
that differentiates the lymphocyte movement behaviour from that of the B-cells (that
blindly seek their cognate antigen), to the T-cells (that home into the chemical signature
of infected tissue).
Tertiary Tissue
Pathogen Dendritic Cell
Lymph
Node
Homing
T-cell
Figure 5.13: Depiction of the courier-like behaviour of dendritic cell’s.
Tissue Modes
The information properties of the system have been described with regard to the trans-
mission of the information substrate (immune cells) between the layers and within the
layers. Table 5.1 lists the abstracted immunological functional behaviour required to meet
the information processing needs outlined for the architecture.
Tier Information Processing Behaviour

1 Generation and feasibility testing Negative Selection
2 Maturation Clonal Selection
2 Homoeostasis Elaborated Clonal Selection
2 Recirculation Recirculation Algorithm
2-3 Courier/Homing Homing Algorithm
2-3 Exposures/Streaming Antigenic Exposure
2-3 Inflammation Recruitment Algorithm
Table 5.1: Summary of information processing needs, and proposed algorithms that meet
those needs.
The information processing constraints of each tissue type, outline a clear respon-
sibility for each tier of the architecture. Primary tissue is concerned with information
generation and the pre-processing of information before it is employed in a learning pro-
cess. Secondary tissue is responsible for learning, for identifying which internal information
is presently useful and attempting to improve it (maturation), whilst at the same time
applying it (effectors). The second tier is also responsible for the generation and main-
tenance of longer term memory, and the selective application of acquired memory. The
tertiary tissue is responsible for providing an interface to the environment, mediating the
application of acquired knowledge to sensory signals, with the filtering of temporally novel
signals to secondary tissues for response generation.
The majority of the clonal selection information processing occurs in the secondary
lymphoid tissue. The other two layers may be easily incorporated into the secondary
165
Tissue Type Responsibility
Primary Tissue Information generation (preparation)
Secondary Tissue Information maturation (learning and memory)
Tertiary Tissue Information application (sensory)
Table 5.2: Summary of the principle responsibility of each tier of the architecture.
tissue, providing only conceptual attributes of the tissue model. This compression of re-
sponsibility has been implicit in previously proposed clonal selection, negative selection,
and other cell-theory based algorithms. Given that the tissues may be compressed, it is im-
portant to consider the utility of separating the proposed responsibilities into a dependant
hierarchy of layers.
Environment
Tertiary Tissue (sensory)
Secondary Tissue (learning )
Primary Tissue (preparation )
Figure 5.14: Depiction of the layered information processing properties of the lymphoid
tissue architecture.
This section proposes three different models that employ the proposed lymphoid tissue
architecture. Each model is inspired by a different aspect of the acquired immune system,
from a lymphatic system (tissue) perspective. The principles for separated responsibility
are (1) the centralised generation of naı̈ve cells facilitating centralised prior knowledge,
(2) consolidated secondary tissue for a networked learning architecture large enough to
meet the needs of the system, (3) a large host consisting predominantly of tertiary tissue,
which provides a distributed sensor platform for receiving and responding to environmental
signals.
Filter Model
The filter model is inspired by the specialisation of the lymphatic system at the primary
entry points for exogenous antigen (pathogen). Examples include the tonsils and related
tissues for addressing pathogen that enter the host via the respiratory system, the Peyer’s
patch tissues in the intestines that addressing pathogen that enter the host via the digestive
system, and the spleen that filters the blood for pathogen. The model contains a single
point of entry for pathogen (tertiary tissue), which is monitored by one or a small collection
of secondary lymphoid tissues. The secondary tissue is supported by a single source of
naı̈ve cells (primary tissue).
Although the model is called a filter (the basis for its inspiration) it is not restricted to
166
Environment
Tertiary
Secondary
Primary
Figure 5.15: Depiction of the simple filter model.
filtering information processing tasks. The focus of this model is on the system possessing a
single point of entry for pathogen (sensory signals) although at a high load (many signals),
and one or a small number of specialised secondary tissues to address the needs of the
tertiary tissue.
Tissue Type Configuration

Primary Single tissue feeding the secondary tissue.
Secondary Single (or a small number) secondary tissues with high capacity of cells.
Tertiary Single point of entry for the system to monitor.
Table 5.3: Configuration of the architecture for the filter model.
Lymph Node Model
The lymph node model is inspired by the network of lymph nodes, connected by the
vascular system, that receive antigen as a stream in the lymph. The model is more complex
than the filter model as it contains a network of moderately sized secondary lymphoid
tissues that recirculate memory and effector cells. Further, unlike the filter model, it has
a large receptive field of tertiary tissue, which provides a single broad sensory system.
Environment
Tertiary
Secondary
Primary
Figure 5.16: Depiction of the lymph node model.
The focus in this model is the large tertiary tissue receptive field, and the network
167
of secondary tissues (lymph nodes), in particular the capabilities of the lymph nodes to
recirculate mature cells (share acquired information), and their capabilities in trafficking
effectors into tertiary tissues (applying acquired information).

Primary Single tissue feeding the secondary tissue.
Secondary A consolidated network of moderate sized secondary tissues.
Tertiary A large and uniform receptive field for input signals.
Table 5.4: Configuration of the architecture for the lymph node model.
Host Model
The host model is an integration of one or more filter models with the lymph node model.
The resultant model is a larger-scale tissue architecture that may be equated to a lymphatic
system of a host. The model is composed of a variety of different tertiary tissue types and
thus different meanings for the secondary tissue to be provided with antigen to present
to resident cells. In addition, the secondary tissue is made of a core network of lymph
nodes, as well as collections of specialised lymphoid tissue to address principle pathogen
penetration areas. Finally, the system has one or more primary lymphoid tissues for the
creation and dissemination of naı̈ve cells to secondary tissues.
Environment
Tertiary
Secondary
Primary
Figure 5.17: Depiction of the host model.
The focus of the model is the integration of the lymph node network (and related
method for sampling antigen from tertiary tissue) with a number of filter-based models
centred around primary pathogen access points. This host model represents a hybrid
integration of the two sub-types of tissue architectures, and culminating (pinnacle in com-
plexity) lymphatic tissue model.
Summary
The lymphoid tissue components, exposures, and architecture models provide a context
for the previously defined adaptive models of the Cellular Paradigm concerning immune
168
Primary Set of primary tissues feeding the secondary tissue.
Secondary A backbone of a consolidated network as well as specialised regions for each
major entry point (diverse.
Tertiary Multiple tertiary tissues of varied size and scope (diverse)
Table 5.5: Configuration of the architecture for the host model
cells (intra-tissue), and a context for novel cell migration adaptive models to be investi-
gated (inter-tissue). The applications of the architecture in the proposed lymphoid tissue
models provides three specific settings for integrated lymphoid tissue-based information
processing. This section suggested at the potential for Tissue Clonal Selection and more
generally tissue-based Artificial Immune System algorithms, as a scale above the state of
classical cell-based Artificial Immune System algorithms.
5.4 Realised Tissue Paradigm

A single immune system of tissues that may be exposed to infections provides an effective
metaphor for considering the problem of the acquisition of information in a decentralised
manner from a distributed environment, and strategies toward the anticipated applica-
tion of acquired information. The presentation of the lymphatic system demonstrated
that some infections can be localised to predictable lymphoid tissues (peripheral immune
system) and others cannot (systemic immune system). Therefore a host immune system
of lymphoid tissues must simultaneously manage predicted, unpredicted, and a variety of
intermediate antigenic infections from its antigenic habitat. Toward this end, the host
immune system considers the differential mobility of cells between tissues and differential
cell types as a strategy for managing so-called anticipated and unanticipated antigenic
infections across all tissues. This section provides a realisation of the concerns of the ab-
stract Tissue Clonal Selection Paradigm presented in Section 5.3. This includes definitions
of a tissue problem in the Infection Exposure Problem and specialisation in colour space,
and the Tissue Clonal Selection Algorithm that provides a basis for investigation into cell
trafficking schemes. A series of tissue-based measures are presented for assessing tissue
algorithms on colour space realisations of infection problems that provide quantitative
indicators regarding the concerns of spatial temporal exposures. Collectively, this section
provides a basis for the implementation and exploratory empirical investigation into the
Tissue Clonal Selection Paradigm, as well as a bridge between the the biology (Section 5.2)
and abstract models (Section 5.3) of Tissue-centric Clonal Selection.
5.4.1 Antigenic Infections

This section considers the realisation of the exposure paradigm outlined in Section 5.3.2.
In particular, this section abstracts the Antigen Exposure Problem considered for the
cellular paradigm for the tissue paradigm called an Infection, and specialises an instance
of the problem in Colour Space. In addition, the discrete properties of tissue exposures
169
are reconsidered and a number of specialised Tissue Exposure Regimes are defined.
Infection Antigenic Exposure Problem
The Antigenic Exposure Problem (AEP) defined in Algorithm 4.1 provides a general prob-
lem definition where a Tissue T is exposed to an Infection I. A given Cellular Clonal Selec-
tion Algorithm may be considered a single T in a Host of Tissues (H = {T1 , T2 , T3 . . . , Tn }).
Therefore the AEP must be generalised for interaction with an H. The Infection Anti-
genic Exposure Problem (IAEP) defined in Algorithm 5.1 provides a generalised exposure
problem in which an H is exposed to a Habitat B of Infections I (B = {I1 , I2 , I3 , . . . , In }).
The parameters Ndeterminants , Nantigen , and Ninf ections define the properties of B with a
symmetrical number of I which are passed to CreateHabitat to create a problem instance.
Exposure refers to a given exposure mechanism, such as the variety of Tissue Exposure
Regimes (TER) defined later in this section. An exposure regime returns a result set Hrs
that represents the state of a given H with regard to its ability respond to the B under
the given TER.
Algorithm 5.1: Infection Antigenic Exposure Problem (IAEP).

Input: H, Ndeterminants , Nantigen , Ninf ections
Output: Hrs
1 B ← CreateHabitat(Ndeterminants , Nantigen , Ninf ections );
2 Hrs ←0;
4 Hrs ← Exposure(H, B);
5 end
6 return Hrs ;
Infection Colour Space Problem
The IAEP may be specialised as an infection extension of the Antigen Colour Space Prob-
lem (ACSP) defined in Section 4.3.1 called the Infection Colour Space Problem (ICSP).
An instance of the ACSP involved the exposure of a T to a set of Colour Space Patterns
(CSP), where the goal of the problem is to minimise average error in the response Trs . The
ICSP specialisation of IAEP exposes a H to a set of sets of CSP (B of I’s), with the goal
of minimising the average error in the response Hrs (defined in Equation 5.1). One may
define a minimal configuration of the IAEP and thus the ICSP that focus the concerns of
the problem on the highest level of abstraction, the Habitat B. This focus is provided by
reducing the configuration at levels below the desired level to minimal levels of complexity.
A Minimal Infection Antigenic Exposure Problem (MIAEP) is defined where Ninf ections is
variable, although Nantigen and Ndeterminants are fixed at 1. In the case of a minimal ICSP,
the number of I defines the number of CSP (A), each of which has a single D comprised
of all three Colour Space Components (Red, Green, and Blue).
170
Habitat
Habitat
Infection Infection ( )
Infection Infection ( )
( ) Antigen Antigen Antigen Antigen
Determinant Determinant
Determinant Determinant ( )
(a) General IAEP Configuration. (b) Minimal IAEP Configuration.
Figure 5.18: Depictions of general and minimal configuration examples of the Infection
Antigenic Exposure Problem (IAEP).
Tissue Exposure Regimes
Section 5.3.2 defined the two important aspects of the problems that Tissue Clonal Selec-
tion Algorithms must address: (1) discrete tissue exposures, and (2) the spatial temporal
properties of exposure anticipation defined by the regularity of a given tissue exposure
regime. Toward this end, this section explicitly defines five different Tissue Exposure
Regimes (TER) for specialising the Exposure(H, B) of the IAEP providing examples
from the range of possible regular and irregular exposures (spatial-temporal exposure of
tissues) with some and all of the information in the problem domain. The schemes in-
clude: Symmetric Exposure, Asymmetric Exposure, Point Exposure, Random Exposure,
and Probabilistic Exposure. The TER provides a mechanism by which to specialise an
the infection exposures of a habitat to the tissues of a host, providing an generalisation of
the limitations imposed by the structure, organisation, and properties of different tissues
as described by the various tissue architectures in Section 5.3.3. Figure 5.19 provides a
depiction of four of the five exposure regimes to a host of tissues.
Symmetric The Symmetric Tissue Exposure Regime (STER) defined in Algorithm 5.2
involves the exposure of each Infection I to each Tissue T (see Figure 5.19a). Thus, each
tissue is exposed to and must address the scope of the antigenic habitat B.
Algorithm 5.2: Symmetric Tissue Exposure Regime (STER).

Input: H, B
1 foreach Ti ∈ H do
2 foreach Ii ∈ B do
3 Exposure(Ii , Ti );
4 end
5 end
Asymmetric The Asymmetric Tissue Exposure Regime (ATER) defined in Algorithm 5.3
involves the exposure of each I Infection to a single Tissue T in a consistent manner, such
171
(a) Symmetric Tissue Exposure Regime (STER). (b) Asymmetric Tissue Exposure Regime (ATER).
(c) Random Tissue Exposure Regime (RTER). (d) Point Tissue Exposure Regime (PTER).
Figure 5.19: Depictions of the four different Tissue Exposure Regimes (TER).
that once established the same T -to-I mapping is used for each exposure (see Figure 5.19b).
The number of T may match the number of I such that each T addresses 1
Ninf ections per
host exposure.
Algorithm 5.3: Asymmetric Tissue Exposure Regime (ATER).

Input: H, B
1 foreach Ti ∈ H ∧Ii ∈ B do
2 Exposure(Ii , Ti );
3 end
Point The Point Tissue Exposure Regime (PTER) defined in Algorithm 5.4 restricts the
exposure of all I to a single and consistent T across all host exposures (see Figure 5.19d).
The choice of T that becomes the point of exposure for the host may be randomly selected
from H.
Algorithm 5.4: Point Tissue Exposure Regime (PTER).

Input: H, B, Tpoint
2 Exposure(Ii , Tpoint );
3 end
Random The Random Tissue Exposure Regime (RTER) defined in Algorithm 5.5 is
similar to the asymmetric exposure regime, although the matching of T with I occurs
randomly with re-selection each exposure (see Figure 5.19c). Thus, a given T may be
exposed to a different subset of the antigenic habitat B each host exposure where it is
possible for the set to be empty.
172
Algorithm 5.5: Random Tissue Exposure Regime (RTER).
Input: H, B
2 Tpoint ← RandomTissue(H);
3 Exposure(Ii , Tpoint );
4 end
Probabilistic The Probabilistic Tissue Exposure Regime (OTER) defined in Algorithm 5.6
provides an intermediate specialisation of the RTER and ATER in which the selection of
T is probabilistic based on past selections providing probabilistic spatial consistency of
re-exposure. In addition, the I-to-T mapping is maintained for an extended number of
host exposures before a probabilistic reselection is made, providing temporal regularity
to the spatial exposures. The duration parameter defines the number of host exposures
(the HostExposureN umber state variable) a selected T is exposed to a given I. The
state variable Histogram keeps track of the frequency each T is exposed to each I, and
biasedRouletteW heelSelection(Histogram) uses a biased roulette wheel selection routine
on the histogram for a given I over all T to select a T to expose for duration host expo-
sures. The Itissues state variable keeps track of which T each I has selected and should
expose.
Algorithm 5.6: Probabilistic Tissue Exposure Regime (OTER).

Input: H, B, duration, Histogram, HostExposureN umber, Itissues
2 if (Epoch mod duration) ≡ 0 then
3 Ti ← BiasedRouletteWheelSelection(HistogramIi );
4 Itissuesi ← Ti ;
5 HistogramIi Ti ← HistogramIi Ti + 1;
6 end
7 Exposure(Ii , Itissuesi );
8 HostExposureN umber ← HostExposureN umber + 1;
9 end
5.4.2 Tissue Clonal Selection

The Cellular Clonal Selection Algorithm (CCSA) defined in Algorithm 4.4 describes the
general clonal selection interaction between an Infection I and a Tissue T . The tissue
paradigm requires a raise in the level of abstraction from a single T to a Host H of T
(H = {T1 , T2 , T3 , . . . , Tn }) that responds to a Habitat B of Infections I. Tissue Clonal
Selection Algorithm (TCSA) defined in Algorithm 5.7 as its names suggests describes a
general tissue clonal selection algorithm that may be specialised based on the concerns
and constraints of the tissue paradigm. In particular, the T issueInteractions operation
provides a facility for specialising the interactions between T such as the trafficking of Cells
(C). An important distinction between the rise in the level of abstraction from CCSA to
TCSA’s is the delegation of concerns. TCSA create and exploit multiple instances of the
173
CCSA without concern for the specifics of the initialisation and response to exposure. This
provides flexibility in TCSA’s to focus the tissue paradigm concerns such as cell trafficking
and potentially making use of a heterogeneous set of CCSA, whilst at the same time forcing
responsibility for tissue-level responses (Trs ) onto the CCSA. The construction of a Hrs is
the linear aggregation of the specific Trs from each T produced through the interaction of
an I and a T as paired according to the TER. The TCSA is immune system centric unlike
the IAEP that is environment centric, providing a flip in responsibility from a passive H
responding to Exposures according to a TER in the case of IAEP, to an active H that
invokes Exposures on a B, although still delegating the TER to the IAEP. This conceptual
flip provides a computationally equivalent system-environment relationship that facilitates
bi-modal realisations suitable for different implementation specific concerns. The use of a
TER provides an important decoupling of the Host from the controlled governance with
the information provided by the antigenic habitat, specifically with regard to the selection
of the Hrs and the adaptation within exposed tissues. This decomposing in which control is
a facet of the specific problem-system implementation provides a clear and simple platform
for investigating the behaviour of decentralised clonal selection distributed across a set of
potentially interacting independent repertoires of cells.
Algorithm 5.7: Tissue Clonal Selection.

Input: B, Ntissues
Output: H
1 H ←0;
2 for i←0 to Ntissues do
3 Ti ← CreateTissue();
4 H ← Ti ;
5 end
7 B.Exposure(H);
8 TissueInteractions(H);
9 end
10 return H;
One may define a minimal implementation of the T issueInteractions(H) that does not
facilitate any interaction between tissues. This Minimal Tissue Clonal Selection Algorithm
(MTCSA) of isolated tissues provides a baseline for comparison with extensions of the
TCSA that do facilitate tissue interactions. The important behaviour that the MTCSA
provides is an example of segregated adaptive tissues, highlighted in the trade-off between
information organisation in a regular TER and information mix in an irregular TER (see
Figure 5.7).

This section defines a series of empirical measures derived from the measures used in the
cellular paradigm (Section 4.3.3) that provide instantaneous information regarding a given
TCSA on colour space specialisations of the Infection Antigenic Exposure Problem. The
174
measures are classified as system that provide general holistic information about the algo-
rithm’s information state and component that provide information about the algorithm’s
component information state. As with the cellular paradigm, an important consideration
with the proposed measures in exploratory experimentation is not their absolute value,
but rather their relative change in value with changes to the systems being investigated.
System Measures
Host Error The Average Cell Error (ACE) defined in Equation 4.4 provides a gener-
alised measure of error for a single Tissue Result Set of cells Trs against a given Infection
of antigenic patterns I. For the purposes of investigating TCSA the ACE measure may
be renamed Tissue Error (TE) that may be calculated for each Trs in a Host Result Set
Hrs and averaged to define a Host Error (HE) (see Equation 5.1). HE provides an indi-
cation of instantaneous average error for each Trs for a given Hrs provided by a TCSA in
response to an Tissue Exposure Regime. HE is in the units of the B problem space, for
example Euclidean distance for a colour space specialisation. HE provides an indication
of a systems capability of addressing a given Habitat problem B under the constraints of
a given TER, in particular the spatial-temporal regularity or lack thereof in exposures.
n
1$
HostError(B, Hrs ) = T issueError(Ii , Trsi ) (5.1)
n i=1
Host Diversity Host Diversity (HD) defined in Equation 5.3 provides an instantaneous
diversity a TCSA. Each Tissue T in a Host H is compressed to a Bit Frequency Histogram
(BFH) defined in Algorithm 5.8 which records the occurrence of each bit for all bit vector
based Cells C. The BFH provides a compression of the information content of a given
T that may be directly compared with that a BFH from another T as the absolute Bit
Frequency Difference (BDF) defined in Equation 5.2. The measure calculates the average
BDF for a given T against all other T in H, which is then averaged across all T . HD
provides a holistic indication of the diversity of the information content of an H. An
increase in the HD suggests a more heterogeneous set of information, whereas a decreases
in HD suggests a more homogeneous set of information that may indicate an improved
system-wide specialisation of resources toward the H.
n
$
BitF requencyDif f erence(BF Hi , BF Hj ) = | BF Hik − BF Hj k | (5.2)
k=1
 
n n
1$ 1$
HostDiversity(H) =  BitF requencyDif f erence(BF H(Ti ), BF H(Tj ))
n i=1 n j=1
(5.3)
175
Algorithm 5.8: Bit Frequency Histogram (BFH).
Input: T
Output: BFH
1 BFH ←0;
3 foreach cj ∈ Ci do
4 if cj ≡ 1 then
5 BF Hj ← BF Hj + 1;
6 end
7 end
8 end
9 return BFH;
Component Measures
Average Tissue Diversity The Average Cell Diversity (ACD) defined in Equation 4.7
provides a measure of Hamming distance for the average cell from all other cells in a Tissue
repertoire of cells T . As such, the ATD may be referred to as Tissue Diversity (TD) which
may be calculated for each T in a Host H and averaged to provide the Average Tissue
Diversity (ATD). ATD provides an instantaneous measure of the diversity of the average
T in a host, measured in bits. ATD provides a general indication about the distribution of
information in the system. A relative increase in ATD suggests that the average T has a
more diverse cell composition, whereas relative decrease in ATD suggests that the average
T has a less diverse cell composition. Decreased homogeneity may suggest an increase
in the organisation of information across the system and a decrease in homogeneity may
suggest an increase in the mixing of the information content across the system respectively.
n
1$
AverageT issueDiversity(H) = T issueDiversity(Ti ) (5.4)
n i=1
Average Tissue Error The Host Error defined in 5.1 provides an indication of a Host
system’s error with regard to the Hrs prepared as a result of where a give TER interacted
with the Host. Alternatively, a Hrs may be prepared systematically for each T in response
to the Habitat B, and averaged across all T in H called the Average Tissue Error (ATE).
ATE provides an instantaneous measure of the error for the average Tissue component T
in the system in the units of the problem space B. ATE provides a general indication of
the capability of the average tissue in the system. A relative increase in ATE suggests the
average T has a more generalise response capability, whereas a relative decrease may sug-
gest a more specialised response capability. Importantly, an increased generalised response
capability suggests an increased mixing of the information content across the system.
 
Tn Bn
1 $ 1 $
AverageT issueError(B, H) =  T issueError(Ij , Trs i ) (5.5)
Tn i=1 Bn j=1
176
Section 5.3.2 highlighted two points of concern regarding the behaviour of tissue-based sys-
tems under different exposure regimes: (1) the spatial organisation of information based
on the spatial temporal regularity of tissue exposures, and (2) the system-wide consistency
of response. These two concerns provide a trade-off in the organisation of cell-based infor-
mation within a multiple tissue model between a uniform mix for a completely irregular
tissue exposure pattern and isolated tissues for completely fixed regular tissue exposure
patterns (see Figure 5.7). This provides a primarily research agenda for the Tissue Clonal
Selection Paradigm for investigating Tissue Clonal Selection Algorithm behaviour under a
variety of different Tissue Exposure Regimes. The results from such an investigation pro-
vide a basis for using such algorithms as tools in problems in which the exposure and/or
use of information is controlled, uncontrolled or some combination (a consideration that
is further investigated in the Host Paradigm, see Section 6.3.2).
Algorithm Mechanism to Behaviour
One may clearly distinguish between the measurable emergent effects, the information
management strategy that such effects represent, and the grounded (component-wise mech-
anisms) of TCSA’s that promote such strategies and effects (summarised in Table 5.6).
Effect Strategy Mechanism

Consistency of Response Everything everywhere Inter-tissue mixing of cells
Spatial Organisation Only at point of use Inter-tissue isolation
Table 5.6: Summary of the difference in effect, strategy and mechanism between the
complementary concerns of tissue algorithms under discrete exposures.
The two identified emergent effects of interest are the systems holistic consistency of
response and the systems spatial organisation. These represent two different information
management strategies, specifically the dissemination and thus promoted availability of all
acquired information across the entire system, and the specialisation and thus availability
of acquired information at the point of use. As discussed in Section 5.3.2, these two
strategies represent two different relationships between acquisition and anticipated need,
specifically local acquisition with global anticipated need and local acquisition with local
anticipated need. The first strategy may be promoted via the global mixing via continuous
high-amplitude inter-component sharing of information, and the second strategy may be
promoted via the segregation and thus isolation of acquired information. These strategies
represent extremes in mechanism and behaviour with many intermediate states promoted
via variation in mechanism, and thus variation in emergent effect.
Exposure Regimes and Trends
The five exposure regimes may be considered with regard to their exposure regularity
and irregularity in terms of the scope of system information acquisition and anticipated
need (summarised in Table 5.7). For example, the asymmetric regimes represents an
177
archetype of regular exposure with local exposure and anticipation whereas the random
regime represents the archetype of irregular exposure with global exposure and antici-
pation. Both symmetric and point regimes expose the system in a regular manner and
thus with expected local availability and anticipatory system expectation. This is the case
for PTER, although given that STER is PTER applied to all components of the system,
there is opportunity for the same information to be useful (anticipated) at any point of
the system. The probabilistic regime provides an intermediate between the asymmetric
regime and the random regime. For example, with a duration of 1 and frequency-biased
re-exposure OTER provides a version of RTER with probabilistically increased regularity
(local re-exposure and anticipation), whereas random is OTER with duration 1 and un-
biased selection and ATER is OTER with an infinite duration. OTER with a moderate
duration provides an explicit intermediate between the global-global behaviour of RTER
with the local-local behaviour of ATER.
Regime Exposure/Acquisition Re-Exposure/Anticipation

ATER Local Local
STER Local-Global Local-Global
PTER Local Local
OTER Local/Global Local/Global
RTER Global Global
Table 5.7: Summary of the exposure and re-exposure behaviour of the five defined tissue
exposure regimes and how they relate to system information availability and anticipated
need.
Relative Measure Trends
The proposed measures provide information content both quantitatively as required, al-
though more usefully qualitatively as relative differences between similar systems and the
same systems under varied exposure regimes. Table 5.8 provides a summary of the sys-
tem and component level measures and an interpretation of the relative movement trends.
These general trends may be used as a guide for assessing results.
Measure Movement Description

DH Up System is more diverse (tissues are more heterogeneous).
Down System is less diverse (tissues are more homogeneous).
HE Up System performance has decreased.
Down System performance has increased.
ATD Up Tissues are more diverse (cells are more heterogeneous).
Down Tissues are less diverse (cells are more homogeneous).
ATE Up Tissues perform worse independently.
Down Tissues perform better independently.
Table 5.8: Summary of general trends and their interpretation for Tissue Clonal Selection
Algorithms.
178
System Behaviour Expectations
Table 5.9 provides expected qualitative comparisons for a given tissue-based system with
global and local anticipation strategies under local-local and global-global tissue exposure
regimes. The qualitative expectations are given in terms of the system and component
level diversity and error, where diversity reflects information content and error ultimately
reflects correctness of anticipation. The ATER may be consider an archetype local-local
(regular) TER given the piece-wise exposure of information to fractions of the informa-
tion environment (antigenic habitat). RTER may be consider the archetype global-global
(irregular) TER given the unbiased mapping between components and exposure. The re-
maining TER’s provide intermediates for investigation. The error and diversity measures
at host and tissue scope may be generalised to system and component, thus making the
extended trends broadly applicable (such as in the case of the Host Paradigm in Chapter
6). As discussed, high-rate inter-tissue mixing provides an example of a global anticipa-
tion strategy, and isolation such as that used in the MTCSA provides an archetype of
a local anticipation strategy. Remaining variations and elaborations of the TER provide
intermediates between these strategies.
Measure Local-Local TER Global-Global TER

Scope Type Global Strategy Local Strategy Global Strategy Local Strategy
System Diversity Decreased Increased Decreased Increased
System Error Increased Decreased Decreased Increased
Component Diversity Increase Decrease Increase Decrease
Component Error Decreased Increased Decreased Increased
Table 5.9: Summary of the expected behaviours of the defined measures with regular and
irregular exposure regimes on systems with mixing and isolated information dissemination
strategies.
Expected system error predictably follows the matching between strategy and TER,
and the inverse of the relationship. Thus, the expectation is that system error is a cor-
related match between TER and information management strategy, whereas component
error is generally invariant to TER and is correlated with the information management
strategy. Both system and component level diversity are also expected to be generally
invariant to the TER and correlated with the strategy. System diversity is expected to
decrease with a global strategy as the system becomes information-homogeneous and be-
come information-heterogeneous with a local strategy. This relationship is expected to be
inverted for component level diversity in which components generally increase in diversity
with the global strategy and decrease with a local strategy given the correlated decrease
and increase in intra-tissue specialisation respectively. A final axis to these expectation
is responses, specifically the time allocated to address an environment given sufficient re-
sources. It is expected that both system strategy types will arrive at the same or similar
final system error after a prolonged period of time. Thus, the expected behaviours for this
measure must be observed after a limited although reasonable amount of time.
179
This section outline a high-level research agenda for exploratory empirical experimentation
for the Tissue Paradigm, considered in the remainder of this chapter.
1. Specific Goals: Confirm Expectations
(a) System-Error correlation between Strategy and TER, specifically Local strategy
with Local TER and Global strategy with Global TER.
(b) System-Diversity and Component-Error and Diversity invariance on Strategy
and TER, specifically Decrease in System-Diversity and Component-Error with
Global strategy and Increase with Local strategy, as well as Increase in Component-
Diversity with Global strategy and Decrease with Local strategy.
2. General Goals: Consider Intermediates
(a) Graceful falloff in measure with intermediate strategies and/or exposure regimes,
generally a predictable continuum between the extremes in behaviour and thus
measure.
(b) Intermediate information management strategies provide varied intermediate
behavioural effects.
5.5 Lymphocyte Recirculation

5.5.1 Recirculation Metaphor and Strategy
Section 5.2 described the recirculation of lymphocytes around the vascular and lymphatic
systems of a host such that a circulating pool of antibodies and cells are able to rapidly
transit and disseminate between locations of need and potential need. Lymphocyte re-
circulation provides a metaphor for a computational mechanism investigating the ‘Con-
sistency of Response’ emergent effect. Collectively, the lymphocytes in all tissues and
thus the tissues themselves represent the acquisition of information (immunity) from anti-
genic habitat. The movement process facilitates the dissemination of information between
the distributed tissue structure of the host. Ideally, all acquired information would be
available in all tissues in all time for the system to be able to effectively respond to the
unknown environment of spatial-temporal exposure patterns. This is not possible given
(1) the limitations of carrying capacity of tissue locations and (2) the vast quantity of
lymphocytes possessed by the entire system. The recirculation mechanism is a strategic
exercise in anticipated information availability. An antigen must physically contact the
receptors of cells to ‘test’ for a reaction, thus smaller lymphocyte pools may provide some
efficiency benefits in terms of local trials. Further, not all information is required every-
where, rather specific information for a given antigen is needed at a specific location at an
unknown time. Bounded lymphocyte recirculation provides a trade-off of these concerns.
The Lymphocyte Recirculation Strategy is defined as a mobile distributed repertoire that
promotes spatially distributed information availability of a limited resource to get the right
information (cells) to the right location (tissue) at the right time (exposure).
180
5.5.2 Recirculation Tissue Algorithm
The tissue interaction step of the Tissue Clonal Selection Algorithm (TCSA) may be spe-
cialised for cell recirculation involving the selection of cells to depart a given tissue and
the integration of cells arriving to the tissue. These may be separate decision processes
although they are clearly related. For example, a simple model involves the removal of
a single lymphocyte (posted to a neighbouring tissue) which frees a position in a fixed-
size repertoire for one-arriving lymphocyte from a neighbouring tissue. The trafficking
of lymphocytes may be elaborated as the configuration of a migration and integration
operators, which may be specialised as the selection of a subset of a given repertoire for
removal and integration into a neighbouring repertoire (a remove-insert strategy). A sim-
plified tissue architecture may be defined as a ring network topology with uni-directional
communication (directed cyclic graph) that superficially resembles a circulatory system
between the tissue nodes. The migration configuration is defined by the number of cells
per migration event (amplitude) and how often migration events occur (frequency). The
number and consistency of the sample of lymphocytes posted to the neighbouring tissue
ideally should be representative of the acquired information content of a given repertoire.
The lymphocytes arriving from neighbouring tissue to be integrated into the repertoire
should have the similar information content, thus meeting the goals of the recirculation
strategy. In addition to providing a suitable sample of the receptors in the repertoire, the
strategy should permit the further development of information in the repertoire. Arriving
lymphocytes must be integrated into the repertoire. This may be more involved if there
is a size mismatch between the number of cells selected and sent compared to the number
arriving (for example if cells arrive from multiple tissue locations). Arriving cells may
fill the vacant slots provided by the departed cells, and use a replacement strategy (as
discussed for clone integration in the elaborated clonal selection algorithm) in the case of
overflow.
Algorithm 5.9: TissueInteractions for Recirculation Tissue Clonal Selection.

Input: H, Nmigration
1 H& ← 0;
3 Ti & ← Sample(Ti , Nmigration );
4 H& ← Ti &;
5 end
6 foreach Ti & ∈ H& ∧ Ti ∈ H do
7 if Ti ≡ TN then
8 Integrate(Ti , T1 &);
9 else
10 Integrate(Ti , Ti+1 &);
11 end
12 end
The Recirculation Tissue Clonal Selection Algorithm (RTCSA) is defined (Algorithm 5.9)
as a specialisation of the T issueInteractions operation in TCSA in which a sample
T i& is taken sample(Ti , Nmigration ) of Nmigration cells from each Tissue T and trans-
181
mitted to a single and consistent neighbouring tissue Ti+1 where they are integrated
integrate(Ti+1 , Ti &). The RTCAS is configured with a remove-insert migration strategy
that selects a consistent amplitude (Nmigration ) random sample of cells for migration,
which is applied each epoch of the algorithm (fixed frequency). Each tissue in the system
participates in the trafficking which is applied in a lock-step manner of selection-removal-
insertion after the exposures of the epoch have been satisfied. The default realisation of
the Sample(Ti , Nmigration ) operation is to take a random sample (Ti &), and the default
realisation of the Integrate(Ti+1 , Ti &) operation is concatenation to the repertoire filling
the positions made by the sampling operation.
5.5.3 Recirculation Empirical Study

Aim
The aim of this empirical study was to investigate RTCSA as a viable mechanism for
realising the ‘Consistency of Response’ emergent effect. Toward this end, the study had
the following goals:
1. Confirm the expectations outlined in Section 5.4.5 using the RTCSA and the MTCSA.
2. Consider intermediates as outlined in Section 5.4.5 using variable migration sample

size (Nmigration ) and the variety of Tissue Exposure Regimes.
Method
Algorithms The study considers two algorithms, the Minimal Tissue Clonal Selection
Algorithm (MTCSA), and the Recirculation Tissue Clonal Selection Algorithm. MTCSA
is a specialisation of the TCSA defined in Algorithm 5.7, that was configured with Ntissues =
10. Each T was an instance of the Replacement Cellular Clonal Selection Algorithm
(RCCSA) defined in Algorithm 4.5, with the configuration Ncells = 50, Nselection = 1,
and Nclones = 5. RTCSA is a specialisation of TCSA with a T issueInteractions(H) that
connects all T into a directed cyclic graph defined in Algorithm 5.9. The same Ntissues
and RCCSA configuration was adopted as was used in MTCSA. Three different migration
sizes were used including: small (RTCSA-SML) with Nmigration = 5 (10% of each T ),
medium (RTCSA-MED) with Nmigration = 20 (40% of each T ), and large (RTCSA-LRG)
with Nmigration = 40 (80% of each T ).
Problems The colour space specialisation of the Infection Antigenic Exposure Paradigm
(IAEP) defined in Algorithm 5.1 was used called the Infection Colour Space Problem
(ICSP). The minimal variation of ICSP was used with Ninf ections = 10, and Ndeterminants =
Nantigen = 1. Each I was a Colour Space Pattern (CSP) randomly generated at the
beginning of each run. The five different Tissue Exposure Regimes (TER) defined in
Section 5.4.1 were used for the ICSP. These included the ATER where the number of T
matched the number of I (one-to-one), STER, RTER, PTER where Tpoint was fixed at
tissue 1, and OTER with duration = 15.
182
fined in Equation 4.8 with M axEpochs = 1000. The four tissue specific measures defined
in Section 5.4.3 were collected from the state of the system after the triggering of the stop
condition. These included the system measures HE and HD, and the tissue measures ATE
and ATD. Each algorithm and problem received a new and different random number seed
each run. Algorithm-Problem combinations were repeated 30 times.
Results
Table A.1 in Appendix A.1.1 provides a summary of results for each algorithm-problem
combination including the mean (x̄) and standard deviation (σ) of collected measure val-
ues. Box-and-whisker plots are provided in which the results for each algorithm are aggre-
gated across all TER for a each measure. Figure 5.20 shows HD, Figure 5.21 shows HE,
Figure 5.22 shows ATD, and Figure 5.23 shows ATE.
Figure 5.20: Box-and-whisker plot of Host Diversity (HD) across all TER for the RTCSA
study.
Analysis
This section provides an analysis of the results summarises in the previous section. Ta-
ble 5.7 was used to consider the general behaviour of each of the specific TER’s. Table 5.6
was used to consider the general behaviour of the global and local information management
strategies of RTCSA and MTCSA. Finally, Table 5.8 was used to qualitative interpret the
summarised results.
183
Figure 5.21: Box-and-whisker plot of Host Error (HE) across all TER for the RTCSA
study.
Figure 5.22: Box-and-whisker plot of Average Tissue Diversity (ATD) across all TER for
the RTCSA study.
184
Figure 5.23: Box-and-whisker plot of Average Tissue Error (ATE) across all TER for the
RTCSA study.
Information Management Trends This section considers the expected trends of the
local and global information management trends under local and global information ex-
posure regimes outlined in Table 5.9 from Section 5.4.4. ATER and RTER are considered
local and global information exposure regimes (exposure and re-exposure) respectively,
and MTCSA and RTCSA-LRG are considered local and global information management
(anticipation) respectively. Table 5.10 provides a variation on the expected trends sum-
mary presented in Table 5.9 in which the specific measures of the selected algorithms are
provided with qualitative comparison. Direct comparison between these tables addresses
the primary aim of the experiment, to confirm or reject the behaviour expectations out-
lined in the tissue paradigm agenda in Section 5.4.5.
Measure ATER RTER

RTCSA-LRG MTCSA Sig. RTCSA-LRG MTCSA Sig.
HD 562.136 (↓) 1610.68 (↑) True 503.919 (↓) 2313.88 (↑) True
HE 0.152 (↑) 0.013 (↓) True 0.114 (↓) 0.121 (↑) False
ATD 76.969 (↓) 85.903 (↑) True 75.347 (↑) 74.7 (↓) False
ATE 0.152 (↓) 0.155 (↑) False 0.111 (↓) 0.123 (↑) True
Table 5.10: Summary of the results for MTCSA and RTCSA-LRG on ATER and RTER
and the relative quantitative difference is measures.
Considered first are the algorithm results for the local information exposure regime
(ATER). From a system perspective both the HD and HE met the expectations in which
the MTCSA resulted in decreased error and increased diversity, whereas the RTCSA-LRG
185
resulted in the inverse behaviour of increased error and decreased diversity. From a com-
ponent perspective the comparison in diversity (ATD) did not meet expectations. Instead
of an increase in average component diversity with a global strategy, this behaviour was
observed by the local strategy. This means that the expectation that the global strat-
egy would decrease component specialisation was interpreted incorrectly, demonstrating
that a general increase in specialisation with a local strategy results in an increase in
the diversity within the average component. Also unexpected was the lack of significant
difference between the average component error between the two strategies. This was a
surprising result as the mixing of acquired information provided by RTCSA-LRG provided
no measured performance improvement over segregated repertoires on the local exposure
regime.
The algorithm results from the global information exposure regime (RTER) are also
considered. From a system perspective the diversity matched the expectation in which
the segregation of MTCSA resulted in an increase in diversity compared to the decreased
provided by recirculation in RTCSA-LRG. The expectation that system error would de-
crease with RTCSA-LRG was not confirmed as there was no significant difference between
the two strategies. This is a surprising result as it generally suggests that consistency
of response may not improve system performance under an irregular global information
exposure regime2 . From a component level the expectation of component error was con-
firmed in which the global strategy resulted in a decrease in HE compared to an increase
seen with the local strategy. The expectation of component diversity was not confirmed
as their was no significant difference between the two strategies. This finding, together
with the lack of significant difference in HE demonstrates that RTER alone was sufficient
to engender a ‘consistency of response’ irrespective of the information management strat-
egy, promoting a reconsideration of cause and effect between management strategies and
exposure regimes. This is considered in the remainder of this section. Table 5.11 provides
a reproduction of the expectations from Table 5.9 of global and local information manage-
ment strategies and information exposure regimes, with those results that did not meet
the expectations highlighted.
Measure Local-Local TER Global-Global TER

Scope Type Global Strategy Local Strategy Global Strategy Local Strategy
System Diversity Decreased Increased Decreased Increased
System Error Increased Decreased Neutral Neutral
Component Diversity Decrease Increase Neutral Neutral
Component Error Neutral Neutral Decreased Increased
Table 5.11: Summary of confirmed behaviours of global and local information strategies
and exposure regimes, highlighting where expectations were incorrect.
Information Exposure Trends This section relaxes the notions of local and global
information exposure regimes and considers general trends across the TER’s in the con-
2
This suggestion is further supported by the lack of significance between MTCSA and all tested varia-
tions of RTCSA on RTER as shown in Table A.1.
186
text of both the local and global management strategies (MTCSA and RTCSA-LRG re-
spectively). Table A.2 in Appendix A.1.1 provides a restricted summary of results that
provides direct comparison and statistical significance between TER’s for the local and
global management strategies in which the effects of the exposure regimes may be teased
from their information management strategies.
Observations may be drawn firstly from the behaviour of the five TER’s with the
local management strategy (MTCSA). From a system perspective, PTER resulted in the
least diverse system whereas RTER resulted in the most diverse system. This suggests
that the holistic adaptation the system was exposed to in RTER and other system-wide
exposure regimes (ATER, STER and OTER) result in a system information content that
is more diverse than the random repertoires under PTER (besides the single exposed
tissue). ATER achieved the lowest error, which is unsurprising given the T -I specialisation
fostered by the TER, and RTER achieved the highest error which was also expected given it
provides the least T -I specialisation (the most irregular regime). From a component level,
PTER achieved the highest ATD and ATE3 suggesting that from a component perspective
point-wise exposure of information results in the expected behaviour of repertoires that
individually perform poorly on average. This behaviour was not restricted to PTER as was
also demonstrated to a lesser degree on the other restricted exposure regimes (ATER and
OTER). RTER resulted in the lowest average component diversity, although with similar
behaviour exhibited by STER, suggesting system-wide exposure of information results in
repertoires that have consistent internal information content on average. STER achieved
the lowest ATE followed by RTER demonstrating that the same effect on component
diversity results in improved tissue capability.
Observations may be summarised from the behaviour of the five TER’s with the global
management strategy (RTCSA-LRG). From a system perspective STER resulted in the
lowest diversity and error. RTER resulted in the highest diversity although was not sig-
nificantly different from many of the other TER’s. Interestingly, PTER resulted in the
highest error suggesting that global management presented a major disruptive influence to
localised exposure regimes, demonstrated by the equally poor HE for ATER and OTER
compared to the much improved STER and RTER. Further, this inverse in behaviour
from the local management strategy results demonstrate the suitability of holistic expo-
sure regimes to the global management strategy, and localised exposure regimes to the
local management strategy. From a component level, PTER achieved the highest average
component diversity and error demonstrating the same trend with the MTCSA manage-
ment strategy. RTER resulted in the lowest ATD with no significant difference in result
with the STER and ATER system-wide exposure regimes. Likewise STER resulted in the
lowest ATE followed closely by RTER mirroring the pattern observed with MTCSA.
Toward teasing apart the influences of TCSA’s and TER’s the observations may be
reduced into the following trends that describe the generalised relationship between infor-
mation exposure strategy and information management strategy:
1. Component Trends are Strategy Invariant

3
Difference in ATE was not significant between PTER and ATER for MTCSA.
187
(a) Localising information exposure increases average component diversity, global-
ising information exposure decreases average component diversity.
(b) Localising information exposure increases average component error, globalising
information exposure decrease average component error.
2. System Trends are Strategy Variant
(a) Globalising information management increases diversity and error under lo-
calised information exposure, and decreases diversity and error under globalised
information exposure.
(b) Localising information management decreases diversity and error with localised
information exposure, and increases diversity and error with a globalised expo-
sure strategy.
The component measure invariance trend with information management strategy demon-
strates that the information exposure regime is the determining factor. This trend was ob-
served independently for both the MTCSA and RTCSA-LRG strategies across the TER’s,
although specific and typically minor differences for strategies on specific TER’s. The
system measure dependence on information management strategy demonstrates the suit-
ability of a localised strategy on localising information exposure and the suitability of
a globalised strategy on globalising information exposure. This trend was observed in-
dependently and complementary (the inverse case) for both MTCSA and RTCSA-LRG
strategies across the TER’s, although with differences in the specific TER in each case.
An important implication regarding TER’s in addition to that of their spatial-temporal
regularity is the scope of their exposure behaviour as either system-wide or restricted,
providing a defining quality for localisation and dissemination management strategies.
These trends suggest that TER’s can predictably influence the component-wise organisa-
tion of information in terms of holistic problem error and information content diversity.
The system-level suitability trends also suggest that MTCSA does demonstrate the spa-
tial organisation (localisation) of information emergent effect and that RTCSA-LRG does
demonstrate the consistency of response (dissemination) emergent effect, confirming the
expectations regarding each approaches mechanism.
Recirculation Trends This section considers information management strategies be-

yond the two extremes already considered. Specifically, this section considers the range of
recirculation sample sizes (Nmigration ) as intermediates between the two extremes, across
all five TER’s. These recirculation trends are drawn from the principle result summarises,
specifically Table A.1, and Figures 5.20, 5.21, 5.22 and 5.23.
From the system perspective, migration size demonstrated the consistent trend of
decreasing diversity (HD) where the larger the sample size the larger the decrease in HD
compared to MTCSA. A consistent trend was also demonstrated with error, where the
increase in recirculation size resulted in increased system error compared to MTCSA.
This generalised error trend applied to all TER’s with system-wise exposure scope (all
tissues participated, not PTER). These system-level diversity and error trends provide
188
clear intermediates in system-level behaviour (localisation and dissemination) as defined
by the captured measures HD and HE. From a component perspective the trends are
less bold as a the system level. ATD demonstrate no clear TER invariant trends with
recirculation size, although there was clear commonality in behaviour. For both the ATER
and OTER exposure regimes, the increase in recirculation size resulted in a step decrease
in average component diversity that seemingly levelled off after the drop from MTCSA
to RTCSA-SML. This demonstrates that a sufficiently small sample size was required to
effect a decrease in the average tissue diversity under those regimes, after which further
increases in sample size had little effect. Interestingly, a different effect was observed for
RTER and STER where a small recirculation size resulted in a small increase in ATD, after
which further increases in size resulted in decreases. These effects were not observed with
PTER, where the increase in recirculation size resulted in the progressive increase in ATD.
The variation in recirculation size demonstrated a common trend across all TER’s except
PTER for ATE, where a small increase resulted in a decreased error, and subsequently
larger recirculation sizes increased error. This trend demonstrates that a sufficiently small
recirculation size is needed to effect a drop in AE under those exposure regimes. The
inverse of this trend was observed with PTER where small recirculation resulted in an
increase in ATE and larger sample sizes decreased this error toward the MTCSA score.
The observed measure-based may be reduced into the following generalised trends
regarding RTCSA recirculation size:
1. Graceful Intermediate Dissemination Effect
(a) The RTCSA recirculation size provides a graceful intermediate in emergent

behaviour (system-level) between localisation demonstrated in MTCSA and
dissemination.
(b) Small recirculation size provides a decrease in average component error where
large recirculation sizes do not on system-wide exposure regimes with local and
global re-exposure.
(c) Recirculation decreases the diversity of the average component on system-wide
exposure regimes with local re-exposure.
(d) Recirculation increases the diversity of the average component on system-wide
exposure regimes with global re-exposure.
Conclusions
This section summarises the findings of the empirical study into the Recirculation Tissue
Clonal Selection Algorithm, in terms of the primitives that were the focus of the study
and the expectations that motivated the study.
1. Study Primitives
(a) The MTCSA is a viable realisation of the spatial organisation effect (localisa-
tion) via an isolation mechanism.
189
(b) The RTCSA is a viable realisation of the consistency of response effect (dis-
semination) via a recirculation mechanism.
2. Confirmations and Findings
(a) The behavioural expectations for the two different management strategies under
two archetype exposure regimes were considered with a mixture of system and
component level expectations being confirmed and rejected.
(b) The localisation or globalisation (exposure/re-exposure) of an exposure regimes
defines the component-level error and diversity that is generally invariant to the
information management strategy
(c) The localisation or globalisation of an exposure regime is suited to a localising
and disseminating information management strategy with regard to system-
level error and diversity.
(d) In addition to the regularity and holistic information exposure/re-exposure
properties of tissue exposure regimes, a third important consideration is the
scope of system exposure, specifically system-wide or constrained exposure/re-
exposure.
3. Intermediates
(a) Recirculation size of the RTCSA provides a viable behavioural intermediate

between localisation and dissemination with predictable systemic effects and
identified component tends on varied exposure regimes.
(b) There is a need for a system to balance localisation and globalisation of acquired
information when presented with an unknown exposure regime.
This final point about intermediates motivates the two remaining investigations in this
chapter, and likely the whole Tissue Clonal Selection Paradigm.
5.6 Lymphocyte Homing

5.6.1 Homing Metaphor and Strategy
Section 5.2.2 considered the directed trafficking of cells, with the specific example of the
imprinting of T-cells by dendritic cells as to the chemical properties of a site of infection.
The educated cells then recirculate until they locate the imprinted chemical signature
and seek their cognate antigen. The RTCSA demonstrated that continual lymphocyte
migration provides a background disruptive behaviour to a tissues ability to specialise a
response to regular exposures. The education and homing mechanism for cells may be
generalised to a specialisation of recirculation that empower cells to express a preferential
residence in tissues for which they may be useful. This bottom-up behaviour is called
a Preferential Residence Strategy and is comprised of Cell Imprinting and Differential
Migration mechanisms to realise the emergent effect.
190
In addition to specialising the resources of the system to the antigenic exposures, the
specialised resources (lymphocyte receptors for antigen) must be positioned such that they
are used in the right locations. The distribution (space and time) of antigenic exposures
is unknown to the system (defined by a TER), thus the system recirculates specialised
resources between the locations of exposure to maximise the chance application of the
resource in an exposure event. Lymphocytes interact with each other and interact with
antigen in this spatial structure. Each tissue type associated with a region of the host has
its own unique chemical makeup, which may be used to differentiate tissues types and thus
regions of the host. Lymphocytes may interact with tissues using specialised receptors that
detect and differentiate between tissue types. This interaction occurs when the lymphocyte
recirculates around the host, and may be used by the cell to travel specific pathways and
to enter and remain in a specific tissue in search of its cognate antigen. These navigation
receptors are not expressed in naı̈ve lymphocytes, instead are expressed in matured cells
that are produced after an interaction with antigen. A mature receptor (memory or
effector) expresses navigation receptors for the tissue surroundings when it is conceived.
Therefore, mature cells are specialised both for an antigenic stimulus, and for the spatial
location (tissue) that is known to have been exposed to the antigenic stimulus. This Cell
Imprinting process occurs irrespective of whether a selected (high-affinity) lymphocyte is
mature or naı̈ve. In effect, specialised resources may be re-educated in that their progeny
may be imprinted with a different location than themselves (presuming the cell was selected
in a tissue different from that it may home to). This imprinting mechanism facilitates both
the spatial anticipation of antigen exposure based on past exposures, and in addition to
recirculation, it facilitates the use and re-education of lymphocytes.
Imprinting empowers individual mature lymphocytes to assert preference for their lo-
cations allowing tissues to discriminate preferential residence during the selection of cells
to traffic. Homing requires an imprinting apparatus, which includes both a unique identity
or signature for tissues, and a way for lymphocytes to be impregnated with that signa-
ture. The match of information between a given lymphocyte and a tissue may be used
to bias the selection of cells from the repertoire to migrate. One may consider a Boolean
match/no-match for lymphocytes in their home tissue, and those not in their home tis-
sue. Those progeny lymphocytes created as a result of the clonal selection process are
already in their home tissue, thus they have a match to their tissue’s signature. All those
lymphocytes that migrate from another tissue will not match. Thus, a migration sample
selection policy that biases toward no-matches and away from matches (1) exercises a pref-
erence for lymphocytes (2) facilitates an automatic recirculation population of homeless
lymphocytes. This process is called Differential Recirculation where the probability that
a mature lymphocyte is selected in an outbound migration sample is biased by the match
between a given lymphocyte and the tissue’s signature. This mechanism facilitates homing
but introduces the problem of recruitment into the recirculation pool. Those cells in the
recirculating pool will continue recirculating until they return their ‘home’, unless they
get used or killed elsewhere. The initial recruitment probability will be uniform across
the repertoire as there are no migrant cells to consider. If a recirculating cell is replaced
191
(killed) by the proliferation strategy at another (non-home) tissue, then that tissue will
fill the gap in the recirculating population.
5.6.2 Homing Tissue Algorithm

The Homing Tissue Clonal Selection Algorithm (HTCSA) is defined as an extension of
the RTCSA with the addition of a cell imprinting and differential migration mechanisms.
Cell Imprinting A pattern-recognition method may be used that is somewhat similar to

the approach used for Mediated Cellular Clonal Selection (Section 4.6) or Network Cellular
Clonal Selection (Section 4.7), although rather than a lymphocyte matching onto other
lymphocytes with a mock ‘surface feature’, lymphocytes are provided with an additional
receptor for detecting and matching the tissue they are in. In keeping with the bitstring
basis of the minimal clonal selection algorithm (and extensions), a simple and scalable
method is to assign each tissue (or tissue group) a random binary string. Each mature
lymphocyte (that is a lymphocyte created from a cloning and maturation process) is then
assigned (imprinted) with a copy of the tissue bitstring specific to the location of its
creation. A less elaborate approach involves assigning a distinct identification number to
each T in H and assigning a tissues identification number to activated cells at the time
of exposure. Algorithm 5.10 defines a generic tissue exposure function in which the cell
repertoire is ordered by its affinity for an Antigen A (OrderByAf f inity(T, Ci )) and a
subset of the Nimprint highest affinity cells for A are imprinted with the current tissue
identification number.
Algorithm 5.10: Cell Imprinting for Homing Tissue Clonal Selection.

Input: I, T, Nimprint
1 foreach Ai ∈ I do
2 foreach Cj ∈ T do
3 Exposure(Cj , Ai );
4 end
5 OrderByAffinity(T, Ci );
6 for j←1 to Nimprint do
7 Cj .TissuePreference ← T.id;
8 end
9 end
Differential Migration The T issueInteractions(H) operation of the RTCSA that gov-

erned tissue interactions in the general TCSA may be specialised to exploit the imprinted
cells by (1) migrating those cells that are not in their preferred tissue or do not have a
preference, and not migrating those cells that are in their preferred tissue. Such pref-
erences may be deterministic or probabilistically implemented. Algorithm 5.11 provides
a re-definition of the T issueInteractions defined for RTCSA in Algorithm 5.9 that sim-
plifies differential migration by not migrating those cells in their preferred tissue. The
SelectSingleCell operation draws cells randomly from the T .
192
Algorithm 5.11: Differential Migration for Homing Tissue Clonal Selection.
Input: H, Nmigration
1 H& ← 0;
3 Ti & ←0;
4 while Ti &N #= Nmigration do
5 C& ← SelectSingleCell(Ti );
6 if C&.TissuePreference #= Ti .id then
7 Ti & ← C&;
8 end
9 end
10 H& ← Ti &;
11 end
12 foreach Ti & ∈ H& ∧ Ti ∈ H do
13 if Ti ≡ TN then
14 Integrate(Ti , T1 &);
15 else
16 Integrate(Ti , Ti+1 &);
17 end
18 end
5.6.3 Homing Empirical Study

Aim
The aim of this empirical study was to investigate the Homing Tissue Clonal Selection
Algorithm (HTCSA) as an intermediate information strategy between high migration
RTCSA and the segregated MTCSA. Toward this end, the study had the following goals:
1. Contrast HTCSA against MTCSA and RTCSA as a mechanism for simultaneously

promoting dissemination and localisation of acquired information.
2. Investigate intermediates of the HTCSA mechanism and subsequent effect by varying

the Nimprint parameter.
Method
Algorithms The study considers three algorithms, the Minimal Tissue Clonal Selection
Algorithm (MTCSA), and the Recirculation Tissue Clonal Selection Algorithm (RTCSA),
and the Homing Tissue Clonal Selection Algorithm (HTCSA). The MTCSA and RTCSA
algorithm are configured as was defined for the RTCSA empirical study in Section 5.5.3
with Ntissues = 10, Ncells = 50, Nselection = 1, Nclones = 5 for MTCSA and RTCSA,
and Nmigration = 5 (10% of each T ) fixed for the RTCSA. HTCSA is an extension of
RTCSA with the addition of an imprinting operation as defined in Algorithm 5.10 for
each exposure and a differential migration as defined in Algorithm 5.11. The imprinting
mechanism was specialised such that a variable number Nimprint of the best-matching
cells were imprinted for each exposure, as follows: small (HTCSA-S) Nimprint = 1 (2% of
T ), medium (HTCSA-M) Nimprint = 5 (10% of T ), and large (HTCSA-L) Nimprint = 10
193
(20% of T ). A probabilistic sample selection scheme was used in differential migration,
changing the configuration from 100% rejection if a cell was housed within its ‘home’ tissue
to a probabilistic 20% (80% chance of home cells not being migrated per migrating cell
selection).
Problems The same Infection Colour Space Problem and Tissue Exposure Regimes
were used as was defined for the RTCSA empirical study in Section 5.5.3.
Experiment The same experimental setup was used as was defined for the RTCSA
empirical study in Section 5.5.3.
Results
Figure 5.24: Box-and-whisker plot of Host Diversity (HD) across all TER for the HTCSA
study.
Analysis
This section analyses the summarised results in Table A.3 in the context of the expecta-
tions from Section 5.4.4 and their confirmation and other findings regarding MTCSA and
RTCSA from the empirical study in Section 5.5.3.
194
Figure 5.25: Box-and-whisker plot of Host Error (HE) across all TER for the HTCSA
study.
the HTCSA study.
195
HTCSA study.
Localisation and Dissemination Trends This section considers a comparison in be-

haviour between the the localising MTCSA, the disseminating RTCSA-S and preferen-
tial residence HTCSA-L. From a system perspective HTCSA-L performed in a similar
way to that as RTCSA-S, both of which were contrastingly different from MTCSA. Both
RTCSA-S and HTCSA-L were less diverse than MTCSA, although RTCSA-S more so than
HTCSA-L. MTCSA achieved consistently lower HE, although HTCSA-L demonstrated a
decreased error on constrained exposure regimes (ATER, OTER and PTER) compared
to RTCSA-S, and no significant difference in error on the remaining (RTER and STER)
exposure regimes. This was an important finding highlighting the increased spatial organ-
isation capability for recirculation with preferential residence over recirculation without
preferential residence on regular exposure regimes. From a component perspective the
performance of HTCSA-L was distinct from MTCSA and similar to RTCSA-S, with gen-
erally little significant difference in the use or non-use of preferential residence. ATD was
increased in HTCSA-L over RTCSA-S for the same constrained exposure regimes in which
an improvement in system error was observed (ATER, OTER and PTER), and no signif-
icant difference for the remaining regimes. This result highlights that under regular and
constrained exposures (different information in different locations in a regular manner),
preferential residence changes the composition of components toward that of segregated
components, whilst under the presence of recirculation, an effect that is not present for
other exposure regimes. Specifically, preferential residence organises information toward
the benefit of the system, when it is possible to do so as determined by the exposure
regime. Homing had no consistent significant effect on recirculation regarding Average
Tissue Error. The observations comparing and contrasting the three strategies may be
reduced to the following general behavioural trends for the preferential residence strategy:
196
1. Homing reduces system error over recirculation on constrained (regular) exposure
regimes whilst remaining consistent with recirculation on irregular and system-wide
exposure regimes.
2. Homing increases system diversity (organisation) compared with recirculation.
3. Decreased system error is reflected in increased average component diversity on

constrained exposure regimes over recirculation, whilst remaining consistent with
recirculation on irregular and system-wide exposure regimes.
4. Improved organisation of homing does not effect average component error compared
with recirculation.
Information Exposure Trends Section 5.5.3 identified the two trends that (1) average
component measures of error and diversity are defined by the localisation (constrained) or
globalisation (system-wide) of the exposure (generally invariant to the information man-
agement scheme), and (2) that globalising management is suited to system-wide expo-
sure regimes (global-global) and localising management is suited to constrained exposure
regimes (local-local). This section considers the HTCSA in the context of these trends,
specifically the degree to which the trends hold or are influenced by preferential residence.
Table 5.12 summarises results for HTCAS-L across all the TER’s, providing a complement
to the same presentation for MTCSA and RTCSA-L in Table A.2.
System Problem HD HE ATD ATE

TCSA TER x̄ σ x̄ σ x̄ σ x̄ σ
HTCSA-L ATER 816.023 53.374 0.084 0.019 83.046 2.228 0.15 0.018
HTCSA-L OTER 756.301 64.802 0.1 0.019 81.077 3.568 0.146 0.015
HTCSA-L PTER 685.395 27.472 0.145 0.017 93.206 0.161 0.17 0.018
HTCSA-L RTER 673.063 58.672 0.114 0.022 75.595 3.745 0.114 0.018
HTCSA-L STER 995.975 114.778 0.046 0.013 79.731 2.071 0.09 0.022
Significance Truea True Trueb Truec
a
False for PTER and RTER
b
False for STER and OTER
c
False for ATER and OTER
Table 5.12: Summary of results from an empirical study into the HTCSA comparing
TER’s with the HTCSA-L information management strategy.
The results show the same trend with HTCSA-L as was shown with recirculation with
a larger sample size in Table A.2 (RTCSA-LRG rather than RTCSA-SML that would be a
fair comparison with any assessed HTCSA). This confirmation supports the comparisons
of RTCSA, MTCSA, and HTCSA and outcomes in the previous section. Specifically
demonstrating that the influence of the TER’s on the system remained generally consistent
with RTCSA, and that the differences between RTCSA and HTCSA are as a result of the
HTCSA strategy.
197
Imprinting Size Trends This section considers the intermediate effects of varying the
number of cells imprinted with each exposure (Nimprint ), and thus the extent of each tissue
that is empowered to express a preference during migration. From a system perspective
some preferential residence was better than none, although the increases in Nimprint gen-
erally did not have a measurable effect. This result was generally demonstrated for both
Host Error and Diversity across all exposure regimes except PTER. The Point Exposure
Regime demonstrated a clear and steady increase in the organisation effect of preferential
residence with the increase in the number of cells imprinted per exposure. This suggest
that homing can counter the disruptive effects of recirculation, a behaviour that gracefully
increases with the number of imprinted cells per exposure. This effect is most pronounced
with PTER because it is the regime that is the most susceptible to the disruptive effects of
blind recirculation. From a component perspective a similar general result was observed
for average component diversity where preferential residence increased diversity with little
difference between the value of Nimprint , and only on the ATER and OTER regimes with
no significant effect on the other regimes. No consistent significant effect was observed for
average component error in varying Nimprint , including a value of zero (RTCSA-S). The
observations regarding the variation of the number of imprinted cells per exposure may
be generalised to the following trends:
1. Homing is generally better than no homing (recirculation) and the specific choice of
Nimprint is not critical.
2. Increase in effect with increase in Nimprint is pronounced on constrained local expo-

sure (PTER), although is not significant under other exposure regimes.
An interesting consideration for future investigation is the correlation of decreased sys-

tem diversity and increased system error with migration size, and the general increase in
organisation and decrease in error with preferential residence. In particular the knowledge
that the homing effect does scale when isolated in PTER. One may investigate the rela-
tionship between the recirculation size (Nmigration ) and the imprint size (Nimprint ) in order
to find general configuration trends toward a trade-off and balance of the dissemination
and preferential organisation effect.
Conclusions
This section summarises the findings of the empirical study into the Homing Tissue Clonal
Selection Algorithm, in terms of the primitives that were the focus of the study and the
1. Homing
(a) The preferential residence (homing) effect realised in the imprinting and differ-
ential migration mechanisms provides a strategy that can exploit the dissemi-
nation of information by recirculation whilst countering the effect by localising
used information.
198
(b) Cell localisation (20% of a repertoire) is insufficient to counter all of the dis-
ruptive effects of continuous recirculation.
(c) Localising used information allows progeny of such information to freely (blindly)
recirculate.
This final point provides a suggestion as to the basis of the mechanism in the homing
approach, and a basis for increasing the probability of migration for cells without a pref-
erence. The second important finding for the homing approach (besides it behaving as
intended) is that even with a large number of cells imprinted each exposure, it is unable
to completly counter the disruptive effects of continuous recirculation, rather it is limited
to complementing the process. Countering the disruptive effects would likely require the
imprinting of some or all of the progeny of exposures, which is expected to impact the
dissemination properties of recirculation. This theme of allowing sufficient specialisation
under regular exposure regimes whilst also disseminating acquired information is further
explored via a Tissue Inflammation metaphor in the next section.
5.7 Tissue Inflammation

5.7.1 Inflammation Metaphor and Strategy
When a tissue is infected by a pathogen that damages the local area, the response is
inflammation at the site. Inflammation has many symptoms, not limited to swelling and
the increased carrying capacity of lymphocytes in the tissue. In addition, the blood vessels
that lead to the area dilate, increasing the flow of lymphocytes (information) to the site,
and typically involves the use of chemical signals or attract recirculating cells and/or arrest
cell movement. An increase in carrying capacity allows more lymphocytes created at the
site of infection to remain there, and more cells to migrate and home into the tissue and
collect at the site of infection. Inflammation provides a mechanism to temporally decrease
the competition between lymphocytes for limited resources within a tissue, providing a
tissue-level direct recruitment mechanism. One may consider the tissue paradigm from the
perspective of active distributed recruitment strategies as opposed to passive distributed
information organisation strategies (already considered). Information is streamed from
the tertiary tissue to the secondary tissue although the streaming is facilitated by courier
cells. These cells are aware of where they collected their material and thus are capable of
imprinting that information onto the produced effector cells. The result is an effect that
differentiates the lymphocyte movement behaviour from that of Recirculation where cells
blindly seek their cognate antigen, and from Homing where cells home into the chemical
signature of infected tissue. The Homing model couriers information about the location
of the pathogen invasion to the cells which are imprinted with that information. This is a
‘fire and forget’ response strategy. Like the recirculating of effectors, the homing effectors
recirculate the system in search of their cognate antigen, although if the cells detect the
imprinted signature of the site of infection, they take up residence and provide a local
search strategy. Further, less effectors are created than in the recirculation model (for
199
example mature T-cells as opposed to antibodies), thus the response is both smaller and
more directed. One may generalise two response strategies as follows:
• Undirected Response: Flood the system with effectors that will generally address the
specific site of invasion and any other sites where the pathogen may have spread, or
penetrated on multiple fronts (recirculation method).
• Directed Response: Produce a smaller set of effectors that home into the specific
site of infection, and have a low chance of addressing additional sites of infection
(homing method).
Both the undirected recirculation model and the directed homing model allocate (re-
cruit) resources to the requirements of a specific antigenic exposure. Further, both strate-
gies address the needs of the distributed information environment (resident receptors and
inbound pathogen information). These two approaches may be considered to describe
Distributed Recruitment Strategies. The undirected approach is a flooding strategy that
does not use any structural information other than the generation of a response in the
vicinity of the pathogens first detection. The directed approach is a homing strategy in
which structural information is used directly to get the resources to where they are re-
quired. Recruitment may be phrased as resource allocation with the desire of maximising
payoff. Thus, Distributed Recruitment may be defined as a strategy in which informa-
tion dispersed across a spatial structure is sought (for example selected) and employed
for a specific task (antigen neutralisation). The two proposed models both relate to the
changing of lymphocyte trafficking behaviour, treating the lymphocyte (a space in the
repertoire) as a limited resource.
5.7.2 Inflammation Tissue Algorithm

Inflammation is a strategy that promotes making more information available in the local
area to address the infection, and there relaxing the competition between receptors for
limited positions in the repertoire. This effect may be realised by temporary increasing
the capacity of a tissue for lymphocytes around the time of exposure. The increase in
carrying capacity would be stepwise at the time of exposure such that the clones created
at the time of exposure may be housed in the tissue with less competition. The decrease
in carrying capacity would not be stepwise, rather it would decay over time (for example
as a linear function of the increased size over time). Ideally, the capacity increase would
facilitate the approximate optimal amount of local information needed to address the
exposure. The decay rate would facilitate the approximate optimal amount of time that
the local information may be required (for example antigen endurance or antigen exposure
frequency). The amount of inflammation (increase in carrying capacity) may be a fixed
amount (step size) and the decay function may also be fixed. In addition, an upper limit of
lymphocytes for a site is required (upper bound on space complexity). For each exposure,
the increase would be applied. The effect would be the allocation of space for lymphocytes
temporally and proportional to exposures. Thus, inflammation allows space complexity to
200
be treated as a resource whose allocation is adapted by the system based on an exposure
(needs) basis. Adaptation of lymphocyte capacity allows not only increases when required,
but also decreases when it is not needed. The following lists some additional mechanisms
that may complement a local increase of the carrying capacity of a tissue:
• Lymphocyte Trapping: Lymphocytes may be prevented from leaving a site of in-

fection for a temporally time. This provides symmetry to the increase in carrying
capacity, which puts less pressure on the cells in the present tissue. This pressure is
reduced further by migrating (out-bound) fewer lymphocytes.
• Upstream Dilation: The tissues ‘upstream’ of an exposed tissue may be influenced

to increase their down-flow migration of lymphocytes. Upstream and down-stream
refers to the uni-directional flow of lymphocytes around the directed cyclic graph
structure in which tissues are arranged. Increased flow of lymphocytes provides
more localised information to address the exposure.
• Effector Discrimination: Tissue behaviour may be extended further such that it

may actively discriminate as to the effectors to accept and integrate into the local
repertoire from migration, and those that are decidedly not useful during an exposure
may be migrated away from the tissue. Thus, tissues may be given the ability to
discriminate between effector lymphocytes.
Algorithm 5.12: Increased Capacity for Inflammation Tissue Clonal Selection.

Input: B, H, Nmaxcells
2 H& ← SelectTissuesToExpose(H);
3 foreach Ti & ∈ H& do
4 Exposure(Ti &, Ii );
5 Ti &.Ncells ← Nmaxcells
6 end
7 end
The Inflammation Tissue Clonal Selection Algorithm (ITCSA) is defined as an exten-

sion of the TCSA with the addition of a mechanism to increase the carrying capacity of
exposed tissues and a mechanism to decay carrying capacity of expanded tissues back to
default levels. Algorithm 5.12 defines a Habitat B exposure of a Host B in which the role
of the TER is generalised to that of the selection of subset (Ti &) of H to expose to each Ii of
B. This generalisation allows the explicit increase in carrying capacity of selected and thus
exposed tissues from the default Ncells to a maximum Nmaxcells . This additional carrying
capacity is filled by progeny created from future exposures in a given expanded tissues,
although alternatively may be used to integrate recirculated cells from a RTCAS special-
isation of ITCSA’s T issueInteractions operation. Algorithm 5.13 provides a definition
of a default T issueInteractions operation for ITCSA in which the carrying capacity of
expanded tissues is progressively decreased by one cell each host exposure (epoch). This
is realised by the selection and remocal of a randomly selected cell from the repertoire.
201
Algorithm 5.13: TissueInteractions for Inflammation Tissue Clonal Selection.
Input: H, Ncells
2 if Ti .Ncells > Ncells then
3 Ti .Ncells ← Ti .Ncells − 1;
4 end
5 end
5.7.3 Inflammation Empirical Study

Aim
The aim of this empirical study was to investigate the Inflammation Tissue Clonal Selec-
tion Algorithm (ITCSA) as an intermediate information strategy between disseminating
RTCSA and localising MTCSA. Toward this end, the study had the following goals:
1. Investigate carrying capacity as a recirculation-independent mechanism for promot-

ing the localisation of acquired information.
2. Investigate the capability of ITCSA to counter the disruptive effects of recirculation.
Method
Algorithms The study considers three algorithms, the Minimal Tissue Clonal Selection
Algorithm, and the Recirculation Tissue Clonal Selection Algorithm, and the Inflammation
Tissue Clonal Selection Algorithm. The MTCSA and RTCSA algorithm was configured as
was defined for the RTCSA Empirical Study in Section 5.5.3 with Ntissues = 10, Ncells =
50, Nselection = 1, Nclones = 5 for MTCSA and RTCSA, and Nmigration = 5 (10% of
each T ) fixed for the RTCSA, referred to as RTCSA-S (small migration). The ITCSA
is a specialisation of the TCSA that increases the carrying capacity of tissues exposed to
infections to Nmaxcells defined in Algorithm 5.12, and progressively decreases the tissue to
Ncells each epoch defined in Algorithm 5.13. Two variants of ITCSA were investigated, one
based on the MTCSA with no T -T interaction referred to as ITCA-N (no migration), and
one based on RTCSA with recirculation (Nmigration = 5) referred to as ITCSA-S (small
migration). Both ITCSA used the same carrying capacity increase of Nmaxcells = 60 (20%
increase over Ncells = 50).
Problems The same Infection Colour Space Problem and Tissue Exposure Regimes
were used as was defined for the RTCSA empirical study in Section 5.5.3.
Experiment The same experimental setup was used as was defined for the RTCSA
Results
202
Figure 5.28: Box-and-whisker plot of Host Diversity (HD) across all TER for the ITCSA
study.
Analysis
This section analyses the summarised results in Table A.4 in the context of the aims
outlined for this empirical study.
Localisation Trends This section considers the localisation effects of increasing car-
rying capacity in ITCSA with out recirculation as compared to MTCSA. From a system
perspective, ITCSA-N resulted in large increases in diversity compared to MTCSA, an
effect expected by a strong localising method (Section 5.4.4). Unexpected was the con-
current large increases in host error across all exposure regimes. This suggests that the
additional resources were exploited by specialising the system, although at the cost of re-
sources that were required for improving the response. From a component level, ITCSA-N
resulted in a general decrease in the average tissue diversity, and a general increase in the
average tissue error. This increased tissue internal homogeneity is expected given that
increased capacity is exploited by local clonal progeny, although the impaired average
tissue capability across the TER’s correlates with the systems increase in response error.
This suggests that an artefact of the mechanism may be responsible for disrupting the
acquisition of information. The observations may be generalised into the following trends:
1. Inflammation facilitates component and system level increases in diversity as ex-

pected by a strong localising effect.
203
Figure 5.29: Box-and-whisker plot of Host Error (HE) across all TER for the ITCSA
study.
the ITCSA study.
204
ITCSA study.
2. Inflammation’s promise of localisation is broken by the increase in system and com-

ponent level error, likely an artefact of the specific implementation.
Counter-Disruptive Trends This section considers the capability of ITCSA (specifi-

cally ITCSA with recirculation: ITCSA-S) to counter the disruptive effects of recircula-
tion to the local exposure regimes such as ATER, OTER and PTER. From a system level
ITCSA-S resulted in an increase in diversity compared to RTCSA-S, although less a dra-
matic increase than between ITCSA-N and MTCSA given RTCSA also decrease diversity
compared with MTCSA. The diversity increase aligns with the expectation of increased
localisation, and with the observation in the previous comparison. ITCSA-S resulted in
an increase in error on ATER and STER and a decrease in system error for PTER and
RTER over RTCSA-S. From a component perspective inflammation resulted in a general
decrease in the average tissue diversity, with an increase observed on ATER. Component
error was increased for OTER and STER, although decreased with inflammation on ATER
and RTER.
The important finding and confirmation that inflammation can counter the disruptive
effects of recirculation is the decrease in HE on PTER and RTER. This finding becomes
clear when one considers the global effects of varying carrying capacity with each exposure.
In the case of regular system-wide exposure regimes (these are regimes under which each
tissue is exposed to some information each host exposure) such as STER and ATER,
the carrying capacity for each tissue is always increased to Nmaxcells . For those regimes
where regular system-wide exposure is not assured, such as OTER, PTER and RTER,
the irregularity promotes variability in the carrying capacity. Given that system error
was not decreased for STER and ATER it suggests that the holistic increase in repertoire
205
size does decrease the disruption of recirculation, rather the inflammation provided an
effect that increased error on ATER and STER. In the case of PTER and RTER system
error was reduced over recirculation suggesting that the variable repertoire sizes across
the host and the inflammation mechanisms method for exploiting the extra size (with
progeny of selected cells) can only be exploited under these circumstances. Interestingly
inflammation provided some increase and decrease in average per-component error that
did not align with the regular and irregular system-wide exposure correlation, and in
particular improved average component capability for both the archetype regular and
irregular exposure regimes (ATER and RTER). These observations are reduced to the
following general trends:
1. Inflammation generally counters the disruptive effects of recirculation with regard

to system diversity, although system-level benefits are observed only under those
exposure regimes that have irregular exposure regimes with regard to the number of
tissues exposed to information per host exposure (epoch).
2. Inflammation under consistent system-wide exposure regimes results in a holistic

increase in repertoire size that under the ITCSA inflammation mechanism decreases
system capability with regard to host error.
Conclusions
This section summarises the findings of the empirical study into the Inflammation Tissue
1. Inflammation
(a) Inflammation promotes localisation that is apparent with and without recircu-
lation, although it generally disrupts system capability (error).
(b) The increase in carrying capacity mechanisms is only effectively exploited as a
counter measure to the disruptive effects of recirculation under exposure regimes
that promote disparity in the resulting effect (non-uniform tissue exposure).
The results suggest at the potential of the emergent effect although raise questions
regarding the specifics of the chosen mechanism. Specifically, the method by which in-
creased carrying capacity is exploited (what cells fill the space), and the method by which
carrying capacity is decreased (what cells are selected and deleted from the repertoire). It
is likely that viable cell lineages are stopped under random cell deletion resulting in the
general increase in system error rather than an expected neutrality and potential decreases
provided by increase specialisation. An alternative mechanism may be increased progeny
tournaments for progeny during integration into the repertoire after creation, providing a
natural mechanism for competition for increasingly limited resources. There is much room
for elaboration on the primitive realisation of the inflammation effect, not limited to the
discussed manipulation of the recirculation effect (increased tissue inflow and decreased
206
tissue outflow), as well as the clear integration of preferential residence retain specialised
resources at the point of use (homing).
5.8 Chapter Summary

It is important to note that tissue models of the immune system have been considered
before by Twycross and Aickelin [400] concerned with the innate immune system, and by
Bentley, et al. [36] concerned with tissue growth models. The principle difference and
contribution of the presented approach is the integrated and subsumed relationship with
the existing clonal selection paradigm. The Tissue Clonal Selection Paradigm (TCSP) was
defined as the investigation of the cellular clonal selection paradigm as constrained by the
concerns of (1) multiple discrete repertoires of cells called Tissues and their interactions,
and (2) the concerns of regimes of discrete repertoire exposures with information called
Tissue Exposure Regimes (TER). The paradigm exploits the metaphor of a holistic immune
system called a Host, arranged into a lymphatic system of lymphoid tissues that may inter-
communicate using the differential trafficking of lymphocyte cells called a Tissue Clonal
Selection Algorithm (TCSA). The principle information processing interest of the paradigm
as gleamed from the metaphor are the information management strategies that may be
employed to address the known and unknown regularities and irregularities in an antigenic
Habitat of information called an Infection Antigenic Exposure Problem (ITEP) toward
continuous acquisition of information and the anticipation of needs of acquired information.
An abstract foundation for the paradigm was defined in terms of model components,
discrete exposures with spatial-temporal regularities, and a series of architectures that
may be exploited for arranging the components of tissue models.
5.8.2 Principles and Findings

The following summarise the important principles and findings from the definition and
investigations into the Tissue Clonal Selection Paradigm:
General Principles
1. Natural organisation based on exposure: Given sufficient time and resources, infor-
mation is naturally organised in a fixed structured system based on regularity and
consistency of its exposure. For example, system-wide exposure naturally promotes
system-wide anticipation and point-wise exposure promotes point-wise anticipation.
2. Problem is in the exposure regime: Resources are not always sufficient, therefore
information management strategies promote more efficient and effective organisation
of information in the face of unknown regularities and irregularities in information
exposure and anticipation.
3. Strategies based on tissue-level immunology: The immune system inspires strate-

gies for this problem as the survival of the host relies on the effective acquisition
207
of information and efficient intra-host application of acquired information to fight
infections.
Specific Findings
1. Recirculation: Scope of system exposure (number of tissues) is best addressed with a

match in scope of information management strategy at a system level, although the
general trend at the component level is invariant to the strategy used. For example
the specialisation of response is suitable for regular constrained exposure, whereas
the unbiased dissemination is suitable for irregular and/or unconstrained exposure,
although such strategies do not counter the general natural component-wise trends
in organisation.
2. Homing: The dissemination of information under a constrained exposure regime

when such dissemination is unneeded provides a disruptive effect that may be coun-
tered by empowering cells to exhibit a preferential residence, whist not impairing
dissemination under those regimes that benefit from the effect.
3. Inflammation: The increase in carrying capacity toward decreasing the competition

for local resources at those points of the system that are exposed provides a second
mechanism for decreasing the disruptive effects of recirculation under those exposure
regimes where disparity of the effect is promoted, at the cost of providing a negative
effect under those regimes that do not.
5.8.3 Integration
The Cellular Clonal Selection Paradigm defined and investigated in Chapter 4 was con-
cerned with the adaptation as a method for information acquisition, and the how that
method is influenced by degenerate information and inter-cell interactions. The Tissue
Clonal Selection Paradigm defined and investigated in this chapter took information ac-
quisition via adaptation for granted, using it as a primitive component in a Tissue Clonal
Selection Algorithm raising the level of abstraction regarding information acquisition and
adaptation from cells to tissues. The interest in this chapter was how different decen-
tralised information management strategies effected the organisation acquisition and use
of information under different information exposure regimes. The abstract tissue paradigm
is grounded in Chapter 8 in the context of two general problem domains, highlighting the
potential benefits the tissue algorithms provide over the cellular clonal selection algo-
rithms. The subsumption theme of raising the level of abstraction in information pro-
cessing is continued in Chapter 6 in which a Tissue Clonal Selection Algorithm is made a
primitive component in a broader Host Clonal Selection Paradigm that is concerned with
information management strategies in a population of hosts.
208
Chapter 6
Host Clonal Selection

This chapter extends the Tissue Paradigm, by raising the level of abstraction for clonal
selection to that of a distributed population of self-governed immune systems interacting
with each other in an antigenic environment. Section 6.2 considers a holistic perspective
of the mammalian immune system and reviews the ways in which populations of immune
systems may interact with each other both directly and indirectly. Specifically, host im-
mune system interactions are considered in the context of the passive and active forms of
immunisation in a population and across generations, as well as the evolutionary concerns
for the basis of immunological function. Section 6.3 abstracts the background biology of
host immunisation and immune system evolution and presents the Host Clonal Selection
Paradigm. The paradigm is presented in terms of shared and evolved immunity interaction
models, controlled and uncontrolled antigenic exposures, and population and generational
host architectures. Collectively these concerns provide a foundation for the series of Host
Clonal Selection Algorithms proposed and investigated in this chapter. Sections 6.5 and
6.6 investigate the inter-population sharing of the trigger and product of acquired im-
munity respectively in a transmission and shared immunity algorithms. Sections 6.7 and
6.8 investigate the the inter-generational sharing of the product and the evolution of the
mechanisms of acquired immunity respectively. The information processing and manage-
ment strategies of the host paradigm demonstrated similar localisation and dissemination
qualities as were demonstrated in the tissue paradigm with the additional level of control
over interactions and decoupling of cellular repertoires promoted by the perspective of
clonal selection.
6.2 Populations of Interacting Immune Systems

Immune systems do not exist in isolation, in terms of the various organs and interrelated
subsystems within the host organism, or in terms of the populations to which the host
organisms belong. In raising the level of abstraction to that of populations of immune
systems, one may be interested in the manners in which immune systems interact. This
transition may be considered as moving from thinking of the Darwinian struggle of lym-
209
phocytes within one host holistically providing a defence to the organism, to populations
of these systems holistically providing a defence to a population (or a species). This sec-
tion considers populations of immune systems, specifically with regard to evolutionary
concerns and natural and artificial immunisation, and ways in which these systems may
interact. This general review is multi-disciplinary crossing such fields of study as popula-
tion and evolutionary ecology, microbiology, epidemiology, and immunology. The intent
of this section is to highlight some systems and processes that underlie the interactions
between holistic immune systems to provide the basis for the inspiration of distributed
computational Artificial Immune System models.
6.2.1 Immunisation
Generally, immunisation implies that a host has been exposed to an antigen (immuno-
gen) inducing an immune response that has educated the system with respect to more
effectively and efficiently detecting and neutralising the antigen in the future. When a
pathogen enters the host and elicits a response (and the host survives), the host gains
a level of immunity to the pathogen. This process can be induced artificially through
inoculation (controlled exposure to the pathogen), and is the basis of vaccination. Inter-
estingly, immunity is not limited to this reactive method, and evolution has designed a
temporary and passive form of immunity that is employed by mothers to protect children
from pathogen before and after birth. This section reviews the various natural and ar-
tificial forms of acquiring immunity in terms of those mechanisms that elicit an immune
response, and those that do not. Table 6.1 provides a summary of examples.
Natural Artificial
Active Infection Vaccination, Inoculation
Passive Maternal Immunity Injection of antibodies
Table 6.1: Summary of the different types of immunity.
A developmental oddity of the acquired immune system is that the mass of the system
is at its peak in early childhood, and decreases with time until early adulthood [289]. This
raises questions as to the usefulness of decreasing the resources of the defence system with
age. An intuitive consideration would suggest that the size of the system would increase
with the increase of acquired knowledge over the lifetime of the host. An answer to this
question provides insight into acquired immunity as a strategy. Children are more vul-
nerable to infection than adults (children under the age of 5 suffer 8 to 24 times more
infection than other age groups). Each pathogen encountered by an infant is novel and
must be responded to anew. As a strategy, acquired immunity learns the hosts’ pathogenic
environment early and quickly. It provides an attempt to increase the hosts’ early and
prolonged survival within its environment. During the early stages of life the system max-
imises diversity, maximises cell proliferation potential, and maximises the rate of memory
cell creation. Cell numbers peak in the first six months of life then steadily decline. Fur-
ther, the developing infant immune system is buffered to the antigenic environment by the
210
temporally inherited maternal immunity (natural passive) gained by the mother over her
lifetime. This form of immunity is discussed in the following section.
Passive Immunity
Passive immunity is the transfer of active humoral immunity in the form of ready-made
antibodies to a host organism. Antibodies are relatively short-lived molecules that have no
reproductive capability, thus the defensive capability provided by them is fast, effective,
and temporary. A common form of natural passive immunity is maternal immunity (for
review see [42, 136, 196, 54]). Antibodies are provided by the mother to the foetus across
the placenta and in breast-fed milk after birth (mucosal immunity). Vaccination of the
mother during pregnancy may directly benefit the child as the antibodies produced by
the mother will be passed to the foetus [172]. During breastfeeding, lymphocytes in the
mother move to lymph nodes near the breast and release large amounts of antibodies into
the breast milk. The antibodies enter the gut of the infant and provide protection against
bacteria [363], and assists in oral tolerance (avoiding allergic reactions to food). The
antibodies provide an initial and powerful boost in defence to the infant, which endures
for a prolonged period after breastfeeding ceases. Breastfeeding stimulates the developing
immune system, improving later response to vaccination, although initially suppresses the
effector response of the infant immune system causing the failure of effective (artificial)
vaccination during this period. It decreases infant mortality rate and provides benefits for
the metabolism and disease resistance later in life [308].
Passive immunity may be provided artificially through the transfer of blood or serum,
and is applied when there is insufficient time for the host to develop an immune response.
Some examples include viral and bacterial infections and poisons such as snake bites
[245]. Cell-mediated immunity may also be transferred in this manner (effector systems
for detecting and neutralising infected cells rather than foreign molecules) and is called
adoptive immunisation. This is the transfer of recirculating lymphocytes rather than
antibodies, although is difficult as such cells may be identified as foreign and attacked by
the host immune system, thus the process is rarely used in humans. Monoclonal antibodies
(mAb) are an artificial passive immunity where large quantities of identical antibodies are
synthesised from identical cloned lymphocytes [407]. Unlike polyclonal antibodies that
vary in their specificity for an antigen, monoclonal antibodies all have the same specificity,
and thus may provide a strong and directed defence. In theory it is possible to synthesise
antibodies to recognise any substance. They are primarily used in biomedical research
for detection and purification tasks, in the diagnosis of disease, and in the treatment of
infections and cancer as delivery mechanisms [414]. The process for producing monoclonal
antibodies was first described in 1975 [260], and later refined for human use [336].
Active Immunity
Active immunity involves the stimulation of the system by an antigen and the develop-
ment of an immune response. Natural active immunity is the normal functioning of the
acquired immune system for neutralising pathogens. Artificial active immunity involves
211
intentionally inducing the immune response by introducing an immunogen in a vaccine.
A vaccine is typically administered before the patient contracts the disease, although it
may be administered after, assuming the immune response to the immunogen is faster
than the response to the infection. There are many different types of vaccines including
inactivated (dead) pathogens, live attenuated (low virulence) pathogens, toxoids (inactive
toxic compounds from pathogens), and subunit or pieces of pathogens. Vaccination does
not prevent the host from contracting the disease, rather like natural active immunity it
provides an improved response to the disease making it harder for it to spread throughout
the population. Vaccination is one of the greatest achievements in medicine and pub-
lic health [344]. The ethics of compulsory vaccination are still debated throughout the
world, although the benefits of immunity to disease and disease eradication are generally
considered to outweigh the costs of a minority of recipients contracting the disease the
vaccine was designed to prevent. An aspect of population vaccination is an effect called
herd immunity [140, 23, 22]. This is where the majority of the population (perhaps more
than 90%) is vaccinated for a disease, which provides protective coverage for the entire
population. An impact of this interesting effect is that vaccination schemes may not be
concerned with immunising the remainder of the un-vaccinated population, rather the
immunisation of enough of the population to reduce the chance (to a desired level) of a
member of the population contracting the disease.
6.2.2 Evolution
The acquired immune system of mammals (such as humans, mice, rats, and cattle) is the
most studied immune system. A consideration of the system in the context of the ‘tree of
life’ reveals that this it belongs to a very small minority of taxa: the jawed vertebrates,
raising the question as to why this system evolved when all other plants and animals
survive without it. Evolution has designed a complex, specialised, and multi-layered de-
fence in the immune system [105], although the system is not perfect: “. . . vertebrates are
exquisitely adapted [. . . ] but not perfectly; autoimmunity and cancers present the costs
and failures of immunity” [337] (page 33). The very fact that mammals are vulnerable
to disease highlights that the immune system is a sub-optimal adaptation [303]. Interest-
ingly, some suggest the acquired immune system may be an evolutionary misstep given
all the positive and negative feedback signals (checks and balances) needed to regulate an
immune response [204]. This section provides a brief review of the innate and acquired
immune systems from an evolutionary perspective.
The Innate and Acquired Immune Systems
The innate immune system is a collection of defence mechanisms, tissues, and specialised
processes that defend the host from its pathogenic environment. The innate defence is
inherited and defined in the genome, thus the specificity for the pathogens it may detect
is ‘hard-wired’ throughout the hosts lifetime. The trade-off for this rigidity is that the
defence it provides is fast, practically instantaneous. In the acquired immune system,
evolution has designed a somatic learning system that may recognise any biochemical
212
molecule, although this tool of defence may also destroy the host which it is intended
to protect. Acquired immunity is a personalised defence, learned within the context of
the hosts’ specific pathogenic environment. It is slow acting, requiring an aggregation of
antigen encounters generally over about a week before it is ready to fight the pathogen (fast
compared to the innate system which specialised over generational time). As a strategy,
the system is anticipatory, assuming that future pathogenic encounters will be the same
or similar to past pathogenic encounters. Table 6.2 contrasts some of the properties of the
acquired and innate immune systems.
Receptor’s Features Immunity Systems

Innate Adaptive
Distribution in cell populations uniform clonal
Specificity low high
Affinity 103 − 104 106 − 1011
Diversity (the number of variants) 102 − 103 109 − 1011
Variability absent high
Ligand origination foreign (non-self) foreign and self
Table 6.2: Summary of the differences between the innate and acquired immune systems,
taken from [257] (page 565).
The acquired immune system is exceedingly complex, such that approximately 5% of

the genome of humans is proposed to be for molecules used in the system. This antibody
immunity is proposed to have evolved gradually over a long period of time from cells and
tissues that previously performed other functions [256]. Klimovich considers some of the
problems related to the evolution of the acquired immune system highlighting: (1) the sys-
tem was formed in the presence not only of pathogen but also coexisting foreign organisms
living in symbiosis, (2) many kinds of activity in the system are not protective in nature,
instead serving alternate functions, and (3) the system performs the learning function be-
fore applying the protective mechanism (so called pre-committed) [257]. Hedrick suggests
that perhaps long-lived complex jawed vertebrates need an acquired immune system, high-
lighting that the acquired immune system may be an approximate solution to the problem
of parasitism, providing a small advantage in reproductive success that invertebrates could
not afford the cost of resources to evolve, maintain, and exploit [204].
6.2.3 Summary
The evolution of immunity is the interaction of immune systems both with their environ-
ment, which may be adversarial in the case of parasites, and with each other in terms
of reproduction and natural selection. Sexual recombination of the genetics of immu-
nity provides interactions of immune systems in terms of new arrangements of existent
traits. The pressure of natural selection results in the implicit competition between het-
erogeneous populations of immune systems in a spatial-temporal pathogenic environment.
Immunisation provides a second example of the interaction between immune systems.
Evolution designed the mammalian acquired immune system to be augmented in infancy
by the lifetime immunological knowledge of the maternal immune system in a temporary
213
Lamarckian-style inheritance. This natural passive immunity provides a robust replace-
ment defence for the developing system. Additionally this process may be reproduced
artificially later in life by the removal of antibodies from one host and then injected into
another to provide a short-lasting immediate defence to a pathogen.
6.3 Abstract Host Paradigm

This section abstracts and elaborates on a host-centric adaptive strategy and proposes
a Host Clonal Selection Paradigm that takes the Tissue Paradigm and implicitly the
Cellular Paradigm for granted, and focuses on the inter-host interaction and information
sharing provided by bottom-up host-population and host-generational immunisation and
evolutionary mechanisms. Conceptually, a host represents a rise in the level of abstraction
from a tissue (conventional clonal selection algorithm or repertoire) to a collection of such
tissues working together to provide holistic immunological function. Therefore a Host
Clonal Selection Algorithm is concerned with a system comprised of a number of Tissue
Clonal Selection Algorithms (Chapter 5) that in effect are the components or units of
adaptation (hosts) within a host algorithm.
6.3.1 Interaction Models

The host paradigm is primarily concerned with the interaction of whole immune systems
where the considerations of what happens within a single host is delegated to the Tissue
Paradigm, and in turn the Cellular Paradigm for the concerns within a single tissue reper-
toire of cells. This section considered the two primary means in which immune systems can
interact and more importantly effect each other: explicitly and implicitly. Explicit inter-
action between host immune systems is defined as Effector Sharing where the information
acquired by an immune system (its effectors) are shared with other immune systems ei-
ther by passive infusion or active response elicitation. Implicit interaction between hosts
involves evolutionary processes over generational time that effect the basis and processes
by which progeny immune systems acquire information called Evolved Immunity.
Effector Sharing
Immunity is acquired and refined over the lifetime of the host given antigen exposures
and resultant clonal selection and affinity maturation. The effectors produced in this
process (antibodies, memory and recirculating lymphocytes) may be shared with other
immune systems. In this context effector is a generalised term that refers to some discrete
information acquired by an immune system such as the product of a clonal selection and
expansion event. This section describes two mechanisms for sharing effectors between
systems in what is referred to as passive immunity: Vertical Sharing and Horizontal
Sharing.
Vertical Sharing Vertical sharing is inspired by natural passive immunity such as the
sharing of antibodies between mother and foetus across the placenta (maternal immunity),
214
and to the infant after birth through breastfeeding (mucosal immunity). Vertical sharing
of immune effectors is a mechanism of inheriting acquired immunity between generations.
A sample of the effectors are removed from a parent system and inserted into a child
system. The shared effectors, such as all cellular components have a finite lifetime and
may perish before being employed by the receiving system. A natural version of this
mechanism is a parent-child relationship, but this is not a requirement. The mechanism
does require (1) a generational population structure, (2) the selection of a donor from the
previous generation, (3) the selection of a receiver from the next generation, and (4) a
sampling method for the donor (effector selection mechanism).
Horizontal Sharing Horizontal sharing of immunity effectors is inspired by artificial

passive immunity where effectors are synthesised and injected into hosts, or removed from
one host and injected into another. This method of effector sharing provides a way of
transferring learned immunity within a population of immune systems. The mechanism
does not require a generational population structure, although, like vertical sharing it does
require (1) a donor of effectors (within the population or synthesised externally), (2) a
receiver system within the population, and (3) a sampling (selection) method for effector
cellular components from the donor.
Evolved Immunity
Another way of sharing immunity is through genetic evolution involving (1) the assignment
of immune system fitness, (2) selection of parent immune systems, (3) the recombination of
parent genetic material to create children, and (4) the mutation of child genetic material.
Those immune systems that demonstrate their relative utility (relative to other immune
systems with the same generational population) may proliferate, and those that do not
demonstrate utility, perish. The evolution of immune systems requires that the expressed
phenotype (thing being evolved such as immune effectors or response strategy) be tran-
scribed from a genome (representation) on which the evolutionary processes of selection,
recombination, and mutation may operate. It requires a generational mechanism for im-
plicitly sharing effectors (seed and process for creating effectors) rather than explicitly as
in the case of vertical and horizontal effector sharing. Two evolved immunity mechanisms
are presented as follows: (1) Evolution of an Innate Immunity, and (2) Evolution of an
Acquired Immunity.
Innate Immunity Innate immunity may be thought of as hard-wired defence mecha-

nisms that are learned in response to pathogens within a species over generational time. A
simple example is a static recognition and response rule system for a host, which although
remains fixed for the lifetime of a single host, evolves to protect the whole species. In this
example, the rule system is fast acting on the lifetime scale, although slow to change on
the same scale implying perhaps the increased expendability of individuals for the greater
good of the species.
215
Acquired Immunity The acquired immune system, although is a somatic learning
system rapidly adapting to the hosts pathogenic environment, may also acquire more
generalised species level traits via evolution. Evolution may be used to evolve the acquired
immune system by biasing the creation of naı̈ve lymphocytes to specific sub-areas of the
shape-space. This bias may accelerate the acquisition of immunity by individuals, requiring
less work by the somatic learning system.

Discrete Host Exposures
Section 5.3.2 extended the Antigenic Exposure Paradigm of Section 4.2.2 to include the
notion of multiple points of exposure on a Tissue Clonal Selection Algorithm (each reper-
toire). The attributes of discrete exposure were bundled into a concept called the Tissue
Exposure Pattern. This concept was specialised to incorporate spatial and temporal regu-
larities of exposure location and duration in order to assess the information acquisition and
dissemination properties of tissue models called an Antigenic Infection. The aggregation
of multiple antigenic infections on a tissue was referred to as situating the tissue model
(host) in an Antigenic Habitat. Like Tissue Clonal Selection Algorithms, Host Algorithms
also provide multiple points of exposure in an antigenic environment and therefore have
similar concerns. Figure 6.1 provides a conceptualisation of a spatial population of discrete
hosts, each situated in their own antigenic habitat.
Figure 6.1: Host Clonal Selection Model and the situated spatial properties that affect
antigenic exposure, where the arrows represent interaction between the discrete compo-
nents (hosts) in the system.
There are two distinct and important differences between the Tissue and Host paradigms
with regard to antigenic exposures: (1) exposures of antigen to hosts may be both con-
trolled (elicited by other hosts) and uncontrolled (environmental) whereas tissue models
have no control over such exposures, and (2) the level of abstraction regarding exposures
is raised from that of an habitat of infections for a single host, to multiple and potentially
different (asymmetric) habitats across multiple hosts called an Antigenic Environment.
216
Controlled and Uncontrolled Exposures
The acquired immune system learns through exposure, detection, and response to antigens,
thus the antigenic environment defines what is learned and when. The immunity that a
system receives through recognising and responding to antigens is called active immunity
and may be induced naturally (uncontrolled) and artificially (controlled).
Uncontrolled Exposure Uncontrolled antigen exposure is a natural way for a system

to achieve active response-based immunity. This form of antigen exposure is an elabora-
tion of the previously defined standard model. An aspect of the distinctiveness of acquired
immune systems in a population is the spatial and temporal variations of antigen expo-
sures. These varied exposure patterns in conjunction with the varied naı̈ve repertoire of
each individual result in distinct initial and learned specificities to antigens. For a given
antigen, exposure may be described as a function over space and time that may be skewed
for many reasons such as antigen virulence, host density, and other pathogen-behaviour
and spatial-population effects.
Controlled Exposure Control over antigen exposure provides control over immunity
of individual systems and virtual immunity over the entire population through effects like
‘herd immunity’. Artificial exposure of systems to antigen provides control of (1) which
systems are exposed to (2) what antigen (3) when. A vaccination exposure strategy may
involve administering a vaccine to a large number of systems quickly. An inoculation
strategy may involve exposing a select few systems to the antigen in a short amount of
time. Control of exposures may also extend to preventative steps such as environment
sterilisation or individual system isolation from the antigenic environment.
Antigenic Habitats and Environments
The Tissue Exposure Regime of the Tissue Paradigm may be extended to a Host Exposure
Regime (HER) that defines which hosts are exposed to antigen and when they are exposed.
This means that the system is less concerned with how well a given host can specialise a
response and internally disseminate acquired information, and more concerned with how
that information is acquired and disseminated between hosts in the population. Unlike
tissues, hosts are not restricted to fixed network topologies like the ring network structure
of the Lymphocyte Recirculation Algorithm, therefore asymmetries in the Host Exposure
Regime and the host interaction patterns define the capability and speed of information
dissemination throughout the population. Given the shift in focus from the internal pro-
cess of information acquisition within a host to the internal information acquisition of a
population of hosts, the antigenic environment may be conceptualised as a Pathogenic
Environment consisting primarily of exogenous antigen. This does not mean hosts may
not respond and communicate information regarding endogenous antigen, rather it sug-
gests that the focus of the Host Paradigm is hosts and their relationships with external
antigen that may come the environment and other hosts. In addition to the Host Expo-
sure Regime, the rise in the level of abstraction in the Host Paradigm facilitates broader
217
concepts regarding antigen as self-contained entities, in particular (1) the co-evolution of
pathogen with a population of hosts, and (2) the contraction of a pathogen by a host and
transmission to neighbouring hosts.
Evolvable Pathogen A pathogen may explicitly evolve or co-evolve with the population
of immune systems. Evolution may refine the strategy of the pathogen affecting configu-
ration such as virulence and transmissibility. An individual acquired immune system may
be open to multiple infections by the same pathogen during its lifetime if the rate of evo-
lution is sufficiently rapid. Such a pathogen may require tracking and ultimately evolving
countermeasures by the host population, perhaps facilitating a host-parasite interaction.
Transmittable Pathogen After exposure to a host, a pathogen may jump from host-to-
host using a number of mechanisms (vectors). Further, transmissibility may be influenced
by any number of epidemic-based host models such as density-dependence, spatial models,
and genetic-similarity models.
6.3.3 Host Architectures

The abstraction of immune system interaction models of immunisation and evolution sug-
gests two distinct Host architectures as follows: (1) Populations of Hosts that focus on the
natural and artificial elicitation of immune response and transfer of acquired immunity,
and (2) Generations of Populations of Hosts that focus on hosts with a finite life-cycle
that include inter-generational elicitation and transfer of immunity and the evolutionary
change of the properties of lifetime innate and acquired immunity. This section outlines
these two distinct host architectures.
Population Host Model
An acquired immune system does not exist in isolation, rather it is a collection of organs
and tissues that collectively represent an integrated part of a host organism, which in turn
belongs to a population of such organisms. One may consider the hosts of a population
interacting with each other, more specifically interacting with regard to their immune sys-
tems. The class of model in which a population of clonal selection-based immune systems
interact with each other is referred to as a Population Host Model. The focus of this model
is the host as the unit of adaptation, each with a distinct starting point, and likely distinct
antigenic environment including interactions with the environment and other hosts. This
diversity provides robustness in the population with regard to information acquisition and
dissemination, abstracting concerns of population survivability in an unknown environ-
ment. A population of non-interacting hosts facilitate the development of multiple varied
perspectives of an antigenic environment that remain isolated. Integration of this informa-
tion without host interaction mechanisms would require augmentation from a higher level
of abstraction (the population level). The isolated perspectives property of the population
model is made more apparent in an asymmetrical pathogenic environment. In such an
environment, not only does the heterogeneous population acquire information in different
218
ways and rates, but the information content of the environment to which they are exposed
also varies. Therefore, although the population may collectively represent a map of the
environment, there is no way for individual systems to share acquired information. This
concern is further highlighted if we consider that the asymmetrical pathogenic exposure
regimes for the systems in the population vary over time. In this case, the information
acquired by one system is directly useful to other systems, although cannot be exploited.
A population of non-interacting hosts provides a foundation on which to incorporate intra-
population (inter-host) interactions. Intra-population re-transmission of captured antigen,
and intra-population effector sharing provide two examples that motivate the investigation
of decentralised intra-population interactions.
Generational Model
Host organisms in a population reproduce such that a generational structure may be de-
fined between hosts where progenitor-hosts produce progeny-hosts. In this mechanism,
the population of immune systems (the present generation) are responsible for creating a
new set of immune systems (the next generation), which replace members of the present
generation. This principle defines the Generational Host Model that provides an extension
to the proposed Host Population Model. Hosts may use an asexual reproduction mecha-
nism where one host contributes a variation of itself to the next generation. Alternatively,
hosts may collaborate in their contributions to successive generations (sexual reproduc-
tion). Further, the contributions of various hosts to successive generations may not be
homogeneous, such that more successful hosts (by some measure of success) may contribute
more than less successful hosts. The principle components of the generational approach
are (1) the reproductive scheme, and (2) the mechanism that triggers the generational
change.
• Reproductive Scheme: The mechanism by which one generation contributes to and

prepares the successive generation. Examples include sexual and asexual repro-
duction where hosts may or may not cooperate in their contributions to the next
generation. Further, the contributions of any one host to the next generation may or
may not be homogeneous with respect to the other hosts of the present generation.
• Generational Trigger : The condition, that once satisfied requires the application of
the reproductive scheme to the present generation to create the next generation of
immune systems. An example is a time trigger in which immune systems are given
a representative exposure to the pathogenic environment.
To be clear, the host generational approach is the same in principle to classical gener-
ational clonal selection algorithms in that one generation represents one iteration of the
systems execution. The important difference is that one generation of the host generation
system involves multiple generations of the classical clonal selection algorithm (cellular)
within each host (hosts’ tissues).
The premise of the generational mechanism is that a given instance of a host immune
system may acquire different information in the same antigenic environment over successive
219
trials (generations). A generational model that does not allow one generation of hosts to
convey information to progeny, rather it provides a clean slate for each system to repeat a
trial. Ultimately such a simplified generational model ‘resets’ each host to initial conditions
at the beginning of each new generation. The restart principle gives a population of
immune systems, each with their unique and ongoing perspective of the environment,
successive chances of learning a perspective of the environment. The inability of the
systems to share acquired information between generations, means that each generation
must re-learn (start from scratch) those things learned in the previous generation. If a
generational-accumulative model is desired, it must be implemented as a mechanism at a
higher scale than the generational model. Such a mechanism is responsible for explicitly
integrating information across generations. The hope of the generational model is that each
fresh-start and subsequent run (trial with the environment) results in a varied perspective
and thus varied acquired information perhaps at the individual system-level, and more
importantly at the population-level. The concern of a non-interacting generational model
is that operating with a fresh-start is likely too computationally intensive (inefficient) with
regard to the information acquired and explicit integrated. More importantly, there is little
guarantee that the ‘varied perspective’ principle will hold for a given environment. The
model may be extended by proposing two different vertical sharing mechanisms by which
populations can share information between generations of the a algorithm: explicitly by
the inter-generational sharing of effectors, and implicitly by the evolution of characteristics
that influence the response to antigenic exposures. The intent of sharing information
between generations is that subsequent generations can learn from the success and failures
of the previous generation, and the generations that came before it. Such mechanisms
allow the generational-population (species) to implicitly integrate information about the
environment over generational time (rather than explicitly by an external mechanism).
Further, such mechanisms do not restrict subsequent generations to the limitations of the
past, rather integrating generational knowledge into the ‘fresh-start’ systems facilitating
the desired trait of the non-interacting generational model of acquiring a varied perspective
(if such a variation in possible).
6.4 Realised Host Paradigm

A population of hosts each with a distinct immune system, exist within an antigenic envi-
ronment where interactions are mediated by an unknown exposure regime. This provides
an effective metaphor for considering the problem of the acquisition and application of
information using a decentralised strategy with distributed information processors. As
was the case with the Tissue Paradigm, information management strategies are concerned
with the effective localisation and dissemination of information, taking the internal acqui-
sition of such information for granted. From the perspective of a population of hosts, the
specifics of information acquisition within each host are taken for granted, relegated to the
investigatory concerns of the Tissue Paradigm. Host immunisation methods provide a rich
metaphor for considering host-host interactions in a population toward eliciting immune
response and sharing the product of the immune response in both a static population
220
and generational population structure. These interactions, in addition to implicit host
interactions in an evolutionary process, motivate the realisation of distributed informa-
tion management strategies in the host clonal selection paradigm. This section provides
a realisation of the concerns of the abstract Host Clonal Selection Paradigm presented
in Section 6.3. This includes a general problem definition that captures the concerns for
investigating systems at the host-level of abstraction, as well as its specialisation in the
colour space domain. A general Host Clonal Selection Algorithm is proposed providing a
basis for population-based host architecture and as well as its extension, the generational
host architecture. Empirical measures are proposed for the investigation of developed
strategies, and general behaviour expectations are proposed motivating confirmatory re-
search.
6.4.1 Antigenic Habitats

This section considers the realisation of a specialisation of the Antigenic Exposure Paradigm
outlined in Section 6.3.2. In particular, this section abstracts the Infection Exposure
Problem considered for the Tissue Paradigm, to a Habitat for the Host Paradigm, and
a specialisation of the problem to Colour Space. In addition, the discrete properties of
tissues exposures are considered in the context of hosts in a population and specialised
Host Exposure Regimes are defined.
Habitat Antigenic Exposure Problem
The Habitat Antigenic Exposure Problem (HAEP) defined in Algorithm 6.1 provides a
general problem definition where a Population P is exposed to an Habitat B. A given Tis-
sue Clonal Selection Algorithm (TCSA) may be considered as a single H in a Population
of Hosts (P = {H1 , H2 , H3 , . . . Hn }). Likewise the Infection Antigenic Exposure Prob-
lem (IAEP) defined in Algorithm 5.1 may be considered a single H in an Environment of
Habitats (E = {B1 , B2 , B3 , . . . Bn }). The parameters of HAEP are the same as IAEP with
the addition of Nhabitats defining the number of B in E. The CreateEnvironment opera-
tion creates a new environment that may be exposed to a P via the Exposure operation
that returns a subset of the population Prs representing the populations ability address the
specific exposure requirements of the environment under the given Host Exposure Regime.
Algorithm 6.1: Habitat Antigenic Exposure Problem (HAEP).

Input: P, Ndeterminants , Nantigen , Ninf ections , Nhabitats
Output: Prs
1 E ← CreateEnvironment(Ndeterminants , Nantigen , Ninf ections , Nhabitats );
2 Prs ←0;
4 Prs ← Exposure(P, E);
5 end
6 return Prs ;
221
Habitat Colour Space Problem
The HAEP may be specialised as a habitat extension of the Infection Colour Space Problem
(ICSP) defined in Section 5.4.1 called the Habitat Colour Space Problem (HCSP). An
instance of the HCSP involves the exposure of a H to a set of Colour Space Pattern Sets
(CSPS) where the goal of the problem is to minimise the average error in the response
Prs . As was the case with IAEP, one may define a minimum configuration of HAEP and
thus HCSP that focuses the concerns on the highest level of abstraction in the problem
(an E of B). The focus is provided by marginalising the configuration at the lower levels.
A Minimal Habitat Antigenic Exposure Problem (MHAEP) is defined where Nhabitats
remains variable, although Ndeterminants , Nantigen , and Ninf ections are fixed at 1. Therefore,
in the case of HCSP, the number of B defines the number of Sets of CSP (I), each of
which is comprised of a single CSP (A) where all three components are treated a single
determinant (D).
Environment
Habitat Habitat ( )
Infection Infection
Antigen Antigen
Determinant Determinant
Figure 6.2: Elaboration on figure 5.18b to depict the Minimal Habitat Antigenic Exposure
Problem (MAHEP)
Host Exposure Regimes
The HAEP may be specialised with a series of Host Exposure Regimes (HER) in the same
manner as the IAEP was specialised with a series of Tissue Exposure Regimes (TER)
in Section 5.4.1. This reuse is facilitated because both the Tissue and Host paradigms
are concerned with strategies for organisation information in the context of unknown
scopes and regularities of information exposure. As such, the series of five TER’s may
be modified for use with the Host Paradigm such that the algorithms remain the same
and the parameters are changed from H and B to P and E. The following Host specific
exposure regimes are defined as AHER, OHER, PHER, RHER, and SHER.
222
6.4.2 Host Clonal Selection
Population Host Clonal Selection Algorithm The Tissue Clonal Selection Algo-
rithm (TCSA) defined in Algorithm 5.7 describes the general interaction between a Habi-
tat B and a Host H. The host paradigm requires a rise in the level of abstraction from
a single H to a Population P of H (P = {H1 , H2 , H3 , . . . , Hn }) that responds to an En-
vironment E of Habitats B. The Host Clonal Selection Algorithm (HCSA) defined in
Algorithm 6.2 describes a general host algorithm that may be specialised based on the
concerns and constraints of the Tissue Paradigm, specifically the HostInteractions op-
eration that provides facility for specialising the interactions between hosts. The focus
of the HCSA is host interactions within a population, therefore to differentiate the ap-
proach from the generational extension (described next), the HCSA may be referred to
as a Population Host Clonal Selection Algorithm (PCSA) to clarify its relationship with
the generational extension. In the tradition of the CCSA and TCSA, one may define a
minimal implementation of HostInteractions in which there are no interactions between
hosts in the population. This Minimal Population Host Clonal Selection Algorithm (MP-
HCSA) provides a baseline for comparison with extensions of the PHCSA that facilitates
intra-population interactions. In addition to no host-interactions, the MP-HCSA codifies
the focus of the Host Paradigm on host interactions, taking intra-host concerns considered
in the Tissue Paradigm for granted. Specifically, each host in the MP-HCSA has a single
tissue (H = {T1 }), allowing each host to be treated as a generic information processing
component in the system.
Algorithm 6.2: Population Host Clonal Selection.

Input: E, Nhosts
Output: P
1 P ←0;
2 for i←0 to Nhosts do
3 Hi ← CreateHost();
4 P ← Hi ;
5 end
7 E.Exposure(P);
8 HostInteractions(P);
9 end
10 return P;
Generational Host Clonal Selection Algorithm In addition to a population-based

host organisation Section 6.3.3 proposed a generational-based host organisation as an
extension to the population-based approach in which a new population of hosts is pe-
riodically created from an existing population. Algorithm 6.3 defines the Generational
Host Clonal Selection Algorithm (GHCSA) as an extension of the PHCSA defined in Al-
gorithm 6.2. The GenerationalChange operation defines the frequency of restarts that
may for example be based on a consistent number of epochs (population exposures). The
223
CreateP opulation operation defines the mechanism for creating a new P from an existing
P that has interacted with the antigenic environment E. A minimal implementation of
the CreateP opulation can be defined in which the creation of the new population of hosts
is independent of the existing population (no interaction). This Minimal Generational
Host Clonal Selection Algorithm (MG-HCSA) provides a baseline for comparison with
extensions of the GHCSA that promotes intra-generational interactions.
Algorithm 6.3: Generational Host Clonal Selection.

Input: E, Nhosts
Output: P
1 P ←0;
4 P ← Hi ;
5 end
7 while ¬GenerationalChange() do
8 E.Exposure(P);
9 HostInteractions(P);
10 end
11 if ¬StopCondition() then
12 P ← CreatePopulation(P);
13 end
14 end
15 return P;

This section defines a series of empirical measures derived from the measures used in
the Tissue Paradigm (Section 5.4.3) that provide instantaneous information regarding a
given HCSA on Colour Space specialisations of the Habitat Antigenic Exposure Problem.
The measures are classified as system that provide general holistic information about the
algorithm’s information state, and component that provide generalised information about
the algorithm’s component information state. As with the Cellular and Tissues Paradigms,
an important consideration with the proposed measures in exploratory experimentation
is not their absolute value, but rather their relative change in value with changes to the
systems being investigated.
System Measures
Population Error The Host Error (HE) defined in Equation 5.1 provides a generalised
measure of the error for a single Host Result Set of tissues Hrs against a given Habitat of
antigenic infection B. This measure may be calculate a Hrs for each H in a P to formulate
an Prs that may be averaged to define a Population error (PE) (see Equation 6.1). PE
provides an indication of instantaneous average error for each Hrs for a given Prs provided
224
by a HCSA in response to an Host Exposure Regime. PE is in the units of the E problem
space, for example Euclidean distance for a colour space specialisation.
n
1$
P opulationError(E, Prs ) = HostError(Bi , Hrs i ) (6.1)
n i=1
Population Diversity The Host Diversity (HD) measure defined in Equation 5.3 pro-
vides an instantaneous diversity a TCSA. The same approach used in HD for calculating
the diversity between repertoires of cells may be used at the population level if the tissues
of each host are compressed to a single repertoire. The measures assumes such a reduction
permitting the direct calculating of a Bit Frequency Histogram (defined in Algorithm 5.8)
for each H, permitting the comparison of Hosts using Bit Frequency Difference (defined
in Equation 5.2), and ultimately the calculation of diversity according to Equation 5.3.
For consistency in the measures used in the host paradigm, this measure is re-branded
Population Diversity (PD).
Component Measures
Average Host Diversity The Average Tissue Diversity (ATD) defined in Equation 5.4
provides an instantaneous measure of the diversity of the average T of a host measured
in bits. This measure may be averaged across all hosts in the population P to provide
a measure of the average host diversity as the average of the diversity of tissues (defined
in Equation 6.2). Alternatively, Average Host Diversity may be calculated as the Host
Diversity (defined in Equation 5.3) averaged for all hosts in the population. This measure
will provide more fidelity for population configuration in which hosts have more than a
single tissue.
n
1$
AverageHostDiversity(P ) = AverageT issueDiversity(Hi ) (6.2)
n i=1
Average Host Error In the same manner as the Average Tissue Error (ATE) in Equa-
tion 5.5, an average error may be defined that exploits the system-level error calculation
for each component of the system. Equation 6.3 defines the Average Host Error (AHE)
that calculates the Population Error (defined in Equation 6.1) for E against each host H
in P , that is averaged by the number of hosts in the population.
 
Hn En
1 $ 1 $
AverageHostError(E, P ) =  HostError(Bj , )Hrs i ) (6.3)
Hn i=1 En j=1

This section considers the expected emergent effects and behavioural trends of host-based
systems in population and generational architectures as defined in the specific realisations.
Although specific to the Host Paradigm, the general behavioural trends are expected to
be qualitatively similar to those expected and observed in the Tissue Paradigm. The
225
primarily reason for this is the reuse of varied exposure regimes as a decoupling mechanism
between systems and environment for the spatial-temporal regularity or lack there of in
the interaction between the two.
Algorithm Mechanism and Strategy
Section 6.3.3 proposed a population and a generational host architectures providing two
different foundational organisations of hosts. This section considers the emergent effects of
these two organisations contrasting their mechanism and information management strate-
gies. Table 6.3 provides summary of the two models in these terms.
Model Mechanism Strategy

Population Population of whole systems. Independent perspectives of problem.
Generation Systems with life-cycles. Independent perspectives with multiple trials.
Table 6.3: Summary of the difference in mechanism and strategy between the complemen-
tary concerns of host algorithms under discrete exposures.
From a population perspective, each host provides a different independent perspective

of the antigenic environment. Inter-host interaction promotes sharing and improvement of
the independent toward improving the competence of individuals and ultimately the en-
tire population. This provides a motivation for information management strategies for the
population model, highlighting the concerns of information dissemination and localisation
in a population (similar to that in the Tissue Paradigm), toward population capability
robustness. From a generational perspective, the population turnover provides an oppor-
tunity for individual perspectives to be refreshed, and for individual host configurations
(starting conditions) to be re-trialled in the same or similar antigenic environment. From
the abstract perspective, the generational model builds upon the population model, tak-
ing the concerns of inter-population interaction for granted, focusing on the mechanism
and effects of inter-generational host interactions. The population in each generation is
independent from the generation before and after it, thus the motivating concerns of in-
formation dissemination and sharing from the population model are directly applicable
for motivating intra-generational information management strategies. Therefore the gen-
eral motivations and expectations of the minimal and extended strategies for both host
architectures may be summarised as follows:
1. Population Strategies
(a) Minimal population promotes multiple and potentially variable perspectives on

a problem.
(b) Intra-population interaction promotes dissemination between isolated perspec-
tives.
2. Generation Strategies
(a) Minimal generational promotes multiple trials from multiple and potentially
variable perspectives on a problem.
226
(b) Intra-generational interaction promotes dissemination about acquired informa-
tion between trials.
Host Exposure Trends
In the tissue paradigm the urgency of acquisition and effective application were promoted
by the metaphor of injury by unattended infection, such urgency is not promoted in the
host paradigm. Instead, hosts unlike tissues are more expendable as long as some of
the hosts population acquire and employ an effective immunity, the ‘system’ (population)
will survive. This disposability of components is fostered more so in the generational
architecture as the mechanism for population assessment and turnover (new component
creation) are explicitly defined. An important consideration in the Host Paradigm is the
independence of perspective, in particular the capability of different hosts to respond in
varied ways to the same stimulus. This may be investigated by varying host configuration
and starting conditions within a population and exposing them to a symmetric antigenic
environment to assess emergent effects. Alternatively, the host exposure regimes provide
an opportunity for investigating the same effects using the opposite case of the same host
configuration exposed to regimes varying in regularity and scope of population exposure.
The integration of these two complementary cases represents the pinnacle host-model
complexity, where the variation of perspective promoted by the variability of exposure
regime provides an idealised and controlled model with strong similarity to model used in
the tissue paradigm. This reuse of exposure regimes in Tissue and Host Paradigms promote
similar expectations as to their general effects in terms of localisation and dissemination
across the components of the system although under different constraints. Specifically, a
population of host systems interact in different ways from that of the fixed intra-tissue
interaction.
Table 6.4 classifies the five Host Exposure Regimes in terms of their general attributes,
including information distribution, nature, and consistency. One may consider the distri-
bution of the exposure of distinct information to the components of the system with the
different host exposure regimes. Symmetrical exposure suggests that each component has
the opportunity of being exposed to any piece of information in the environment, whereas
asymmetrical constraints this expectation. The opportunity may be deterministic, such as
a certainty of being exposed to all or some of the information environment or probabilistic
in those regimes with a stochastic element. Scope provides a general notion as to how
much of the system has the opportunity for interacting with the antigenic environment
irrespective of the specific information it may be exposed to. The nature of the regime
defines the basis of the exposure mechanism which directly relates to the consistency of
the component-information match ups each each exposure event.
Localisation and Dissemination Trends
Information dissemination promotes generalised capability whereas information localisa-

tion promotes specialised capability. In the tissue paradigm these two contrasting manage-
ment types were described with regard to their emergent effects in the context of a tissue
227
Regime Distribution Scope Nature Consistency
AHER Symmetrical System-Wide Deterministic Regular
OHER Symmetrical System-Wide Probabilistic Semi-Regular
PHER Symmetrical Constrained Deterministic Regular
RHER Asymmetrical System-Wide Probabilistic Irregular
SHER Asymmetrical System-Wide Deterministic Regular
Table 6.4: Classification of Host Exposure Regimes in terms of general attributes.
model, specifically: the ‘consistency of response’ and ‘spatial organisation of information’

respectively. These two effects may be rephrased in the context of the host models as the
following:
1. Dissemination: Consistency in host perspective of the environment across the pop-

ulation and generations.
2. Localisation: Specialisation in host perspective of the environment across the popu-

lation and generations.
The general observational trends for both management types are expected to hold for
hosts as they did for investigations into Tissue Paradigm. Specifically, an increase in the
population generalisation results in Average Host Error that approaches the Population
Error with decrease in system diversity as components are more similar, contrasted with
a relative decreased in system error with population specialisation resulting in an increase
in system diversity as component become more heterogeneous.

The research agenda for this chapter and perhaps the Host Clonal Selection Paradigm is
to investigate the population generational models in the context of the expected emer-
gent effects and behavioural trends. Specifically the relative qualitative comparison of
the localisation and dissemination of information of different decentralised information
management strategies under a variety of information exposure regimes. This may be
summarised as follows:
1. Investigate strategies to promote intra-population (population model) localisation

and dissemination of acquired information.
2. Investigate strategies to promote intra-generational (generational model) localisation

and dissemination of acquired information.
6.5 Population Elicited Immunity

6.5.1 Antigen Transmission
Eliciting an immune response is the way in which the immune system acquires information
about the antigenic environment. As reviewed in Section 6.2.1 this active form of immu-
nisation may occur naturally through normal interactions with pathogen and artificially
228
by exposing a system to harmless molecules with similar features to the antigen of inter-
est that elicits the same response. The artificial elicitation of immunity is the principle
behind vaccination’s against disease. In both the artificial and natural cases, eliciting an
immune response provides control over what information a host should acquire and when
it should be acquired, although this control may or may not reside with the host. This
control may be abstracted to the general sampling of antigen from the environment by a
host and future transmission to other hosts in the population. This sampling and trans-
mission provides a mechanism by which a given immune system may share information
about the antigenic environment with other immune systems in the population. Individ-
ual immune systems must collect and store samples of antigen to which they are exposed
for transmission to other systems in the population. This mechanism is referred to as
the Antigen Sampling Scheme. The sampled antigen must then be passed on to other
immune systems with which a given immune system interacts. This passing on is referred
to as the Antigen Transmission Scheme. The antigen sampling mechanism is dependent
on the antigenic environment, more explicitly the host antigenic exposure regime the sys-
tem is subjected to by the environment and other hosts (Antigenic Habitat). The antigen
transmission mechanism is dependent on the intra-population (inter-host) interactions fa-
cilitated by the population structure, host environment, and any interaction intentions of
the individual systems themselves.
The specialisation of the Population Host Clonal Selection Algorithm (defined in Al-
gorithm 6.2) with an antigen sampling and antigen transmission scheme is called the
Transmission Host Clonal Selection Algorithm (THCSA). Algorithm 6.4 defines a generic
antigen sampling scheme where all habitats exposed to a population are recorded by the
hosts for future transmission. The principle of the algorithm is for hosts in a population
to inform other hosts in the population in either a controlled or uncontrolled manner re-
garding antigen of interest, and to allow those hosts to elicit their own internal response.
The algorithm is reasonably general, such that it may be specialised in a number of dif-
ferent ways inspired by population-immune system interactions. Two specialisation of the
THCSA are defined: (1) uncontrolled transmission scheme inspired by pathogen dynamics,
and (2) a host controlled transmission scheme inspired by vaccination dynamics.
Algorithm 6.4: Stored Exposures for Transmission Host Clonal Selection.

Input: E, P
1 foreach Bi ∈ E do
2 P & ← SelectHostsToExpose(P);
3 foreach Hi & ∈ P & do
4 Exposure(Hi &, Bi );
5 Hi &.exposures ← Bi ;
6 end
7 end
229
Pathogen Transmission Dynamics
A first simple specialisation of the THCSA is to that of an infectious pathogen that

may be transmitted from host-to-host. Transmission is achieved via host contact, the
number of hosts, and the time of infectiousness (storage time for transmission), all of
which are a function of the pathogen. The approach facilitates pathogen behaviour not
considered at either the cellular or tissue scales. So-called ‘host-mediated’ pathogenic
exposure regimes facilitate emergent information processing that reformulate a given Host
Exposure Regimes such as epidemics (rapidly spreading), and pandemics (complete or
close to complete population exposure). The specialisation of the THCSA is defined as
follows:
• Sampling Scheme: Samples are collected when a host is exposed to pathogen from
the environment or another host. The longevity of the sample in storage is a property
of the infectiousness of the pathogen. A host may be a carrier for more than one
pathogen at a given time.
• Transmission Scheme: A host may transmit a carried pathogen to any systems to

which it comes into contact with (for as long as the host is a carrier). Interactions
with other immune systems may be random or mediated by proximity, such as in a
spatial organisation.
The pathogen dynamics inspired approach to the transmission algorithm is called the
Pathogen Transmission Host Clonal Selection Algorithm (PT-HCSA). The concerns of
pathogen infectiousness and host-carriers are simplified to the collection of all exposed
antigenic information (as per Algorithm 6.4), the random selection and transmission of
sampled antigen, and the clearing of sampled antigen for each population exposure (al-
gorithm epoch). Two variations of the PT-HCSA are proposed: (1) in which transmis-
sions occur between random pairings of infected hosts, and (2) where transmissions occur
between random neighbouring hosts in a one-dimensional fixed-spatial environment. Al-
gorithm 6.5 defines a specialisation of the HostInteractions for random pairing based
transmission.
Algorithm 6.5: Random Pairings for Pathogen Transmission Host Clonal Selection.
Input: P
1 foreach Hi ∈ P do
2 if Hi .exposures #=0 then
3 H& ← 0;
4 while H& ≡ Hi do
5 H& ← SelectRandomHost(P);
6 end
7 B& ← SelectRandomExposure(Hi .exposures);
8 Exposure(H&, B&);
9 end
10 end
230
Algorithm 6.6 defines a specialisation of the HostInteractions for the random spa-
tial pairings approach to pathogen transmission where hosts are arranged into a one-
dimensional ring structure that limits the random pairing of hosts. Those hosts with
sampled exposures transmit a single exposure to a randomly selected spatial neighbour.
Algorithm 6.6: Spatial Pairings for Pathogen Transmission Host Clonal Selection.
Input: P
2 if Hi .exposures #= 0 then
3 H& ←0;
4 if RandomDouble() < 0.5 then
5 if Hi ≡ Hn then
6 H& ← H1 ;
7 else
8 H& ← Hi+1 ;
9 end
10 else
11 if Hi ≡ H0 then
12 H& ← Hn ;
13 else
14 H& ← Hi−1 ;
15 end
16 end
17 B& ← SelectRandomExposure(Hi .exposures);
18 Exposure(H&, B&);
19 end
20 end
This specialisation of THCSA highlights an important property of the transmission

principle. Specifically that the antigenic environment (exposure regime to which a pop-
ulation is exposed) may be augmented by intra-population interactions between hosts.
Further, the causal influences to inter-population interactions may also be considered in-
direct influences of host-transmitted pathogenic exposures, such as a hosts proximity to
other hosts, host density, or interaction with other hosts from ‘high pathogenic exposure’
areas of the environment. A feature of host-based exposures is that a given host may
be a carrier for one or more pathogens to which an interacting system may be exposed.
In effect, a given host may represent a micro-environment of antigenic exposures, each
potentially providing an exposure regime in and of themselves.
Vaccination Transmission Dynamics
The pathogenic dynamics configuration provides a specialisation of the transmission algo-

rithm, in which the sampling scheme is defined by the pathogen, and is thus outside the
control of the host system. In this configuration, the host is provided with control over
both the sampling and transmission mechanism. A host may select both the antigen to
collect and store, and a small sample of host systems from the population to which to
transmit the antigen (exposure). This specialisation of the algorithm is inspired by vacci-
231
nation and inoculation principle from immunology in which a small sample of pathogen is
given to a healthy system to illicit immunity to the pathogen. The scheme is simpler than
that of the pathogen dynamics scheme in that (1) the host has control over selecting which
pathogen to transmit, and (2) the host has control over selecting a ‘healthy’ system to
which to transmit a sampled pathogen. The operations for the vaccination specialisation
of the THCSA are defined as follows:
• Sampling Scheme: Samples are collected when a host is exposed to pathogen, al-
though the host has discriminatory control over which pathogen are collected and
stored and which are not. Further, the host may decide to discard stored pathogen
in the acquisition of additional information from the environment (such as exposure
frequency).
• Transmission Scheme: A host interacts with systems according to the properties of

the environment and the population structure, although the host has control over
selection of which encountered system to transmit a sample pathogen to. An example
is that a host may select to infrequently inoculate those systems that have a lower
probability of being exposed to a given pathogen.
As with PT-HCSA, a simplified exposure sampling scheme is used where a randomly

selected sampled exposure is transmitted, where sampling is refreshed each epoch. This
Vaccination Transmission Host Clonal Selection Algorithm (VT-HCSA) (defined in Algo-
rithm 6.7) has a parameter Nvaccinate that defines the number of hosts in the population
a single randomly selected host may vaccinate with a single sampled exposure.
Algorithm 6.7: HostInteractions for Vaccination Host Clonal Selection.

Input: P, Nvaccinate
2 P & ← SelectAllExposedHosts(P);
3 H& ← SelectRandomHosts(1, P &);
4 B& ← SelectRandomExposure(H&.exposures);
5 Psubset ← SelectRandomHosts(Nvaccinate , P);
6 foreach Hi ∈ Psubset do
7 Exposure(Hi , B&);
8 end
9 end
The host selection mechanism provides a powerful tool for the population of immune
systems to artificially manipulate the pathogenic exposure regime. This manipulation
will likely have the intention of improving ‘coverage’ by the population for the inocu-
lated pathogen. A host may decide to push the limits of the transmission scheme (mass
vaccinations) either under special circumstances (adaptive transmission), or all the time.
Vaccination of a large sample of the population may provide benefits to the entire popu-
lation given a ‘herd immunity’ effect.
232
6.5.2 Transmission Empirical Study
Aim
The aim of this empirical study was to investigate the Transmission Host Clonal Selection
Algorithm, specifically the pathogen and vaccination configurations in the context of their
information dissemination and localisation capabilities as compared to the Minimal Popu-
lation Host Clonal Selection Algorithm under a variety of host exposure regimes. Toward
this end, the study had the following goals:
1. Investigate the relative effects of the two defined host interaction methods for the
uncontrolled pathogen transmission scheme.
2. Investigate the relative effects of the controlled elicitation of immunity of the vacci-
nation transmission scheme under different vaccination sample sizes.
Method
Algorithms The study considered the Minimal Population Host Clonal Selection Al-
gorithm (MP-HCSA) and the Transmission Host Clonal Selection Algorithm (THCSA),
with two specialised configurations: the Pathogen Transmission Host Clonal Selection Al-
gorithm (PT-HCSA) and the Vaccination Transmission Host Clonal Selection Algorithm
(VT-HCSA). MP-HCSA is a specialisation of the HCSA defined in Algorithm 6.2, that
was configured with Nhosts = 10. Each H was configured with a single tissue (Ntissues = 1)
that was an instance of the Replacement Cellular Clonal Selection Algorithm (RCCSA) de-
fined in Algorithm 4.5, with the configuration Ncells = 50, Nselection = 1, and Nclones = 5.
The method for sampling all exposures in THCSA defined in Algorithm 6.4 was used
for both specialised transmission regimes. Two different pathogen transmission schemes
were investigated, the random pairings scheme defined in Algorithm 6.5 referred to as
(PT-HCSA-RP), and the random spatial parings defined in Algorithm 6.6 referred to as
(PT-HCSA-SP). The vaccination regime defined in Algorithm 6.7 was used with a small
vaccinated set (VT-HCSA-S) Nvaccinate = 1 (10% of the population) and a large vaccinated
set (VT-HCSA-L) Nvaccinate = 5 (50% of the population).
Problems The colour space specialisation of the Habitat Antigenic Exposure Paradigm
(HAEP) defined in Algorithm 6.1 was used called the Habitat Colour Space Problem
(HCSP). The minimal variation of HCSP was used with Nhabitats = 10, and Ndeterminants =
Nantigen = Ninf ections = 1. Each B was a Colour Space Pattern (CSP) was randomly
generated at the beginning of each run. The five different Host Exposure Regimes (HER)
defined in Section 6.4.1 were used for the HCSP. These included the AHER where the
number of H matched the number of B (one-to-one), SHER, RHER, PHER where Hpoint
was fixed at tissue 1, and OHER with duration = 15.
fined in Equation 4.8 with M axEpochs = 1000. The four host specific measures defined
233
in Section 6.4.3 were collected from the state of the system after the triggering of the stop
condition. These included the system measure PE and PD, and the host measures AHE
and AHD. Each algorithm and problem received a new and different random number seed
each run. Algorithm-Problem combinations were repeated 30 times.
Results
ues. Box-and-whisker plots are provided in which the results for each algorithm are ag-
gregated across all HER for a each measure. Figure 6.3 shows PD, Figure 6.4 shows PE,
Figure 6.5 shows AHD, and Figure 6.6 shows AHE.
Figure 6.3: Box-and-whisker plot of Population Diversity (PD) across all HER for the
THCSA study.
Analysis
This section provides an analysis of the results from the empirical study into the THCSA
summarised in Table A.5. These analyses exploit the trends and expectations outlined in
Section 6.4.4, and in particular the attributes of host exposure regimes in Table 6.4.
Pathogen Transmission Trends This section considers the constrained (spatial) and
unconstrained random host pairing methods for the pathogen transmission scheme com-
pared to the MP-HCSA that does not facilitate inter-host communication. From a system
perspective the effect of pathogen transmission was restricted to those exposure regimes
with an asymmetric distribution of information to the hosts in the population. Pathogen
transmission resulted in an increase in population diversity and population error compared
234
Figure 6.4: Box-and-whisker plot of Population Error (PE) across all HER for the THCSA
study.
Figure 6.5: Box-and-whisker plot of Average Host Diversity (AHD) across all HER for the
THCSA study.
235
Figure 6.6: Box-and-whisker plot of Average Host Error (AHE) across all HER for the
THCSA study.
to MP-HCSA on AHER, OHER, and PHER, with no significant difference between the
three approaches at the system level on RHER and SHER. The random pairing method
resulted in a higher system diversity and system error (increased effect) compared to the
constrained spatial random pairing method. The same trend was observed at the com-
ponent level with regard to the symmetry in information distribution by the exposure
regimes. No significant difference was observed between the three approaches on RHER
and SHER with regard to AHD and on SHER with regard to AHE. A relatively large
decrease in component diversity and component error was observed with both pathogen
transmission schemes on AHER, OHER and PHER. As as was the case with the system
measures, the component measures observed an increase in effect with the unconstrained
random pairing method than the constrained random spatial pairing method. The ob-
servations regarding the two pathogen transmission schemes may be generalised to the
following trends:
1. Independent perspectives of the information environment are promoted by asymmet-

rical and not symmetrical information exposure with the chosen host configuration.
2. Pathogen transmission increases system diversity and error, and decreases compo-
nent diversity and error on asymmetrical information exposure regimes.
3. The pathogen dissemination effect was more pronounced with random pairing com-
pared with spatial pairing.
Vaccination Size Trends This section considers the vaccination transmission scheme
compared to the MP-HCSA, and in particular the effect of varying the number of hosts
selected for vaccination per population exposure (Nvaccinate ). From a system perspective
236
the vaccination scheme demonstrated the same general trend as the pathogen scheme
with regard to the restriction of the dissemination effect to those exposure regimes with
an asymmetrical information distribution. Vaccination resulted in an increase in system
diversity compared to MP-HCSA, and generally no significant change to system error
other than a small increase in OHER. The large vaccination size demonstrated a larger
increase in system diversity over MP-HCSA than the smaller vaccination size. From a
component perspective vaccination demonstrated a general decrease in the Average Host
Diversity and Average Host Error on the asymmetric information distribution exposures.
This observed dissemination effect increased with the increase in the vaccination sample
size. This result relates to the herd immunity effect, as effective information dissemination
is required to facilitate majority coverage in the population under a probabilistic exposure
regime. A stronger connection to the herd immunity effect may be investigated in the
future by considering a probabilistic variation of PHER with spatial-temporal consistency
like the OHER where the dissemination effect for a one or a number of antigenic habitat’s
could be assessed under a variety of Nvaccinate values. The expectation is that the herd
immunity effect would be correlated with the improvements to the already measurable
dissemination effect (relative decreases in average component diversity and error). The
observations regarding the vaccination transmission scheme with varied vaccination sizes
may be generalised to the following trends:
1. Vaccination promotes dissemination of information that increases system diversity,

whilst at the same time decreases average component diversity and error only on
exposure regimes with asymmetric information distribution across the population.
2. The effects of dissemination on the average component is more pronounced with the
larger vaccination size (lower relative Average Host Diversity and Error).
Dissemination Trends Table A.6 in Appendix A.2 contains the results from the ran-
dom pairing pathogen transmission configuration and the vaccination transmission with
large Nvacinations . These two approaches demonstrated better information dissemination
for their respective approaches. This section compares these two different approaches
across the five exposure regimes in terms of information dissemination. From a system
perspective there was little significant difference in system diversity and error between
the pathogen and vaccination transmission schemes. The only observed significant differ-
ence at the system level was on the PHER with an asymmetric information distribution.
The average component perspective provided more insight into the observable differences
between the two transmission schemes. The pathogen scheme demonstrated a slightly
lower average host diversity on AHER, and a relatively slight increase on the PHER com-
pared the vaccination. The vaccination scheme demonstrated a lower average error on the
PHER, whereas the pathogen approach demonstrated a slight decrease in error over vac-
cination on AHER. This differentiation between the two methods suggests that pathogen
transmission provides some benefit on exposure regimes with asymmetric distribution and
a system-wide scope, where as vaccination is suited to the constrained scope exposure
with asymmetric distribution. This is an intuitive result as one-to-many dissemination is
237
preferred for constrained exposures, and one-to-one dissemination for system-wide expo-
sures. The observations made regarding pathogen and vaccination transmission may be
generalised to the following trends:
1. There was no significant difference between the pathogen transmission and the vac-
cination transmission approaches on exposure regimes with symmetric information
distribution.
2. Pathogen transmission schemes provide improved dissemination on system-wide ex-

posure with asymmetric information distribution, whereas the vaccination transmis-
sion scheme provides improved dissemination on constrained exposure with asym-
metric information distribution.
Conclusions
This section summarises the findings of the empirical study into the Transmission Host
1. Primitives
(a) Pathogen transmission with random pairing provides a viable metaphor for
information dissemination via elicitation of response.
(b) Vaccination transmission provides a viable metaphor for information dissemi-
nation via elicitation of response that improves with vaccination size.
2. Transmission
(a) There was generally little difference in the dissemination effect between PT-
HCSA-RP and VT-HCSA-L, except the pathogen approach demonstrated a
relatively slightly better effect on system-wide exposure as opposed the the
relatively slightly better effect observed with the vaccination approach on con-
strained exposure with asymmetrical information distribution.
(b) Exposure regimes with an asymmetric information distribution are required to
engender a varied perspective of the antigenic environment in hosts permitting
the observation of the dissemination effect.
This last point highlights the important finding from the results that confirmed a
base assumption of the empirical investigations into host-based system, which is that the
chosen minimal configuration does not result in significantly independent perspectives
on the same antigenic environment. This result was demonstrated in the results for the
symmetrical RHER and SHER regimes. Alternatively, this result may be interpreted as
the confirmation that an exposure regime with an asymmetric information distribution
is required to observe varied perspectives of the antigenic environment with the chosen
host configuration, allowing the dissemination effect to be observed and measured. The
configuration of transmission schemes were simplistic, providing much room for elaboration
238
and improvement. An important consideration to motivate further investigation is the use
of both pathogen and vaccination based transmission schemes in parallel providing the
benefits of varied cardinality-based dissemination (one-to-one, one-to-many) suitable for
an unknown exposure regime.
6.6 Population Shared Immunity

6.6.1 Effector Transmission
An alternative to actively eliciting an immune response is to allow hosts to sample and
share the product of immunisation. Section 6.2.1 defined this as passive immunisation
highlighting that the immunity provided by inter-host transfer of antibodies is fast, effec-
tive, and temporary. This section considers the artificial transplantation of passive im-
munity between the hosts as a strategy for disseminating specialised acquired information
about the antigenic environment between hosts in the population. The Shared Immunity
principle is defined as the sharing of acquired immunity between hosts in a population by
explicit sampling their own cellular repertoire and transmitting sampled cells to other se-
lected hosts in the population. The sharing requires a sampling scheme of a hosts internal
cellular repertoire, and the explicit selection of one or more other sibling host systems in
the population to which sampled cells are transmitted. This sampling and transmission
facilitates horizontal sharing of acquired immunity within the population. The sampling
scheme is responsible for selecting those mature cells most likely to be useful to those
host systems to which the sampled cells are transmitted. There is expected to be a tight
coupling between the selection (sampling) of a hosts cellular repertoire and the selection of
the recipient host or hosts. Sampled cells will be mature in that they have been produced
as a result of an interaction with the antigenic environment. More useful, are those cells
that are created from a series of exposures, and thus represent more refined and presently
useful information about the environment. The selection of recipient host systems via a
transmission scheme is the selection of those systems in the population that would most
benefit from the sampled acquired immunity. Examples include the transmission of cells
from a healthy system (good performance with regard to some system-environment mea-
sure), to a system or set of systems that are less healthy (as defined by the same measure).
In the absence of such a measure, a good heuristic is the transmission of a representative
sample of acquired information to a host that is geographically distant (with regard to
the antigenic environment). This heuristic provides a general sharing scheme in which the
hosts of the system seek to provide general coverage to the population by sharing acquired
immunity. After hosts are selected and the cells are transmitted, the recipient hosts must
integrate the received cells into their cellular repertoire.
• Sampling Scheme: The identification (selection) and collection of acquired immune

cells from a given hosts cellular repertoires for transmission to one or more other
hosts (as defined by the hosts cell transmission scheme). Those cells sampled for
transmission should be mature with regard to the antigenic environment, and rep-
resentative of information acquired during the hosts recent exposures.
239
• Transmission Scheme: Requires the selection of sibling hosts within the systems
population of immune systems, and the injection (transmission) of sampled cells
to selected hosts. The principle concern in the selection of recipient hosts is that
the transmitted cells are likely to be useful. This likelihood may be increased by
selecting sibling immune systems in the population that are in turn likely to have
been subjected to different pathogen and/or at different frequencies (assuming an
asymmetric antigenic environment). Once selected, the transmitted cells must be
integrated into the recipient systems cellular repertoire.
Algorithm 6.8: HostInteractions for Shared Immunity Clonal Selection.

Input: P, Nsharers , Nrecipients , Nsharedcells
1 for s←0 to Nsharers do
2 Hs ← RandomHost(P);
3 T & ← RandomTissue(Hs , Nsharedcells );
4 P & ← P;
5 while P &n #= Nrecipients do
6 H& ← RandomHost(P &);
7 P &.Remove(H&);
8 end
9 foreach Hi & ∈ P & do
10 Hi &.Integrate(T &);
11 end
12 end
The Population Host Clonal Selection Algorithm defined in Algorithm 6.2 may be
specialised with the the cell sampling and transmission of shared immunity, referred to as
the Shared Immunity Host Clonal Selection Algorithm (SI-HCSA). Algorithm 6.8 provides
a definition of the HostInteractions operation in which a number of hosts (Nsharers ) select
a given number of other hosts in the population (Nrecipients ) and share a fixed number
(Nsharedcells ) of randomly sampled cells. Transmitted cells are copied from the sharing
host and are integrated into the receiving hosts using an integration operation (Integrate),
and host selection both with regard to sharers and receivers is random (RandomHost).
This provides a simple realisation of the shared immunity algorithm without cell or host
selection bias.
6.6.2 Shared Immunity Empirical Study

Aim
The aim of this empirical study is to investigate the Shared Immunity Host Clonal Selection
Algorithm in the context of the information dissemination and localisation capabilities as
compared to the Minimal Population Host Clonal Selection Algorithm under a variety of
host exposure regimes. Toward this end, the study had the following goals:
1. Investigate the relative effects of of small and large sharer and recipient sizes under
shared immunity
240
2. Investigate the relationships between sharer and recipient sizes and information dis-
semination.
3. Assess whether the trend of information dissemination restriction of exposure regimes

with asymmetric information dissemination holds.
Method
Algorithms The study considered the Minimal Population Host Clonal Selection Algo-
rithm (MP-HCSA) and the Shared Immunity Host Clonal Selection Algorithm (SI-HCSA).
MP-HCSA is a specialisation of the HCSA defined in Algorithm 6.2, that was configured
with Nhosts = 10. Each H was configured with a single tissue (Ntissues = 1) that was
an instance of the RCCSA defined in Algorithm 4.5, with the configuration Ncells = 50,
Nselection = 1, and Nclones = 5. SI-HCSA was defined in Algorithm 6.8, where the in-
tegration operation (Integrate) was specialised to the Euclidean-based replacement as is
used in the RCCSA. The number of shared cells was fixed Nsharedcells = 5 and the number
of sharers and recipients were varied. A small number of sharers Nsharers = 1 (10% of
the population) and a large number of sharers Nsharers = 5 (50% of the population) were
assessed with a small number of recipients Nrecipients = 1 (10% of the population) and a
large number of recipients Nrecipients = 5 (50% of the population).
Problems The same Habitat Colour Space Problem and Host Exposure Regimes were
used as was defined for the THCSA empirical study in Section 6.5.2.
Experiment The same experimental setup was used as was defined for the THCSA
Results
ues. Box-and-whisker plots are provided in which the results for each algorithm are ag-
gregated across all HER for each measure. Figure 6.7 shows PD, Figure 6.8 shows PE,
Analysis
This section provides an analysis of the results from the empirical study into the SI-HCSA
Section 6.4.4 regarding information dissemination between the hosts in the population.
Sharer Size Trends This section is concerned with the effect of the number of hosts
sharing acquired immunity in the population (Nsharers ), irrespective of the number of re-
cipients (Nrecipients ). As such, the observations are taken as the general trends in the small
and large number of sharer hosts. From a system perspective the Population Diversity
241
SI-HCSA study.
Figure 6.8: Box-and-whisker plot of Population Error (PE) across all HER for the SI-
HCSA study.
242
Figure 6.9: Box-and-whisker plot of Average Host Diversity (AHD) across all HER for the
SI-HCSA study.
SI-HCSA study.
243
was generally lower with sharing compared to MP-HCSA across all exposure regimes. The
increase in sharer size generally resulted in a decrease in diversity. A similar relationship
was observed with system error, where sharing generally resulted in an increase in sys-
tem error compared to MP-HCSA, an effect that increased with the number of sharers
(increase in share size increased system error). From a component perspective shared im-
munity resulted in a general decrease in the Average Host Diversity than MP-HCSA. This
effect decreased with the increase in number of sharing hosts, such that a AHD approached
MP-HCSA on the AHER, OHER, and RHER domains. Interestingly, a similar effect was
observed with the Average Host Error that was generally lower than MP-HCSA, although
increased with the increase in the sharer size. This correlation in relationships at the
system and component level with share size demonstrates a disruptive and information
dissemination effect at the system and component level respectively that generally results
in increased disruption at both scales with the increase in the number of hosts sharing
immunity. The observations may be generalised to the following trends:
1. System
(a) Shared immunity results in a generally lower population diversity and generally
higher population error than MP-HCSA.
(b) Increase in sharing size results in increase in system level effects, specifically
decreases in system diversity and increases in system error.
2. Component
(a) Shared immunity results in a general increase in Average Host Diversity, and a
general decrease in Average Host Error compared to MP-HCSA.
(b) Increase in a decrease in component level effects with an increase a decrease in
Average Host Diversity, and a general increase in Average Host Error.
Recipient Size Trends This section considers the effects of varying the number of
recipient hosts (Nrecipients ) irrespective of the number of hosts sharing. As such, the
results are generalised across the configuration values for the number of sharers (Nsharers ).
From a system perspective Population Diversity and Error exhibited the same trends
both with regard to the comparison to MP-HCSA and with regard to the increase in
the number of recipients increasing the disruptive effect of sharing at the system level.
The same relationship was demonstrated at the component level with a general increase
in diversity and decrease in error compared to MP-HCSA, and the increased disruption
(decreased average diversity and increased average error) with the increased in the number
of recipients. These observation demonstrate an important symmetry in general behaviour
between the number of hosts sharing acquired immunity and the number of hosts receiving
acquired immunity. In particular increasing either results in a increase in the disruption
at the system level and disruption to the dissemination of information at the component
level. These observations may be summarised in the following generalised trends:
244
1. The system and component-level effects of recipient size mirror the effect trends of
number of sharing hosts.
(a) System diversity decrease and system error which increases with recipient size.
(b) Component diversity increases and component error decreases, an effect that
decreases with recipient size.
Sharer and Recipient Relationship Trends This section considers the trends in the
relationships between the number of hosts sharing and the number of hosts receiving shared
immunity, not considerating the comparison the MP-HCSA. From a system perspective
the highest population diversity was observed with a one-to-one relationship, whereas the
lowest population diversity was observed with a many-to-many configuration. Generally
there was no significant difference between one-to-many and many-to-one configurations
across the exposure regimes, with PHER as the exception. The same general trend was
observed with regard to population error, where the lowest error was achieved by one-
to-one and the largest system error was achieved by many-to-many. Interestingly, no
significant difference in system error was observed between SI-HCSA-SS and MP-HCSA
on all exposure regimes except RHER where one-to-one sharing resulted in a lower system
error. The component perspective demonstrated the same general trend where the highest
average component diversity and least average component error were observed with the
one-to-one configuration, and the most disruption in diversity and error caused by the
large scale sharing in the many-to-many configuration. Also consistent was the lack of
significant difference in effect between the one-to-many and many-to-one configurations.
The consistency of the dissemination effect in the relationship between both the number of
sharer and recipient hosts was confirmed both in the extreme cases with the least and most
disruption caused by the least and the most amount of sharing, and in the intermediate
cases confirming the symmetry in behaviour. These observations may be summarised in
the following generalised trends:
1. Generally no significant difference in assessed measures between one-to-many or

many-to-one with regard to the number of sharers and recipients across all HER.
2. The least disruptive system effect was observed with one-to-one, whereas the largest
disruptive system effect was observed with many-to-many sharing.
3. The best dissemination effect at the component level was observed with one-to-one
sharing, whereas the worst disruptive effect at the component level was observed
with many-to-many sharing.
Conclusions
This section summarises the findings of the empirical study into the Shared Immunity
Host Clonal Selection Algorithm, in terms of the primitives that were the focus of the
1. Large scale sharing provides a disruptive effect at the system and component level.
245
2. Small scale sharing minimises the disruptive effect at the system level and maximise
the dissemination effect at the component level.
An interesting aspect of artificial passive immunity is the concern that such shared
information is initially useful although such utility is temporary. This property is expected
to vary with exposure regime and provides an important extension to the investigation
into the SI-HCSA.
6.7 Generational Maternal Immunity

6.7.1 Generational Transmission
Maternal immunity involves progeny receiving antibodies and immune cells from the
mother, both across the placenta and in breast milk (mucosal immunity). This type
of immunity is referred to as natural-passive immunity because the conferred acquired
immune information is passed between the hosts without eliciting an immune response.
Maternal immunity provides a strategy for one generation of immune systems with lifetime
acquired immunity to transfer such acquired knowledge to the following generation. As
with intra-population sharing of acquired immunity via passive means in Shared Immu-
nity (Section 6.6), a cell sampling and transmission scheme are required. The concerns of
cell sampling are predominantly similar to the concerns highlighted in the Transmission
and Shared Immunity Host Clonal Selection Algorithms. It is important to highlight the
trade-off of the sampled cells and their effect on freshly instantiated host systems. A
selection of a sample that is too large or contains many dominant (with respect to clonal
selection) pieces of acquired information (memory cells) will cause the recipient systems
to (in effect) represent continuances of the transmitting host. The sample should be di-
verse and representative, and likely contain many effector cells. The selection of effector
cells for transmission is useful for a number of reasons. Firstly, the biological inspiration
(maternal, and more importantly mucosal immunity) confers mostly this type of acquired
immunity in the form of antibodies.
• Sampling Scheme: The selection of mature acquired immune cells from host systems
to be removed and transmitted to host systems of the subsequent generation. The
selection scheme should draw a representative sample of the information acquired
by the system over the course of its lifetime (trial period).
• Transmission Scheme: The selection of instantiated host systems of the subsequent

generation by hosts of the present generation to which sample acquired immune
systems cells will be transmitted. A simple host selection scheme is the selection of
progeny host systems.
The host transmission scheme requires the selection of hosts in the subsequent genera-
tion to receive the sampled cells. A natural implementation of this scheme is the selection
of progeny hosts. Specifically, the selection of hosts in the subsequent generation by hosts
in the present generation, to which they are responsible for instantiating. The parent-child
246
transmission scheme may result in the formation of independent generational host-lines
given the asexual basis of the approach. The concern of this effect is that if a given host
system is lost, then the information acquired by that hosts generational line is also lost.
This concern may be addressed by relaxing the parent-child transmission constraint, and
allowing parent-hosts to potentially transmit to any host in the child generation. A biolog-
ical basis for this configuration may be the use of manufactured formulae, or the use of a
wet-nurse. An example of a decoupled transmission scheme is random host selection, with
reselection. The reselection allows a given host in the child generation to potentially receive
acquired immune information from more than one parent host system. The transmission
of cells to non-child hosts provides redundancy between host lines, if there is a relationship
between parent-child that may affect host-loss (such as reproduction in a hazardous spa-
tial environment). A one-to-many transmission scheme between the generations will also
foster redundancy of acquired information between the generations. The concern in a host
system receiving acquired immune information from more than one host, is that the child
system may become overly biased by the previous generation. A parent-child (one-to-one)
transmission scheme limits the scope of received acquired immune information to a single
host of the previous generation’s perspective.
The Minimal Generational Host Clonal Selection Algorithm (MG-HCSA) of the genera-
tional algorithm defined in Algorithm 6.3 provides a basis for comparison with a maternal
immunity approach as it promotes independent generations. Algorithm 6.9 provides a
definition of the CreateP opulation operation for the MG-HCSA. Equation 6.4 provides a
generational change trigger based on a number of population exposures or epochs, where
a user specified parameter Ngenepochs defines the maximum number of epochs before a
generational change is triggered.
Algorithm 6.9: CreatePopulation for the Minimal Generational Clonal Selection.

Input: P, Nhosts
Output: P &
1 P & ←0;
4 P & ← Hi ;
5 end
6 return P &;
GenerationalChange(Epochi ) = (Epochi ≥ Ngenepochs ) (6.4)
A Maternal Immunity Host Clonal Selection Algorithm (MI-HCSA) may be defined as

the realisation of intra-generational sharing of acquired immunity that may be specialised
to parent-child based sharing. Algorithm 6.10 defines the CreateP opulation operation for
parent-child based maternal immunity, where the amount of inter-generational sharing is
defined by a user defined parameterL Nmaternalcells .
247
Algorithm 6.10: CreatePopulation for Maternal Immunity Clonal Selection.
Input: P, Nmaternalcells
Output: P &
1 P & ←0;
3 T &i ← SelectRandomTissue(Hi , Nmaternalcells );
4 H&i ← CreateHost();
5 H&i .Integrate(T &i );
6 P & ← H&i ;
7 end
8 return P &;
6.7.2 Maternal Immunity Empirical Study

Aim
The aim of this empirical study is to investigate the Maternal Immunity Host Clonal
Selection Algorithm in the context of the information dissemination and localisation capa-
bilities as compared to the Minimal Generational Host Clonal Selection Algorithm under
a variety of host exposure regimes. Toward this end, the study had the following goals:
1. Compare and contrast generational clonal selection with and without inter-generational
sharing.
2. Investigate the dissemination of information effect and its disruption by varying the
amount of inter-generational sharing.
Method
Algorithms The study considered the Minimal Generational Host Clonal Selection Al-
gorithm (MG-HCSA) and the Maternal Immunity Host Clonal Selection Algorithm (MI-
HCSA). MG-HCSA is a specialisation of the HCSA defined in Algorithm 6.3, that was
configured with Nhosts = 10. Each H was configured with a single tissue (Ntissues = 1)
that was an instance of the RCCSA defined in Algorithm 4.5, with the configuration
Ncells = 50, Nselection = 1, and Nclones = 5. The CreateP opulation operation for the
MG-HCSA was realised as the creation of a new population of hosts each generational
change as defined in Algorithm 6.9, and the epoch generational change condition defined
in Equation 6.4 with Ngenepochs = 100. The MI-HCSA with assessed with the parent-child
specialisation of the CreateP opulation operation defined in Algorithm 6.10. The number
of cells selected for inter-generational sharing was assessed with three different values as
follows: small Nmaternalcells = 5 (10% of the hosts repertoire), medium Nmaternalcells = 10
(20% of the hosts repertoire), and large Nmaternalcells = 25 (50% of the hosts repertoire).
used as was defined for the THCSA empirical study in Section 6.5.2.
248
Experiment The same experimental setup was used as was defined for the THCSA
empirical study in Section 6.5.2, except the maximum number of epochs of the EOSC
stop condition was decreased from 1000 to 999. This configuration change was made such
that the final measures recorded reflected the state of the system after 10 generations of
100 epochs before the last generational change (at the tenth generation).
Results
gated across all HER for a each measure. Figure 6.11 shows PD, Figure 6.12 shows PE,
MI-HCSA study.
Analysis
This section provides an analysis of the results from the empirical study into the MI-HCSA
Generational Trends This section compares and contrasts the minimal generational
approach without sharing with the maternal immunity approach with inter-generational
sharing. MG-HCSA is compared against MI-HCSA-L as it is expected to exhibit the
largest effect of sharing and therefore difference to the minimal approach. From a system
perspective MI-HCSA resulted in a generally population diversity and lower system error
than MG-HCSA across all five of the exposure regimes. Interestingly this trend was not
249
Figure 6.12: Box-and-whisker plot of Population Error (PE) across all HER for the MI-
HCSA study.
Figure 6.13: Box-and-whisker plot of Average Host Diversity (AHD) across all HER for
the MI-HCSA study.
250
MI-HCSA study.
observed on the symmetric exposure regime, where no significant difference was observed.
From a component perspective, maternal immunity generally achieved a lower Average
Host Diversity and lower Average Host Error than the minimal approach. This was gener-
ally observed across the five HER’s, except on SHER and PHER with regard to AHD and
AHE respectively. The results show a clear localisation effect with decreased error and in-
crease in diversity at the system level, with the interesting side effect of decreased average
error at the component level. The general lack of effect on the SHER suggests that the
inter-generational sharing does not provide any benefit under such a strongly specialising
exposure regime. The observations may be generalised to the following trends:
1. Maternal immunity results in improved localisation compared to the minimal gen-

erational approach.
2. Parent-Child inter-generational sharing provides an inter-generational localisation

method resulting in highly-specialised hosts.
Maternal Immunity Size Trends This section considers the effect of varying the num-
ber of cells shared between the generations defined by different values for the Nmaternalcells
user parameter. From a system perspective the results demonstrated an increase in effect
with the number of transmitted maternal cells. Specifically, increase in Nmaternalcells re-
sulted in an a increase in system diversity and decrease in system error. The effect was
less pronounced with small sample sizes, typically with little significant difference between
MG-HCSA and MI-HCSA-S as well as between MI-HCSA-S and MI-HCSA-M. From a
component perspective the increase in the number of transmitted sampled cells resulted
in relative increase in the effect as measured by decreases in Average Host Diversity and
decreases in Average Host Error. The same general lack of significance was observed with
251
small, medium and empty sample sizes. The maternal cell size trend demonstrates that
parent-child generational sharing promotes a strong localisation effect that increases with
the amount of information communicated between the generations across the variety of
exposure regimes. Importantly the effect is not apparent (significant) until a sufficiently
large number of randomly selected cells are transmitted. The observations may be gener-
alised to the following trends:
1. Localisation effect increases with the number of transmitted maternal cells.
2. The number of cells must be of a sufficient size (50%) for the effect to be consistently
significant across the exposure regimes.
Conclusions
This section summarises the findings of the empirical study into the Maternal Immunity
Host Clonal Selection Algorithm, in terms of the primitives that were the focus of the
1. Generational
(a) The assessed specialisation of maternal immunity resulted in improved inter-

generational specialisation than no inter-generational sharing.
2. Parent-Child Sharing
(a) Maternal Immunity with parent-child sharing is an inter-generational localisa-

tion method.
(b) The concern of maternal immunity that sufficiently large sample sizes would re-
sult in homogeneous generations was confirmed to the benefit of the localisation
effect resulting in reduced system and component-wise error.
An important consideration not addressed is the differentiation from generational

learning and lifetime learning. As mentioned in the method, the results were taken from
the end of the run before a generational change, thus reflect generational learning over
9 generations of 100 epochs, and lifetime learning over 99 epochs in the final generation.
Elaborative investigations may consider such a distinction and assess generational learning
as the performance of the population before lifetime learning (at the epoch of creation),
and lifetime learning before the generational change. When such a distinction is not made
as in the case of this and the next empirical studies, the effects are still present, although
aggregated (confounded) together. The parent-child approach assessed is perhaps the sim-
plest of such maternal immunity approaches. One may consider shifts in the cardinality
between sharers in the previous generation and receivers in the next generation, with a
clear relationship to the investigation into SI-HCSA and whether the results hold across
generations. Two additional important questions that may motivate future work include
(1) the consideration as to how much of the new generation must be seeded with maternal
cells for the effect to be observed, and (2) the effect of positioning children in different parts
252
of the antigenic environment (effecting the specialisation of children under the exposure
regimes).
6.8 Generational Evolved Immunity

6.8.1 Generational Inheritance
The evolution of the immune system provides a inter-generational method for propagating
functional aspects of the immune system based on lifetime performance. Unlike maternal
immunity, evolution does not share acquired immune information directly in the form of
cells, rather it shares it indirectly through selection and inheritance of the mechanisms
that create and utilise those cells. An inherited generational algorithm requires a selection
and reproduction mechanisms for hosts. The important differences in this specialisation of
the generational algorithm is that a given host may contribute more or less than the other
hosts in its generation. This is a realisation of natural selection, where a measure of the
hosts in the population discriminates the reproductive fitness and genetic contribution to
the next generation. The principle difference of this specialisation is the use of a genetic-
basis for traits that effect hosts acquired immune system. The principle components of
the inherited generational algorithm are defined as follows:
• Selective Scheme: A specialisation of the host selection scheme of the minimal gen-
erational algorithm that realises the principles of natural selection. A given hosts
contribution to the subsequent generation is differentiated based on assessed fitness
against an inherited trait that effects a hosts acquired immune system.
• Genetic Basis: A genetic code (genome) is used to define a trait of a hosts acquired
immune system. This genetic code provides the basis of natural selection (differenti-
ated reproductive success), and the medium for reproduction (genetic inheritance).
• Reproductive Scheme: The reproductive scheme of the minimal generational population-

based algorithm, that may take into account the broader considerations of asexual
and sexual reproduction. Reproduction provides the basis for inheritance with re-
gard to the genetic basis of the trait or traits that effect a hosts acquired immune
system, and manipulations to that genetic representation in the form of genetic
mutation and/or genetic recombination.
The reproductive scheme provides the mechanism for inheritance of the genetic basis
of an acquired immune system trait. Reproduction provides a duplication of the parents
genetic material which may be modified in minor ways by genetic mutation (likely a lower
rate of mutation than that of hypermutation in the immune response). This introduces
variations of the trait for natural selection to differentiate reproductive success. A property
of the host populations acquired immune system is defined by a parameter or set of
parameters, which is encoded in a genome. The genome is the basis for inheritance, and the
expressed trait is the basis for host selection. For example, an inheritance mechanism may
be defined that encodes information that defines the generation of naı̈ve immune cells in a
253
hosts acquired immune system. Unlike the maternal sharing scheme that directly reinforces
the receptor configurations that are useful, an evolutionary algorithm indirectly reinforces
receptor configurations in the generation of a hosts base repertoire, and ongoing naı̈ve cells.
This is a broader form of sharing that provides the flexibility for the selection mechanism
and scope of the genetic basis for the trait to define the regions of ‘receptor configuration
space’ that are beneficial for a hosts untested acquired immune system to sample. The
mechanism provides an interesting example that combines both genetic-based learning over
generational time and somatic-based learning over a hosts lifetime. Therefore, this class of
host algorithm facilitates learning at two scales, specifically learning at the host-lifetime
scale in somatic adaptations via clonal selection, and learning at the population-generation
scale in genetic adaptations via natural selection. Examples of other traits that may be
subjected to the pressures of evolution by natural selection include (1) the sensitivity of
matching in the clonal selection algorithm (2) and the organisation and connectivity of
tissue types in the lymphoid tissue algorithm. In addition to opening up parameterised
models in the framework to the process of evolution via natural selection (generational
learning), the minimal natural selection algorithm provides a connection of the host-level
of the acquired immune system framework with the field of genetic algorithms.
Algorithm 6.11: CreatePopulation for the Evolved Immunity Clonal Selection.

Input: P, Pmutation , Pcrossover
Output: P &
1 P & ←0;
3 Hi F itness ← HostError(Hi );
4 end
5 Pparents ←0;
7 Hi ← BiasedRouletteWheelSelection(P);
8 Pparents ← Hi ;
9 end
10 foreach Hi , Hi+1 ∈ Pparents do
11 Hi child1 ← Crossover(Hi , Hi+1 , Pcrossover );
12 P & ← Hi child1 ;
13 Hi child2 ← Crossover(Hi , Hi+1 , Pcrossover );
14 P & ← Hi child2 ;
15 end
16 foreach H&i ∈ P & do
17 Mutate(H&i , Pmutation );
18 end
19 return P &;
The Evolved Immunity Host Clonal Selection Algorithm (EI-HCSA) is defined as a spe-
cialisation of the Generational Host Clonal Selection Algorithm (defined in Algorithm 6.3)
that uses a genetic basis, host selection and host reproduction schemes as defined in the
CreateP opulation operation in Algorithm 6.11. The Pmutation parameter defines the prob-
ability of mutating a component in the hosts genetic representation during reproduction,
254
and Pcrossover defines the probability of creating two new host from a cross of two parental
hosts genetic basis. The reproduction scheme is a rudimentary realisation of the classical
genetic algorithm (discussed in Section 3.4.1) for a hosts initial repertoire.
6.8.2 Evolved Immunity Empirical Study

Aim
The aim of this empirical study is to investigate the Evolved Immunity Host Clonal Selec-
tion Algorithm in the context of the information dissemination and localisation capabilities
as compared to the Minimal Generational Host Clonal Selection Algorithm under a variety
of host exposure regimes. Toward this end, the study had the following goals:
1. Compare and contrasted generational clonal selection with and without an inherited
initial repertoire.
2. Investigate the localisation and/or dissemination properties of inter-generational in-

heritance.
Method
Algorithms The study considered the Minimal Generational Host Clonal Selection Al-
gorithm (MG-HCSA) and the Evolved Immunity Host Clonal Selection Algorithm (EI-
HCSA). MG-HCSA is a specialisation of the HCSA defined in Algorithm 6.3, that was
configured with Nhosts = 10. Each H was configured with a single tissue (Ntissues = 1)
that was an instance of the RCCSA defined in Algorithm 4.5, with the configuration
Ncells = 50, Nselection = 1, and Nclones = 5. The CreateP opulation operation for the
MG-HCSA was realised as the creation of a new population of hosts each generational
change as defined in Algorithm 6.9, and the epoch generational change condition defined
in Equation 6.4 with Ngenepochs = 100. The MI-HCSA with assessed with the evolutionary
specialisation of the CreateP opulation operation defined in Algorithm 6.11. Host fitness
was assigned using the AHE measure. The probability of crossover was fixed at 0.90
(90%) per host-pair creation. The mutation duration host creation was 1
192 bits which is
one mutation cell or 50 mutations per population creation ( 9600
50
bits).
used as was defined for the THCSA Empirical Study in Section 6.5.2.
Experiment The same experimental setup was used as was defined for the MI-HCSA
empirical study in Section 6.7.2, expect the maximum number of epochs on the EOSC
stop condition was increased from 999 to 9999. This configuration change was made such
that the final measures recorded reflected the state of the system after 100 generations of
100 epochs before the last generational change at one hundredth generation.
255
Results
gated across all HER for a each measure. Figure 6.15 shows PD, Figure 6.16 shows PE,
EI-HCSA study.
Figure 6.16: Box-and-whisker plot of Population Error (PE) across all HER for the EI-
HCSA study.
Analysis
This section provides an analysis of the results from the empirical study into the MI-HCSA
256
Figure 6.17: Box-and-whisker plot of Average Host Diversity (AHD) across all HER for
the EI-HCSA study.
EI-HCSA study.
257
Generational Trends This section compares and contrasts the evolved immunity algo-
rithm against the minimal generational approach. From a system perspective the evolved
immunity algorithm generally exhibited a decrease in population diversity compared to the
minimal generational algorithm, although the trend was not significant on the symmetric
exposure regime. A difference in population error between the approaches was only sig-
nificant on OHER and RHER, where a decrease and increase were observed respectively.
From a component perspective there was practically no significant difference between EI-
HCSA and MG-HCSA with regard to Average Host Diversity and Error, other than a
small increase in diversity in the evolved approach on AHER. The results demonstrate
little significant difference between the two approaches other than a general decrease in
population diversity and some changes to system error. The change in population diver-
sity provides confirmation that the evolutionary mechanisms were providing a convergence
effect, resulting in increased homogeneity between the hosts in the population.
1. Generally there is little difference between inherited immunity and independent gen-
erations for the chosen configuration.
2. Decreased population diversity may suggest at an evolutionary effect that promotes

a a population of homogeneous genetic material (evolutionary convergence).
Conclusions
Two considerations that may have resulted in the lack of significance between the two
approaches include (1) the amount of evolution permitted given the complexity of the
problem, and (2) the noisiness of the fitness function. In the first case the configuration
permitted 100 evolutionary generations each with 100 epochs providing a testing ground for
the host genetic basis. Given that this genetic basis consisted of 9600 bits, it is more than
likely that an increase in the number of generations (such as a factor of 10 or 100) would
result in a significantly increased evolutionary effect toward a homogeneous host genetic
footprint. In the second case, the objective function used (genetic system competence via
host error) may have proven too noisy given the reduced amount of interaction with the
problem (100 environment exposures). In addition, such an objective, although desirable
may have been convoluted by the asymmetry of information exposure by some of the
exposure regimes providing a potentially second level of noise. This may be tested by
again increasing the number of generations and/or the number of epochs per generation,
as well as testing on simpler symmetrical and deterministic exposure regimes.
The proposed inherited immunity approach aligned with the general progression in
complexity with the proposed host clonal selection algorithms, although in the context of
the potential of combining abstractions of evolution and the acquired immune system was
very simplistic. The principle of combining the approaches begs for elaborated investiga-
tion.
258
6.9 Chapter Summary
The Host Clonal Selection Paradigm (HCSP) was defined as the investigation of the Tissue
and Cellular Clonal Selection Paradigms as constrained by (1) multiple holistic immune
systems called Hosts and their interactions, and (2) the concerns of regimes of discrete
host exposures of information from an antigenic environment called Host Exposure Regimes
(HER). The paradigm exploits the metaphor of a collective of holistic acquired immune
systems called a Population and the turn-over of populations called Generations. Intra-
population communication is mediated by the hosts themselves as well as by pathogen
that may spread through host-to-host contact. Intra-generational communication is me-
diated by parent-progeny relationships both explicitly and implicitly via an evolutionary
process. The principle information processing interest of the paradigm as gleamed from
the metaphor are the information management strategies that may be employed to address
the known and unknown regularities and irregularities in an antigenic Environment of in-
formation called an Habitat Antigenic Exposure Problem (HAEP). The concern of such
strategies is the effective intra-population and intra-generational dissemination of informa-
tion between spatially isolated hosts and temporally isolated populations toward improved
host-wise and population-wise capabilities whilst maintaining the intrinsic localisation of
the adaptive units independence.
6.9.2 Principles and Findings

The following summarises the important principles and findings from the definition and
investigations into the Host Clonal Selection Paradigm:
General Principles
1. Host-mediated dissemination: Unlike the Tissue Paradigm that provides a natural

intra-component communication channel, the focus of the Host Paradigm is the
development of information strategies to promote the controlled dissemination of
acquired information toward improved population capability.
2. Strategies based on host-level immunology: Population immunology such as immu-

nisation and evolutionary immunology inspire information strategies for acquiring,
localisation, and disseminating information in an unknown environment as the sur-
vival of a population of organisms with immune systems rely on such strategies for
survival.
Specific Findings
1. Transmission: Elicited immunity via pathogen transmission exhibits an information

dissemination effect on exposure regimes with an asymmetric information distribu-
tion with little difference in effect whether immunity is elicited by random-pairing
based pathogen transmission or one-to-many based vaccination.
259
2. Shared Immunity: Shared immunity via effector transmission results in a disruptive
effect at the system and component level with large numbers of shared effectors,
which is minimised with small scale intra-population sharing maximises the dissem-
ination effect at the component level.
3. Maternal Immunity: Generational transmission via maternal immunity in parent-

child relationships results in an inter-generational localisation method exploiting the
generational homogeneity promoted by the mechanism.
4. Evolved Immunity: Evolved immunity via generational inheritance demonstrated

comparatively little difference over generational restarts other than a suggestion
at evolutionary convergence toward a consistent genetic footprint, suggesting at the
potential of the under investigated combination of generational learning via evolution
with lifetime learning via the acquired immune system.
6.9.3 Integration
The Tissue Clonal Selection Paradigm in Chapter 5 took the Cellular Paradigm from
Chapter 4 for granted, using it as a principle component in the Tissue Clonal Selection
Algorithm, focusing on how different decentralised information management strategies
effected the organisation acquisition and use of information under different information
exposure regimes. In the same manner the Host Paradigm in this chapter took the Tissue
Paradigm for granted, subsuming the scope of inter-tissue interaction concerns and using
the system in its entirety as a principle component (Host) in a the Host Clonal Selection Al-
gorithm. The population provided the first abstraction of multiple hosts, concerned with
decentralised information dissemination strategies, whereas the generational-population
abstraction subsumed the concerns of a given population, highlighting the related but
different concerns of inter-generational decentralised information dissemination strategies.
These two abstractions represent the pinnacle in the exploration of the clonal selection
paradigm presented in this work. The following chapter considers the aggregation of all
three paradigms (cellular, tissue, and host) in a unifying framework that provides both
explanatory power for presented algorithms, as well as predictive power in highlighting
unrealised algorithms and other facets of the considered abstractions. The abstract host
paradigm is grounded in Chapter 8 in the context of two general problem domains, high-
lighting the potential benefits the tissue algorithms provide over the cellular clonal selection
algorithms.
260
Chapter 7
Hierarchical Frameworks

The previous chapters outlined three distinct perspectives on the clonal selection adap-
tive strategy phrased as a Cellular (Chapter 4), Tissue (Chapter 5), and Host (Chapter
6) Clonal Selection Paradigms. This chapter considers all three paradigms together and
aggregates the models, algorithms, and problems into a series of hierarchical frameworks.
Specifically, Section 7.2 integrates the clonal selection models and algorithms, Section 7.3
integrates the antigenic exposure paradigm, and Section 7.4 combines the clonal selec-
tion and antigenic exposure frameworks into an integrated hierarchical clonal selection
framework. The frameworks provide both explanatory tools for the design decisions and
limitations of the investigated algorithms, as well as predictive power as to organisations
and configurations that may reveal new and interesting findings regarding the underlying
adaptive strategy. Finally, Section 7.5 considers the methodology behind the framework,
the application of the framework to the broader field of AIS, and the general applicability
of the methodology to related computational intelligence sub-fields.
7.2 Clonal Selection Framework

7.2.1 Paradigms
The three paradigms considered in this work provide tangible contributions to the field
of Artificial Immune Systems, including: (1) varied perspectives of clonal selection as an
adaptive strategy, and (2) algorithms within those perspectives inspired by the computa-
tional and information processing attributes of the clonal selection theory. The perspec-
tives of clonal selection were selected to promote a natural hierarchical relationship where
the tier that followed encompassed the concerns of the tier that proceeded. This section
considers the integration of the clonal selection models and algorithms from the Cellular,
Tissue, and Host Clonal Selection Paradigms into a Hierarchical Clonal Selection Frame-
work (HCSF). The focus of this framework is on the explicit integration of the information
processing concerns of the each of the paradigms as abstractions of the clonal selection
theory at difference scales and scopes of the acquired immune system. This integration
261
facilitates (1) inter-paradigm contrast and comparison, and (2) analysis and reasoning on
the framework providing both explanatory and predictive capabilities regarding abstrac-
tions of the clonal selection theory. The focus of the framework is on the systems side of
the system-environment relationship, and as such the concerns of the antigenic exposure
paradigm may be compressed to a generic cross-paradigm antigen exposure problem (see
Figure 7.1). This simplification of the environment directs focus on the elaboration, and
inter-relationships on the systems.
Clonal Selection
Host
Tissue Antigenic Exposure
Cellular
Figure 7.1: Depiction of the focus on systems and the compression of the environment in
the Hierarchical Clonal Selection Framework (HCSF).
7.2.2 Comparison
This section directly compares the three paradigms beyond the obvious and superficial
concerns of the perspectives and the algorithms within. Toward this end, the adaptive
systems formalism from Section 3.5.1 is used as a basis to motivate the comparison. This
formalism was selected because it explicitly provides a tool for considering the similarities
and differences between adaptive systems (paradigms), and importantly between strategies
(algorithms within a paradigm). The primary objects of the adaptive systems formalism
provide a conventional perspective on the components of an adaptive system (see Ta-
ble 3.5). From this high-level perspective the subtleties of the three paradigms are lost
and one may consider the generalities of the clonal selection adaptive strategy, specifi-
cally: the antigenic environment (e), the clonal selection strategy in cells (s), and the
affinity of cells to antigen (U ). This consideration compresses the hierarchy to the essen-
tial concerns of the adaptive strategy (for example see Figure 7.2). This is an important
perspective, as (1) it highlights the important primitives required in all so-called ‘Clonal
Selection Algorithms’, (2) provides a context for comparing the general approach to other
adaptive strategies such as genetic adaptive plans (the basis of genetic algorithms), and
(3) provides a context for relating and integrating the general approach to other Artificial
Immune System approaches (for example clonal selection as a sub-strategy in Negative
Selection and Immune Network algorithms).
The secondary objects promote the consideration of the subtler details of an adaptive
system, providing the detail of the formalism (see Table 3.6). Table 7.1 provides an inter-
pretation of the three paradigms in the context of the secondary objects of the formalism
262
Clonal Selection Antigenic Exposure
Figure 7.2: Depiction of the compression of both the system and the environment when
the HCSF is compressed to the primary objects of the adaptive system formalism.
specifically motivated by Holland’s fundamental questions listed in Table 3.7.
Cellular Tissue Host

A Cell Tissue Host
E Antigen Infections Habitats
O Interaction of cells Movement and capacity of Movement of antigen and
cells cells
S Single repertoire Discrete tissue and con- Discrete Hosts
strained connectivity
X Composition and capability Acquisition and anticipation Acquisition and anticipation
of need of need
I Selection Spatial-temporal selection Spatial-temporal selection
M Proportional resource allo- Spatial organisation Spatial organisation
cation
Table 7.1: Mapping of the cellular, tissue, and host clonal selection perspectives onto the
secondary objects of the adaptive systems formalism.
The naming convention adopted for the three tiers highlights the structures or units
under adaptation within each tier. This naming convention is mirrored in the areas of
the environment to which each structure is adapting. The consideration of the adaptive
structures highlights the already apparent subsumption of cellular concerns by the tis-
sue tier, and the tissue concerns by the host concerns. Although discussed throughout the
investigation of the paradigms, the implications of this point are less intuitive. The consid-
eration of a tissue as a structure under adaptation requires a consideration of a repertoire
of cells in its entirety as a unified structure. Adaptation of a tissue structure (T → T &)
may be considered to occur in response each infection exposure (aggregation of antigenic
exposures). This transition is less concise when one considers continuous movement of
cells via recirculation as a second transformation operator. In this context, the ‘struc-
tures’ of a single Host (tissue system) are considered in a constant state of constrained
change that is distinctly different from the adaptation of tissues provided by infection
exposure. Likewise, a similar state exists at the Host-level with the adaptive transition of
hosts (H → H&), the sharing of acquired information between adaptive units. One may
generalise this transition to say that adaptation of structures occurs with each change
at the finest level. This suggests new tissue and host structures are created with each
shared piece of information and each clonal response to antigenic exposures. The scope
(level of abstraction) of each perspective provides a frame of reference in which to choose
when to aggregate such fine-grained changes and delineate or not (continuous) structure
263
transitions. An elaboration of the concerns of each tier may consider the subsumption
of degenerate cells subsumed by the cellular tier, and the population subsumed by the
generational variation of the host tier. Although the investigation of these concerns was
tied to the cellular and host paradigms respectively, further elaboration of the sub-cellular
and generational clonal selection algorithms may result in new perspectives with as much
potential as the canonical three presented in this work (this consideration is elaborated in
Section 7.4.2).
The comparison of strategies within each perspective (X) is interesting as both the
tissue and host tiers are concerned with the effective acquisition of information in the
context of spatial and temporal anticipation of need. This assessment is based on historic
spatial-temporal selection ultimately of cells in infection and habitat exposures across the
distributed components of each class of system (I). Importantly, this information and
assessment motivate the strategies in each of these tier to (1) localise acquired knowledge
for specialised spatial-temporal anticipation, (2) dissemination acquired knowledge for
generalised system-level anticipation. Such organisation of information is directly reflected
across the aggregation of the units of adaptation at each tier and thus may be considered
a secondary memory (M ). The spatial concerns of these two tiers is compressed in the
cellular tier to that of the historic and thus anticipated temporal-properties of exposures.
As with the other tiers, the aggregation of the structures under adaptation facilitate an
additional super-structure (repertoire) in which the organisation, in this case proportional
resource allocation under varied constraints, differentiate the strategies within the tier.
7.2.3 Bracketing Analysis

An important principle in defining the three perspectives of clonal selection was high-
lighting the specific constraints imposed at each level. This principle is referred to in
this work as the bracketing of concerns after the suggestion by Goldberg in his integrated
methodology for investigating adaptive systems (see section 2.7.3). Specifically, the brack-
eting of investigated systems occurred in two ways which are discussed in this section: (1)
the bracketing of systems by the cardinality of structures under adaptation, and (2) the
bracketing of systems by the interactions of managed structures under adaptation.
Component Cardinality
The hierarchical relationship between the tiers results in the adaptive units at one tier
being comprised of an aggregation of the adaptive units from the previous tier. This
section considers the cardinality in terms of 1 and many (N ), and the bracketing of these
components within specific algorithms. Table 7.2 summarises the principle algorithms from
across the three tiers and their component cardinality in the full context of the previously
defined Cells (C), Tissues (T ), Hosts (H), and Populations (P ). An additional component
is defined below the cellular level called a sub-cellular component (c) that is used to
differentiate a degenerate component in the context of degenerate cellular algorithms.
The table reveals much about (1) the specific component cardinality selected for the
investigated classes of algorithms, and more importantly (2) suggests at the possibility
264
Paradigm Algorithm c C T H P
Cellular DCCSA N 1 1 1 1
CCSA N N 1 1 1
MCCSA 1 1 1 1 1
∗CCSA 1 N 1 1 1
Tissue TCSA N N N 1 1
∗TCSA 1 N N 1 1
Host PHCSA N N N N 1
∗PHCSA 1 N 1 N 1
GHCSA N N N N N
∗GHCSA 1 N 1 N N
Table 7.2: Summary of the component cardinality across the principle abstract and realised
clonal selection algorithms, where ∗ represents a wild-card for descendant algorithms of a
given base algorithm.
space of viable permutations of component cardinality across the hierarchical clonal se-
lection framework. From a cellular perspective the Table clearly shows the trade-off in
cardinality of the simplified minimal algorithm as a realisation of a (1+1) MHCA compared
to degenerate and cellular algorithms. The explicit aggregation of degenerate components
c into holistic cells for assessment is distinct compared to the less restricted CCSA, and the
bracketed range of assessed cellular algorithms in Chapter 4. Strictly, the c was bracketed
at c = 3 for the three colour components for investigated cellular algorithms and cellular
primitives in later tiers. The table records this as c = 1 as there was no choice for selecting
among the three c for each cell, the ordering (selection or aggregation) was specifically
defined and fixed, therefore treating all three components as a single component. The
investigation of tissue and host algorithms clearly show this same choice in bracketing,
enforcing the aggregation and thus holistic adaptation of colour space patterns. The algo-
rithms of the host tier also show an important bracketing of the tissue-level components.
This choice of bracketing was made explicit to remove the spatial acquisition and anticipa-
tion of information effects that are the concern of the tissue paradigm, and compress them
to the simpler terms of the cellular paradigm. This is reflected in the cardinality table,
clearly showing each host in the population of investigated host algorithms as a single
tissue repertoire of cells. An artefact of this choice in bracketing is that the investigation
of host algorithms did not reveal any information on the compounded concurrent effects of
spatial acquisition and anticipation of information in a population and within individual
hosts.
This last point highlights the reason for bracketing as a tool for investigating adap-
tive systems, specifically to bound the complexity of an adaptive model to the areas of
interest for an investigation. From this perspective, one may consider the cardinality of
all algorithms irrespective of their categorising scope and consider the investigation clonal
selection over the last three chapters as the relaxation of cardinality from the sub-cellular
to the population level. This is evident if one considers the shift in the N (the shift in fo-
cus) from the left to the right of adaptive structures as the list is traversed top-to-bottom
in Table 7.2. The power of bracketing the cardinality of the system is apparent when
265
one considers the alternative of investigating clonal selection at the three selected levels of
abstraction from the perspective of a population or generational host clonal selection algo-
rithm (see Figure 7.3). Such unbounded cardinality would likely make assessing principle
behaviours difficult given the confounded effects of the delegation of responsibility through
the hierarchy. This final point suggests at the focus of the Hierarchical Clonal Selection
Framework, which is: the investigation and predictable application of an integrated hier-
archical system motivated by the emergent effects of a bottom-up clonal selection adaptive
strategy.
Population of Hosts
Tissue of Cells
Host of Tissues
10010101010
Cell of bits
Figure 7.3: Depiction of an expanded realisation of a hierarchical clonal selection system

with all three tiers.
Component Interaction
This section provides a complement to the analysis of cardinality bracketing by considering

the bracketing of the interactions of the structures under adaptation across the investigated
paradigms and algorithms. Interaction refers to direct competition or pattern recognition
between structures (such as in the cellular paradigm), or the communication of information
between structures (such as in the tissue and host paradigms). Bracketing of interaction
refers to the specific constraints imposed on the interaction between structures at different
scales within clonal selection algorithms. Table 7.3 provides a summary of which com-
ponents interact across the algorithms of the three paradigms. The general algorithms
for each paradigm, specifically CCSA, DCCSA, TCSA, PHCSA, and GHCSA explicitly
list interactions for all structures, whereas specialised algorithms list interactions only for
relevant structures.
Naturally, the suggested and adopted bracketing of interactions mirrors the bracketing
of cardinality. Restricting cardinality was designed to directly restrict the effects of inter-
action, and the elaboration of the cardinality was designed to directly provide a context for
investigation. An important point implicit in the bracketing of cardinality and explicit in
the interactions is that cells and their interactions are required (almost) universally by all
clonal selection algorithms, with the DCCSA as the single exception. This is because the
structures under adaptation at higher levels of abstraction ultimately all subsume cells,
specifically receptors on and excreted by cells. Beyond the verification of the primary
objects of the adaptive systems formalism discussed in Section 7.2.2, this reinforces the
266
Paradigm Algorithm c C T H P
Cellular DCCSA yes no no no no
CCSA yes yes no no no
MCCSA no no
∗CCSA no yes
Tissue TCSA yes yes yes no no
MTCSA yes no
∗TCSA yes yes
Host PHCSA yes yes yes yes no
MP-HCSA yes no no
∗PHCSA yes no yes
GHCSA yes yes yes yes yes
MG-HCSA yes no no no
∗GHCSA yes no no yes
Table 7.3: Summary of the component interaction across the abstract and realised clonal
selection algorithms, where ∗ represents a wild-card for descendant algorithms of a given
base algorithm.
fact that both the clonal selection theory and the inspired computational principles are
defined in a cellular context. The implications of this fact reinforces the ‘perspectives on
the clonal selection adaptive strategy’ consideration of the three investigated paradigms
in this work. The interactions between structures was defined in Section 7.2.2 as the oper-
ators providing the defining characteristic between strategies within each paradigm. This
is highlighted in the naming convention adopted where the Minimum of each algorithm
class is defined with no interaction between the structures under adaptation, providing a
baseline for comparison of the effects of such interactions.
7.2.4 Summary
The integration and analysis of the three clonal selection paradigms with a compressed
antigenic exposure paradigm in the HCSF provided a context for both (1) retrospectively
explaining generalised design and methodology principles for clonal selection algorithms,
and (2) promoting consideration of permutations of cardinality and interaction that al-
though are predicted by the framework were not investigated in this work. This section
reviews the explanatory and predictive power of the ‘system biased’ Hierarchical Clonal
Selection Framework, as follows:
1. Explanatory Power
(a) Cellular Basis: Irrespective of the specific perspective taken (level of abstrac-
tion), the clonal selection theory and computational principles is a bottom-up
pattern recognition-based adaptive strategy involving cells (receptors or detec-
tors) and antigen (complementary structures).
(b) Bracketing: The forward and backward bracketing of the cardinality in the
context of a hierarchical complex adaptive system has been demonstrated to
provide a powerful tool for investigating such a system.
267
(c) Naming Conventions: A perspective invariant naming convention promotes the
naming of the perspective for the structures under adaptation, and the design
of a minimal strategy of bracketed interactions between such structures as a
baseline for measure the effects of such interactions.
2. Predictive Power
(a) Varied Bracketing: An analysis of the specific bracketing considered in the

investigation highlighted the bracketing that was not considered, as well as
the removal of bracketing all together to consider a broader hierarchical clonal
selection system, providing a rich basis for future investigation.
(b) Compound Effects: The investigation and application of an unbracketed clonal
hierarchical selection based system provides an opportunity for identifying and
assessing the compound and expected non-linear aggregate effects of cross-
component and cross-tier interactions.
(c) Alternative Perspectives: The consideration of the three paradigms in aggrega-
tion provided a context to consider (1) the potential splitting of the Cellular
and Host paradigms into new and/or sub-paradigms, and (2) suggested at the
potential for perspectives of clonal selection not yet considered.
7.3 Antigenic Exposure Framework

7.3.1 Paradigms
The antigenic exposure paradigm was defined as a complement to the clonal selection strat-
egy, and elaborated in scale along with the adaptive systems across the three paradigms.
Although the clonal selection abstractions had a basis in reviewed immunophysiology,
the antigenic environments were contrived from the general properties of endogenous and
exogenous antigen. The reason for this was that although an antigenic environment is
required to stimulate a clonal selection system, it was not the focus of the investigation.
This section integrates all of the scales of the antigenic exposure paradigm investigated
in this work from across the three paradigms, specifically: antigen (and determinants),
antigenic infections, and antigenic habitats. The integration of these concerns is referred
to as the Hierarchical Antigenic Exposure Framework (HAEF). As with the hierarchical
clonal selection framework, the focus of this integrated framework is on the explanatory
and predictive power provided via an analysis of the constraints imposed (referred to as
bracketing) at each level in the hierarchy. Unlike the clonal selection framework, this
section is not concerned with comparing and contrasting the tiers in the context of an
adaptive systems formalism. The reason for this is that the antigenic environment is not
an adaptive system, rather a required component of the clonal selection adaptive system.
As such, the perspective on the antigenic environments is elaborated, compressing the
concerns of clonal selection to provide an inverse perspective as to that considered in the
integration of clonal selection (Figure 7.4)
268
Antigenic Exposure
Habitat
Clonal Selection Infection
Antigen
Figure 7.4: Depiction of the focus on environment and the compression of the systems in
the Hierarchical Antigenic Exposure Framework (HAEF).

Bracketing in the antigenic exposure paradigm involved constraints imposed on the cardi-
nality of the triggers for adaptation and the regimes that governed the piece-wise exposure
of environmental information to a given system. This section considers the bracketing of
configurations imposed on these two antigenic environment concerns.
Component Cardinality
The bracketing of the cardinality of the information in the environment was configured
to match the scale of the units of adaptation in the adaptive systems. Unlike clonal se-
lection, the antigenic environment was not motivated in its interaction with the systems
(for example adversarial, dynamic, and/or adaptive). The neutrality of the environment
was intentional to focus attention on the emergent system behaviour under low-complexity
environment concerns. Table 7.4 summarises the cardinality of the abstract antigenic envi-
ronments and the realised colour space antigenic environments in terms of the previously
defined: determinants (D), antigen (A), infections (I), habitats (B) and environments
(E).
Paradigm Problem D A I B E
Cellular AEP N N 1 1 1
ACSP 1 N 1 1 1
DCSP N N 1 1 1
Tissue IAEP N N N 1 1
ICSP 1 1 N 1 1
Host HAEP N N N N 1
HCSP 1 1 1 N 1
Table 7.4: Summary of the component cardinality across the principle abstract and realised
antigenic exposure problems.
The table shows the potential of each abstract exposure problem and the strong brack-
eting of the colour space realisations for the structures under adaptation in the systems in
each tier of the hierarchy. The ACSP, ICSP, and HCSP all exploited the same bracketing
269
method, specifically the treating of each Colour Space Pattern (colour) as a trigger for
adaptation at the respective tier. For example, a set of antigen (ACSP) is equivalent to
a set of infections (ICSP) and a set of habitats (HCSP), in that they can all be reduced
to a set of Colour Space Patterns. This equivalence in problem was designed to min-
imise the complexity of the information to be acquired across the tiers to a level that was
easy to interpret, whilst introducing complexity the way in which the information was
exposed and expected through varied exposure regimes. This simplification of problem
motivated the corresponding cross-cutting cellular-based clonal selection concerns in the
systems by reducing all information to single antigen. One may consider the alternative
of such a minimisation in a less constrained colour space realisations of the the expo-
sure problems. For example, an infection of antigen may involve sets of antigen that are
similar (intra-infection), although distinctly different between the sets (inter-infection).
This could be realised through minor variations for intra-infection colour space patterns
(similar colour), and major variations for inter-infection Colour Space Pattern centroids
(inter-infection). This same example scales to the habitat level, with sets of sets of antigen,
where intra-habitat infections may be more similar than inter-habitat infections. These
examples highlight what the bracketed cardinality provided, specifically exemplar pat-
terns (set size of one with no variations) for each infection and habitat with promoted
inter-trigger dissimilarity (minimum hamming distance). This chosen level of cardinality
bracketing had the effect of removing the degeneracy, (specifically the redundancy of the
triggers of adaptation) requiring an immune response for each antigenic trigger in the
antigenic environments. The alternatives provide an opportunity for re-introducing such
degeneracy although at the cost of the specificity of a system in an environment (gener-
alised response against variations for each antigenic determinant at the lowest level).
Exposure Regimes
Antigenic exposure controlled the piece-wise revealment of information to the systems as

well as the piece-wise spatial and temporal expectation of future exposure. In the context
of forced response to a fixed information environment, the exposure regimes managed the
complexity of the problems to be addressed by the systems. The complexity of exposure
scaled with the increase in the level of abstraction starting with exposures, multiple expo-
sures and multiple antigen in the definition of the exposure paradigm (Section 4.2.2), and
the raise in the level of abstraction of such exposures to include spatial concerns across
multiple points of exposure in infection exposures (Section 5.3.2) and habitat exposures
(Section 6.3.2). This section considers the bracketing of exposures with regard to (1) the
spatial points of exposure, and (2) the temporal concerns of the expectation of exposure.
Table 7.5 considers the bracketing of the spatial concerns of an exposure, in particular
the triggers for adaptation in terms of the scope of the points of exposure in a system (1 or
many). A point of exposure in a system is unit of adaptation, therefore given the chosen
cardinality bracketing in both systems and environments, there is always a many-to-many
relationship between the units and triggers of adaptation at the focused level of detail
in the hierarchy. At the antigen level, the scope of the system (all points of exposure) is
270
ACSP 1 N 1 1 1
DCSP N N 1 1 1
Tissue IAEP N N 1-N 1 1
ICSP 1 N 1-N 1 1
Host HAEP N N 1 1-N 1
HCSP 1 N 1 1-N 1
Table 7.5: Summary of the bracketing of spatial exposure across the abstract and realised
antigenic exposure problems.
always exposed. This has the effect of requiring a mechanism to constrain a polyclonal
activation (shown to be strong selection in Section 4.4.3). The reason for this design
decision was the relative small scale of a single repertoire of cells in the context of multiple
of such repertoires (tissues) and multiple such hosts of tissues. The same strong selection
was required at the determinant level, and given the antigen focus at all levels, this had the
effect of requiring explicit aggregation of response toward holistic solution to address the
resultant polyclonal response. This unbounded antigen-cell exposure and explicit strong
selection and aggregation is bracketed in the tissue and host levels in the introduction
of discrete Antigenic Exposure Regimes (AER) providing deterministic and probabilistic
patterns of interaction between the units and triggers of adaptation (referred to as 1 − N
in the table). This had the effect of requiring the systems under such regimes to manage
the spatial organisation (localisation and dissemination) of acquired information to best
address the patterned interaction. One may consider the alternatives of such bracketing
of spatial exposure, for example the spatial selection of cells in a repertoire considered
in spatial up-front selection in Section 4.5. The symmetrical exposure regime at the
tissue (STER) and at the host (SHER) levels provide a counter example of the effects
of simplifying exposure complexity to that of system-wide with respect to the points of
exposure of a system.
ACSP N N 1 1 1
DCSP N N 1 1 1
Tissue IAEP N N N 1 1
ICSP N N N 1 1
Host HAEP N N N N 1
HCSP N N N N 1
Table 7.6: Summary of the bracketing of temporal exposure across the principle abstract
and realised antigenic exposure problems.
Spatial exposures without a time component are a compression of the patterns of

exposure over the scope of the units of adaptation. Therefore, temporal exposures can
be considered an opposite case of the compression of spatial location and a focus on
271
the systems expectation of information on a timeline. Table 7.6 summarises the temporal
expectations of the abstract and realised exposure problems with regard to the triggers for
adaptation. The table clearly shows that the temporal expectations of exposure extended
backward through the hierarchy from the focus of each tier. This simple but important
perspective on the antigenic exposure paradigm highlights the motivating complexity for
the corresponding Hierarchical Clonal Selection Framework, specifically the compounded
increase in temporal expectation of information. This table also highlights the need for
bracketing the information cardinality of antigenic environments across the hierarchy to
bound the complexity of the effects of the compounded temporal expectations.
7.3.3 Summary
The integration and analysis of the antigenic exposure paradigm across the three levels
of abstraction in the context of a compressed clonal selection system provided a context
for both (1) retrospectively explaining generalised design and methodology principles for
the of the problems that motivated the investigation of clonal selection algorithms, and
(2) promoted the consideration of permutations and relaxation of cardinality, spatial,
and temporal concerns of antigenic exposure. This section reviews the explanatory and
predictive power of the Hierarchical Antigenic Exposure Framework, as follows:
(a) Forced Response: The strong bracketing of the cardinality of antigen across the
hierarchy resulted in a forced response expectation from systems with regard to
the trigger for adaptation. Such forced response may be relaxed by loosening the
bracketed cardinality constraints at a given tier in the hierarchy, in particular
for the infection and habitat levels.
(b) Compounded Expectation: The focus of the investigation using the antigenic
exposure paradigm was on compounded expectation with varied spatial scope
of exposure at the tier of interest, and strong bracketing of the spatial scope at
tiers above and below the tier of interest.
2. Predictive Power
(a) Cardinality Complexity: An alternative to the design of clonal selection sys-

tems in response to exposure complexity, is that the design of such systems
can be motivated by variations in the cardinality complexity instead and/or
combinations of cardinality and exposure complexity. Specific examples of car-
dinality complexity include sets of antigen in each antigenic infection, and sets
of infections in each antigenic habitat.
(b) Compounded Spatial Scope: The spatial scope of exposure was constrained at
tiers of the hierarchy above and below a focused level, the relaxation of this
bracketed spatial exposure will provide an additional axis of problem complexity
to explore in addition to compounded expectation.
272
7.4 Integrated Hierarchical Framework
7.4.1 Frameworks
Given an assessment of integrated hierarchical clonal selection and antigenic exposure
frameworks, this section considers the aggregation of the two frameworks towards a unified
abstraction and framework for clonal selection in the context of an antigenic environment.
This is a natural integration as the abstract and realised problems and algorithms in both
sides of the integration were designed as a natural fit for each other. This integration is
referred to as the Integrated Hierarchical Clonal Selection Framework (IHCSA) to differ-
entiate it from the concerns of either independent framework, and to highlight the strong
bottom-up clonal selection focus. As such, the focus of the framework is the hierarchical
integration of the elaborated concerns of both the system and the environment, depicted
in Figure 7.5
Host Habitat
Tissue Infection
Cellular Antigen
Figure 7.5: Depiction of the focus on system and environment in the Integrated Hierar-
chical Clonal Selection Framework (IHCSF).

This section focuses on the bracketing of concerns that have an effect both on the system
and environment hierarchies. Specifically, this section is concerned with the matching or
integrated cardinality across the hierarchy, and the effect of variations of this matching.
Also considered in this section is the extension of the framework through the integration
of additional tiers to the existing framework.
Integrated Cardinality
The focus of this subsection is the bracketing of the cardinality of both the system and
environment hierarchies as they pertain to the interaction between the hierarchies. Specif-
ically, the interaction of the triggers for adaptation with the structures under adaptation
at both the scope of their intended focus (for example cells and antigen in the cellular
level), and scopes above and below that focus.
Table 7.7 summarises the intended or so-called ideal integrated cardinality cases for
the environmental triggers and system units of adaptation. The table clearly shows the
273
c C T H P
D 1 N N N N
A 1 N N N
I 1 N N
B 1 N
E 1
Table 7.7: Summary of the suitability match between clonal selection and antigenic expo-
sure components
response suitability in terms of the one-to-one relationships (upper limit) and the general
capability in terms of the one-to-many relationship of system structures to environmental
triggers. It should be pointed out that the upper limit (one-to-one) of the suitability
between the frameworks was not investigated throughout this work as an initial assessment
may suggest. For example the cellular paradigm considered a ‘tissue of cells’ against an
‘infection of antigen’ rather than a cell against an antigen or the mismatch case of a
‘tissue of cells’ against a single antigen. Table 7.8 elaborates on the ideal cases, and
summarises the components of each hierarchy and indicates the scope of concerns of each
paradigm, and the cases that were embodied in algorithms that were investigated (bold).
Paradigms are marked with the first letter of the structures under adaptation, with the
division of Cellular into Degenerate (D) and Cellular (C), and Host into Host (H) and
Generational (G). Perspectives or tiers are defined by the interaction of the designated
triggers and structures for each paradigm, for example cells (C) and antigen (A) in the
cellular paradigm. The italicised cases demonstrate those cases that could be addressed
from tiers above and below a given tier.
c C T H P
1 N 1 N 1 N 1 N 1 N
D 1 D D
N D C C
A 1 D C C
N C T T
I 1 C T T
N T H H
B 1 T H H
N H G G
E 1 H G G
N G
Table 7.8: Summary of the component cardinality for both the clonal selection and anti-
genic exposure models.
An important point highlighted by this table are the overlapping concerns between the
paradigms. For example N cells may be regarded as a single tissue (minimal cardinality
tissue model), both of which can address N antigen or 1 infection (minimum cardinality
infection model). The designed subsumption inherent in both independent and the inte-
274
grated frameworks promote this maximum-minimum principle at the transitions between
tiers in the hierarchy. This principle allows similar concerns to be investigated from two
different perspectives. For example in the cellular-tissue case, one can investigate the
the systems composition and capability of a repertoire in response to a set of N antigen
with explicit aggregation of the best matching cells for each antigen from the cellular per-
spective, or the response capabilities of a tissue against a single antigen from the tissue
perspective. Both examples provide a different context (different level of abstraction with
regard to the acquired immune system) that may provide varied strategies for considering
the same problem. The table clearly shows such cases were investigated using the maxi-
mum of the previous tier rather than the minimum of the given tier. The highlighting of
investigated cases clearly shows the scope of future work with regard to integrated cardi-
nality cases. An interesting point is that the generational case did not follow the trend
and investigate multiple environments with multiple populations. Rather, the EI-HCSA
adapted multiple populations against a single antigenic environment (set of habitats). The
table also provides a clear indication of those cases delineated to specific scopes of interest
that were not investigated, and those cases that could not be investigated. For example,
the maximum trigger for a given tier (N ) with the minimum structures of adaptation (1)
is an infeasible combination defining the upper limit on cardinality complexity for each
system.
Figure 7.6 depicts the integrated cardinality bracketing in a hierarchical way to en-
force the subsumed responsibility across the tiers, providing a standard representation of
the frameworks and basis for extrapolating from the integrated relationships. The fig-
ure is information dense capturing intra- and inter-tier relationships for both the system
and environment frameworks, and clearly highlighting the important maximum-minimum
principle at tier-boundaries (diagonal lines). An interesting observation is that of the
relationship between the cells and antigen, showing that a cell and a receptor do not
differ in terms of their information content, and that a cells receptor is specialised for a
given determinant on an antigen (base principles of the clonal selection theory). This is
important given the compression of this relationship (bracketing) of one-antigen-to-one-
determinant (and thus one receptor or cell) throughout the majority of the investigations.
Only the DCCSA considered the relaxation of this bracketing, allowing cells (receptors)
to be selected for by the determinants of an antigen. This specialisation was classified
as apart of the cellular paradigm, although if such specialisations are elaborated and in-
vestigated (specifically the one-to-many relationship between an antigen and receptors as
in the DCCSA in Section 4.4.3), one may refer to this as the Receptor Clonal Selection
Paradigm (RCSP) to clearly distinguish it from the cellular paradigm. The figure also
suggests at tiers above and below the the host and tissue tiers respectively, outside of the
scope of what was investigated. These extrapolated additional tiers are considered in the
next section.
275
Species n-Populations 1 1 n-Environments

Species Niche
1
1 1
n
n n
Host n-Hosts 1 1 n-Habitats
Population Environment
1
1 1
n
n n
Tissue n-Tissues 1 1 n-Infections
Host Habitat
1
1 1
n
n n
Cellul ar n-Cells 1 1 n-Antigen
Tissue Infection
1
1 1
n
n n
n-Substructures 1 1 n-Determinants
Cell Antigen
1 1
1
1 n
Receptor 1 1 n-Substructures
Receptor Determinant
1
1 1
n
n n
Paratope 1 1
Substructure Substructure
Figure 7.6: Depiction of the integrated hierarchical clonal selection and hierarchical anti-
genic exposure frameworks, showing subsumed relationships and component cardinalities
across the investigated paradigms.
Additional Tiers
A point stressed in the integration of the three investigated clonal selection paradigms
in this Chapter has been that each represents a single perspective on the application of
clonal selection as an adaptive strategy. Explicitly selected in the standard form depiction
of the integration in Figure 7.6 was the capability for additional vertical tiers above and
below the host and cellular tiers respectively, as well as the potential for integration of
additional perspectives of a given tier. This section extrapolates examples of such addi-
tional perspectives on clonal selection and integrates them into the hierarchical framework
of such perspectives.
Population Clonal Selection A perspective beyond that of a population of hosts and

an environment of habitats may be that of a species and an ecological niche, where a
a set of populations interacts with a set of environments called the Population Clonal
Selection Paradigm (PCSP). Like the generational variation of the host paradigm, the
population paradigm involves maintenance of populations as the structures of adaptation,
although the separation between the populations is spatial rather than temporal. As a tier
276
beyond the host level, it subsumes the concerns of the host level including the generational
turn-over, introducing inter-population interaction concerns such as those from population
genetics (such as migration), and epidemiology (such as the spread of disease).
Paratope Clonal Selection Given that the clonal selection theory describes the obser-
vations of immune cells, a level below the cellular level may not be a reasonable proposition.
One may consider the development of a single receptor of sub-structures for a single anti-
genic determinant of sub-structures as the maximum in complexity of this preceding tier.
The scale of the process is that of the tertiary confirmation of the immunoglobulin proteins
and related immunochemical forces of their interaction with complementary structures of
the determinants on macro-molecules and proteins, governed by the sequence amino acids
that make up both molecules. This tier is referred to as the Paratope Clonal Selection
Paradigm (ACSP) 1 and may be concerned with the specific chemical basis for immunoglob-
ulin synthesis in cells, the valance of immunoglobulin molecules, and the physical forces
that govern the molecular interactions and resultant affinity and avidity for determinants.
7.4.3 Summary
This final hierarchical integration represents the pinnacle of the abstractions presented.
The integration and analysis of the integration framework provided a context for both (1)
retrospectively explaining generalised design and methodology principles that motivated
the investigation of clonal selection algorithms, and (2) promoted the consideration of un-
realised permutations and relaxation of cardinality and additional tiers of the framework.
This section reviews the explanatory and predictive power of the Hierarchical Integrated
Clonal Selection Framework, as follows:
(a) Framework as a Map: The integrated framework facilitates the explicit spec-
ification of immunological and antigenic concerns in terms of their scope of
influence relative to each other.
(b) Maximum-Minimum: Tier transitions are defined by the maximum cardinality
complexity of one tier that is reduced to the minimum cardinality complexity of
the next tier. Such boundaries allow the same information content (problem) to
be addressed from the perspectives of the tiers above and below the boundary,
revealing insights and alternative solution strategies.
(c) Simplest Case: The investigations considered the average case of each paradigm,
specifically the many-to-many relationship between triggers for adaptation and
units of selection and adaptation.
2. Predictive Power
(a) Varied Interpretation: The backward competence of a paradigm with its sub-
sumed levels suggests at non-contigious interpreted cases of cardinality and
1
Paratope was chosen as the term for a sub-structure of a receptor molecule for lack of a better term.
277
interaction along both the system and environmental axis. The realisation
of such models facilitate the use of tier-specific mechanisms (algorithms) on
backwardly-compatible (subsumed) problem domains.
(b) Additional Tiers: The integration of the frameworks facilitated the elaboration
of the cellular and host tiers to include concerns beyond the scope of these two
limits of the framework.
(c) Additional Mapping: The extensibility of the framework suggests at the po-
tential for mapping all manner of immunological and antigenic concerns into
the appropriate scope of the framework and defining relationships in terms of
cardinality and interaction of structures and triggers. For example, the Nega-
tive Selection, Danger Theory, and Immune Network Artificial Immune System
paradigms.
7.5 Methodology and Hierarchical Framework

The integrated Hierarchical Clonal Selection Framework was not conceived independently
of the tiers of the framework, nor were the tiers conceived independently of each other,
rather they were systematically co-developed. This section considers (1) an overview of
the methodology used to develop and investigate the clonal selection framework, (2) the
application of the framework as a template for investigation in the broader field of Artificial
Immune Systems, and (3) the framework as a template for investigation in the separate
but related Computational Intelligence sub-fields.
7.5.1 Partition-Reduction Methodology

Methodology Overview
In considering the applicability of the chosen hierarchy for clonal selection and the method-
ology for applying such a methodology, it is critical that the motivation and adopted
methodology be clearly elucidated. The general systems theoretic approach used involved
three important considerations (1) the distillation of the core computational principles
and strategy from the clonal selection theory, (2) the division of the clonal selection con-
cerns into that of system and environment, and (3) the reduction of the holistic acquired
immune system to that of the cellular concerns of the clonal selection theory. These steps
are summarised as follows:
1. Metaphor : Identify the relevant observations, theory, and abstractions of the selected
metaphor.
2. Strategy: Distil the metaphor to the core computational principles and information
processing strategy.
3. Partition: Divide the concerns from the principles and strategy into a framework of
systems and environments.
278
4. Reduction: Relate the components and their interactions at the lowest and highest
levels (within scope of the metaphor).
The distillation of clonal selection theory resulted in an adaptive strategy concerned

with information acquisition, the concerns of which were partitioned into the immune sys-
tem and the antigenic environment. The reduction from immune system to the cellular
concerns of the theory focused on the three different and inter-related functional organisa-
tions of cells under clonal selection: the simplest case of antigens and cells in a small group
(cellular ), many groups of cells using structural concerns to influence antigenic interac-
tions (tissue), and groups of immune systems influencing the transmission of disease and
evolving (host). The selection of the levels in the reduction was also generally motivated
based on the units of selection consideration from selectionist theory, providing a cellular
basis in the reduction given that the lymphocyte is the unit of selection and adaptation
in the clonal selection theory. Importantly, the general functional reduction was taken a
step further by considering the cardinality of the units of selection and their interaction
within each tier, and focusing on the limits of each.
Relation to Methodologies
The adopted methodology is comprised of concerns of the three reviewed approaches for
investigating biologically inspired computation, immune information processing, and adap-
tive systems in Section 2.7. This section briefly highlights the selected aspects of the mo-
tivating methodologies and how they contributed to the integrated adopted methodology.
The general approach was a realisation of the five-step conceptual framework from
biologicall-inspired algorithms (2.7.1). Specifically (1) the selection of the immune sys-
tem as the motivating metaphor, (2) the review of immunochemistry and immunophys-
iology (probes) as a basis for information processing principles, (3) the investigation of
models as abstraction of the information processing properties from models, (4) the ap-
plication and verification of the computational models (considered next in Chapter 8),
and (5) resulting computational tools. The adopted methodology drew heavily from this
approach, in particular with regard to the general process for realising viable computa-
tional tools, and with regard to the strong focus on the transition from probes (metaphor )
to information processing models that were investigated (results of the reductions). The
immune information processing approach (Section 2.7.1) provided a specific context to
motivate the conceptual framework, in particular to focus on the information processing
concerns of immunology. Goldberg’s so-called small models methodology for adaptive sys-
tems (Section 2.7.3) provided a strong influence in the adopted methodology, specifically
focusing the modelling effort on the decomposition of general information processing to
separate sub-problems (reduction) and the ultimate re-integration (the IHCSF). Along
with the reductionist consideration, Goldberg’s methodology promoted a strong focus on
the bracketing of high-order phenomena which was realised in particular with regard to
the cardinality and interactions between structures. It is important to clarify that the
small models methodology anticipates system modelling is achieved through the use of
mathematical methods. The approach taken in this work involved the use of algorithmic
279
computational modelling and simulation. An additional alternative modelling approach
that could have been used is diagrammatic object and state modelling, such as the Unified
Modelling Language (UML).
Alternative Interpretations
The specific reduction was an clonal selection-specific interpretation of the immune sys-
tem based on the functional aggregation of cells at three intuitive levels of detail. This
suggests that alternative cellular-based reductions may be considered, as well as reduc-
tions not based on the units of selection and adaptation in the clonal selection theory. An
additional consideration for a reduction may include an antibody protein focus perhaps
including the concerns of other macromolecules that influence the behaviour of the im-
mune system such as cytokines and cellular receptors for such self-molecules. This would
provide a focus on the interaction of components for signalling rather than adaptation.
Alternatively, one may focus on the division of the immune system into adaptive and
innate, and focus on the structures and functions within each ‘sub-system’ promoting a
stronger consideration of the concerns of the innate immune system and its interrelated
processes with the acquired immune system. The viability of a given reduction may be
assessed based on the insights it provides regarding the computational principles under
consideration. As such, an alternative already considered is the ‘immune system ↔ cel-
lular ’ reduction common across paradigms in AIS, suggesting that the contributions of
the IHCSF may be the host and tissue considerations and the levels of comparison they
provide.
7.5.2 Hierarchical Artificial Immune Systems

The field of Artificial Immune Systems is concerned with the investigation of computational
tools inspired by the structure and function of the immune system (see Section 2.3.2).
Given the pervasiveness of molecular and cellular concerns in immunology and the follow-
on effects of this focus into Artificial Immune Systems, one may consider the application
of the proposed Integrated Hierarchical Clonal Selection Framework as a general tool
in AIS. Specifically, the hierarchical perspective may be used as a framework to focus
effort for the integration existing seeming disparate computational methods inspired by
immunology, and potentially highlighting new opportunities for investigating AIS.
Negative Selection
The negative selection paradigm exploits the clonal selection process for adaptation-based
information acquisition although focus on the specific concerns of using pattern recogni-
tion to discriminate self from nonself and the preparation of such pattern detectors (see
Section 2.4.2). The first important consideration is that of negative selection as adap-
tive strategy (modelling in the complement space with centralised preparation and with
decentralised opportunistic specialisation using clonal selection). The strategy focus of
the paradigm highlights the constraints under which the strategy must operate (satis-
fice) within the immune system at varied levels of abstraction. Such consideration may
280
provide insight to advance the paradigm beyond the current highly-constrained cellular
perspective, and the limitations of such approach in practical application [374].
Scope Negative Selection

Cellular Centralised detector generation and application.
Tissue Centralised/Decentralised detector generation and decentralised application.
Host Controlled antigenic exposure, biased restarts of environmental perspective, and
adaptation of generation and application structures and processes.
Table 7.9: Summary of the Negative Selection AIS paradigm mapped onto the three tiers
of the IHCSF.
Table 7.9 provides a summary of the mapping of the negative selection paradigm onto
the IHCSF. Given the close ties with clonal selection, the negative selection paradigm has
generally the same concerns. Specifically, effective adaptation in the cellular tier, and ef-
fective localisation and dissemination in the tissue and host tiers. Also consistent with the
clonal selection paradigm is the antigenic environment, where the information in the envi-
ronment encapsulates the scope of patterns that must be differentiated from self patterns.
The important differences between the paradigms is the strong focus on cellular prepara-
tion required in negative selection that may be achieved locally or globally. For example,
the corpus of self-patterns may be centrally managed at the cellular scope providing a
focus on effective non-self pattern generation, and the general application of such detec-
tors. At the tissue scope, the preparation of the detectors may be centralised in a thymus
model where a specific centrally connected tissue is designated the role of managing the
self-pattern corpus and detector generation, with system-wide decentralised application
and localised refinement of the recirculated detectors. Alternatively the management of
the corpus may decentralised and distributed throughout the tissues of the system for
localised (specialised) preparation and dissemination, likely a more scalable configuration,
particularly if the self corpus is accumulated in an opportunistic manner. These concerns
were considered by Hofmeyr in his dissertation (see Section 2.5.2) and provide a impor-
tant research agenda for negative selection at the tissue scope of the integrated hierarchy.
The host has similar concerns as that of the tissue, although with the important shift
in focus from a strongly integrated series of perspectives to strongly independent holistic
perspectives with weak integration. Pattern generation is centralised within each perspec-
tive although some control is provided over the exposure to information, the restarting of
perspectives with biasing, and the long-term adaptation of the structure and processes of
detector preparation and application.
Immune Network
The immune network paradigm is concerned with intra-receptor interactions with and
without the presence of antigen and was proposed to be a specialised case of clonal selec-
tion at the cellular level (Section 4.7). This paradigm was investigated (1) briefly, and (2)
only at the cellular level, suggesting an opportunity for further elaboration and integra-
tion with the scope of artificial immune network approaches. The information processing
281
strategy distilled form the theory was that of repertoire maintenance using intra-cellular
interactions. Such interactions form networks of cellular relationships which may be con-
sidered intra-repertoire higher order structures, that may be exploited for storing more
complex information.
Scope Immune Network

Cellular Formation and maintenance of intra-repertoire relationships.
Tissue Interaction and integration of inter-tissue cell-based relationships.
Host Generational refinement of structures as well as the processes and structures used to
form and maintain such structures.
Table 7.10: Summary of the Immune Network AIS paradigm mapped onto the three tiers
of the IHCSF.
Table 7.10 provides a summary of the mapping of the immune network strategy to the
three tiers of the IHCSF. The antigenic environment involves the scope of the standard
clonal selection antigenic environment at each tier in the framework, in addition to the
scope of intra-repertoire interactions that may result. As these relationships are scaled-up
from the cellular level, they may be considered to occur between tissues, and between
hosts, with the latter providing a truly abstract proposition for the theory. From the tis-
sue perspective, the localisation and dissemination of acquired information extends from
information about antigenic infections, to information about the relationships within each
tissue. The dissemination of cells involved in inter-cell relationships provides an opportu-
nity for the interaction and ultimate integration of localised relationships between tissues,
providing new contexts for response, and likely initially disrupting the maintenance of
localised relationships. The host perspective, in particular the generational specialisation
provides a context for considering the biasing (so-called ‘seeding’) of cellular relationships,
and ongoing refinement of these relationships over extended time periods across multiple
host-lineage in a population using the maternal immunity approach. The generational spe-
cialisation also provides an opportunity for adapting the processes and structures used in
the formation and maintenance of relationships, allowing potentially for the development
of new relationship types (structures).
Other
The generality of the reduction of functional immunology from the scope of a holistic
immune system to the theory-laden cellular level may provide a powerful abstraction
for many unrealised and/or budding Artificial Immune Systems. Examples may include,
danger theory, innate immunity, cytokine molecules, and the large number of cell types
and their interactions. The so-called power of the abstraction comes from (1) the context
it provides for framing the information processing qualities of an immunological structure
or process, (2) the strong focus on identifying and bracketing of complexity in resultant
models promoting clear schedule for experimentation and elaboration, and finally, (3) a
context for the integration of findings into a coherent model of the motivating paradigm.
282
7.5.3 Hierarchical Framework Methodology
This section considers the exploitation of the proposed methodology in different but related
biologically-inspired Computational Intelligence sub-fields. In considering this question,
a brief assessment of two general fields are considered, specifically (1) Evolutionary Com-
putation and (2) Connectionist Intelligence. The fields are considered in the context of
their metaphor, strategy, partitioning into system and environment, and reduction. Given
the enormity of the research into the selected Computational Intelligence sub-fields the
hierarchical phrasing of the three adaptive strategies is not likely to reveal any insights
into the respective paradigms, rather the expectation is that the reductionist methodology
may provide a framework for better integrating findings from existing approaches.
Evolutionary Computation
The metaphor for the well studied evolutionary computation field is neo-Darwinian evo-
lution, taking into account the genetic concerns of micro- and macroevolution. The dis-
tilled computational principles include selection and descent with modification typically
via the recombination and mutation of structures. The computational strategy may be
summarised as adaptation toward relative improvement under the constraints of an en-
vironment (see Section 3.4.1). The summation of the adaptive strategy highlighted the
intuitive separation of the population of structures under adaptation, and a environment
responsible for defining the scope of ‘fitness’, typically apportioned directly in terms of
quantitative scalar values that are interpreted by the adaptive strategy as the proportion-
ate allocation of available resources. Table 7.11 provides a summary of an interpreted
reduction of evolution from the species level to the DNA level, constraining both the gen-
eral adaptive strategy and fitness-assigning environment. The reduction is based on the
units of selection and adaptation using the same general approach as was adopted in the
reduction of clonal selection. The chosen units of selection include genomes, populations,
and species, which forms an Integrated Hierarchical Evolutionary Framework.
Tier Environment System

Genome Habitat of Localities. Population of Genomes, adapting in context of the scope
of information being adapted and its mapping to an as-
sessable form.
Population Niche of Habitats. Species of Populations, localising and disseminating
adapted genomes.
Species Ecology of Niches. Ecosystem of Species, adapting and competing for ecolog-
ical niches.
Table 7.11: Summary of an interpretation of the field of Evolutionary Computation in

terms of the hierarchical reductionist methodology used to investigate clonal selection.
The genome level encompasses the scope of standard evolutionary approaches, and
is bounded to a single population of individuals each with a distinct genome. The en-
vironment at this level is concerned with the relative difference between members of the
population based on the information content of their genome. It is concerned with the
283
tools of DNA, information decomposition from holistic genome to functional gene, and
ultimately to the base pairs of the representation. This is an important consideration as it
provides a bracketing of information representation complexity, including but not limited
to the concerns of the scope of the information mapped and assessable in the phenotype,
intra-gene relationships (gene or allele linkage), and other representation complexities.
The mapping of genotype to phenotype provides another critical axis of complexity for
bracketing including concerns of development and time-scope of assessment. As such the
mapping may vary between direct assessment of genome, unicellular and multicellular or-
ganisms, to the full concerns of developmental biology and lifetime-based assessment of
the resulting organism. To summarise, the important axis for bracketing include: (1) in-
formation representation such as one-to-many relationships in base-pairs, genes, and the
number of and number of sets of chromosomes (ploidy); and (2) information mapping such
as the transformation and time-horizon on the assessment of the genotype and potentially
resulting organism.
The population level shifts the concerns from that of organisms in a population, to a
set of discrete populations in a species. The environment is scaled in turn, considering
the relative improvements of populations through the aggregation of the relative improve-
ments of the genomes in each population. The environment is generally the same for
all populations, limiting the divergence of populations into different species. This level is
reminiscent of hierarchical genetic algorithms (for example see [319, 119]) and parallel evo-
lutionary algorithms (for example [68]), with concerns of inter-population interaction in
the form of migrations. Examples of inter-population interactions provides an important
axis for complexity for bracketing beyond the cardinality of genomes in each repertoire.
This constraint is lifted in the species level that raises the concerns from that of popu-
lations in a species, to that of a species in an ecosystem. The environment provides a
variety of ecological niches that may be exploited and competed for by the species in the
ecosystem. This introduces the concerns of the divergence of species called speciation (for
example see speciation genetic algorithms [123]), and the invasion of sudden exploitation
of underexploited ecological niches adaptive radiations (for example see niching genetic
algorithms [279]). Both the interaction between species, and the cardinality of ecological
niches provide contexts of complexity that may be bracketed for investigation, with inter-
esting combinations such as species themselves representing a part of an ecological niche
in a predator-prey relationship.
The hierarchical framing and relationships between the concerns of evolution highlight
the specific bracketed configurations chosen and propagated for sub-fields of the paradigm.
For example in the case of niching genetic algorithms and speciation genetic algorithms
the effects are realised within a single population. This observation does not violate the
constraints of the proposed framework, rather suggests a cross-hierarchical bracketing
where the diversity concerns of multiple populations or multiple species are compressed to
that of genomes within a population. This form of bracketing is similar to that used in the
HAEF with antigen from the cellular to the host levels, although in this case is used on the
system side to bring features from the higher levels of abstraction down to the lowest level
284
of abstraction by compressing the cardinality of the units of adaptation and selection.
The compression of cardinality is the basis of population genetics that considers gene
frequencies, flows and drifts in populations and species of individuals. The contribution of
the framework is subtle, as it does not suggest perspectives of the strategy unconsidered
by the field. Rather, the contribution is that the focus of each tier provides a context for
bracketing the complexity of the strategy at different levels of abstraction. This focused
bracketing highlights the axis of complexity that may be bounded for investigation, and
more importantly it suggests at the effects of the constraint by considering the extremes
of the alternatives.
Connectionist Intelligence
The metaphor for the field of connectionist intelligence are brain cells called neurons that
ultimately provide the biological basis of human-level intelligence when arranged into
vast interconnected networks in the central nervous system. The distilled computational
principles include collections of neurons with synapses that connect the neurons into a
network structure, and the propagation of signals through the network. Propagation of
signals involves sensory (input) neurons which propagate received signals through the
network in parallel. Signal propagation is governed via the aggregation of inbound signals
to a given neuron in the network which may or may not activate and transfer (froward
propagate) the received signal. The computational strategy involves the adaptation of
the ways signals are propagated via adjusting the weights between the neurons. The
strategy is divided into the central nervous system of neurons and their interconnections
and the environment that provides the stimuli to which the system receives and adjusts
its behaviour. Table 7.12 provides a summary of an interpreted reduction of connectionist
intelligence from the holistic nervous system to the neuron level, constraining both the
general adaptive strategy and the sensory input environment. The reduction is based on
levels of abstraction regarding the aggregations of brain cells which provide interesting
constraints on the information processing that may occur including a network of cells
referred to as a neuronal group that responds to a habitat of stimulus, groups of networked
cells referred to as a host (nervous system), and finally a population of hosts with nervous
systems.
Tier Environment System

Cellular Stimulus of Inputs. Group of Cells, providing a focus on the organisation and
interaction of neuronal cells.
Group Habitat of Stimulus. Host of Groups, focusing on the interaction between
groups of cells such as the dissemination of information.
An alternative name for a group may be a cortex.
Host Environment of Habitat. Population of Hosts, focusing on the interaction of hosts
each with their own nervous system.
Table 7.12: Summary of an interpretation of the field of Connectionist Intelligence in

terms of the hierarchical reductionist methodology used to investigate clonal selection.
285
The cellular level is representative of the standard forms of Artificial Neural Networks
(ANN) (for example see [334]). The environment at this level involves a stream of input
patterns referred to as a stimulus (for example a series of multivariate vectors). Therefore,
the scope of supervised and unsupervised learning schedules for adapting weights, as well
as recurrent and forward propagating structures belong to the cellular level. Bracketing
involves bounds on the cardinality of the units of adaptation (neurons), and their interac-
tions (synapses). Much like the clonal selection of immune system cells, the information
content from the environment is exposure-based, although in this case governed by the
sensors that perceive and communicate the information to the neural network. In addi-
tion to the trigger-based adaptation of intra-network connections, recurrent connections
provide a source of self-stimulation within the structure, much like the immune network
paradigm.
The group level is concerned with multiple functional groups of neurons that collec-
tively may form a complete nervous system. As such, the functional units are exposed
to an array of stimuli called a habitat. This provides a scale above the standard neural
network models, and considers multiple network approaches. There has been much work
on such models in the field not limited to hierarchical neural networks [34], and modu-
lar and ensemble neural networks [224, 357], and may others. The rise in the level of
abstraction may also provide opportunities for exploiting varied neuronal cell and con-
nection types including relay, motor, and sensory cells as well as the variety of neural
receptors and transmitters. The axes of complexity that may be bracketed include the
number of functional groups of cells as well as the amount of interaction between the
groups. The host level is concerned with populations of interacting nervous systems that
respond to an environment of habitats of stimuli mediated by the each host’s perception.
The environment includes the stimuli that may be provided by other nervous systems as
well as self-generated stimuli, and some control over preferential sensing from the environ-
ment. The scope, sources, and spatial temporal properties are expected to provide the the
greatest source of complexity at this level, requiring careful bracketing of the environmen-
tal information, habitat-mediated perception of information, and the interaction control
exerted by nervous systems. The population perspective permits the consideration of a
generational population structure that may adapt the structure and processes that govern
the propagation and manipulation of signals in networks of neurons. This is related to the
field of evolutionary artificial neural networks (for example see [440]), that is concerned
with the application of evolutionary processes to the concerns of neural networks at the
neuron level such as the architecture, the connections weight, and the learning rules used
to adjust the connection weights.
Other
The methodology, specifically the division and reduction concerns result in a clear hier-
archical system-environment perspective of a biological inspired computation paradigm
provides common ground for comparing and contrasting such paradigms at multiple lev-
els of abstraction. From the perspective of clonal selection, a brief interpretation of the
286
related fields of evolutionary computation and connectionist intelligence provides a infor-
mation processing context (more than superficial similarity) for directly relating models
and algorithms at the same level of abstraction (for example the relationship of ensemble
neural networks (mixture of experts) and island population genetic algorithms with the
tissue clonal selection paradigm), for exploiting the findings of such related approaches,
and identification of relevant and important application problem domains.
A clear difficulty in this approach is the elucidation of a discrete environment that
has meaning at each level of abstraction. For the connectionist paradigm in particular,
such environments may be conceptually intuitive, although difficult to articulate concisely
resulting in arbitrary description. The pattern of this mapping of external concerns to
an environment follows a clear trend across clonal selection, evolution and connection-
ist paradigms, which is the bracketing of a general environmental concern at the highest
level (antigenic, ecological, and sensory environments) to the scale of a specific functional
group of interest in the system. The comparison may be strength through the phrasing
of additional Computational Intelligence paradigms with regard to the division and re-
duction aspects of the methodology. For example one may consider the following Swarm
Intelligence paradigms: Particle Swarm Optimisation with reduction based on hierarchical
groups of swarming units, and Ant Colony Optimisation based on hierarchical groups of
foraging units.
7.5.4 Summary
The section summarises the generalised principles deduced from the consideration of the
adopted methodology, the application of the resulting framework to the broader field of
Artificial Immune Systems, and the general applicability of the methodology to related
Computational Intelligence sub-fields.
1. Methodology
(a) Four Step Process: The four-step process (metaphor, strategy, division, reduc-
tion) provides a clonal-selection focused integration and augmentation to the
three reviewed standard methodologies (see Section 2.7), promoting the con-
cerns of abstraction, modularity, extensibility, bracketing, and ultimate reinte-
gration.
(b) Alternatives: The specifics of cell-based reduction of the immune system (the
cellular, tissue, and host tiers) is one interpretation of such an reduction that
was demonstrated to be useful for the investigation of clonal selection. Alter-
native reductions exist both with regard to the specific aggregations of cells,
and with regard to the use of alternative criterion from which provide a basis
for reduction.
2. Broader Applicability
(a) Beyond Cells: The framework with its basis in aggregation of cells, provides an
intuitive and modular perspective on the immune system providing a context
287
for considering the scaling concerns of an array of cellular and molecular-based
computational metaphors from immunology.
(b) Transferable Findings: Given the central role clonal selection plays with regard
to the negative selection and immune network paradigms, the findings from the
investigation clonal selection are expected to be generally transferable, specifi-
cally with regard to the localisation and dissemination concerns investigated at
the tissue and host scales.
3. General Applicability
(a) Cross-Framework Perspective: The brief interpretation of evolutionary compu-

tation and connectionist intelligence in the context of the hierarchical frame-
work provided a common ground for comparison at multiple levels of abstraction
that revealed related algorithms and problem domains that may be exploited
and integrated in the clonal selection framework.
(b) Hybrid Approaches: The modularity of algorithms across paradigm hierarchies,
and the control over constraints provided by bracketing of complexity facilitates
the interchangeability of algorithms between frameworks, for example, the use
of a genome-level GA or neuron-level ANN at the cellular level in the IHCSF
when considering tissue or host level.
7.6 Chapter Summary

This section considers the broader principles and findings from the three proposed hier-
archical frameworks and the adopted systematic methodology that ultimately developed
the frameworks.
7.6.1 Frameworks Overview

The Hierarchical Clonal Selection Framework provided the integration of the cellular, tis-
sue, and host perspectives of the clonal selection adaptive strategy, and the suppression
of antigenic exposures to that of a generic exposure environment. The framework high-
lighted the fact that each tier imposes a distinct set of constraints on the cellular-focused
strategy, where bracketing provides a tool to manage subsumed complexity of the interac-
tions. It further suggested that the unbracketed complexity across the hierarchy results in
compounded interaction effects, revealing a broader hierarchical system and promotes con-
sideration of additional perspectives with which to constrain the cellular-focused strategy.
The Hierarchical Antigenic Exposure Framework provided the integration of the arbitrary
defined antigen, infection, and habitat perspectives of the piece-wise exposure of domain
information to a generic (suppressed) trigger-response based clonal selection system. The
framework showed that the expected capability of a system is compounded as additional
perspectives are considered, where strong bracketing of environmental information con-
tent bounds such complexity at the expense of forced response to each exposure. The
piece-wise exposure of an information environment may be investigated (and bracketed)
288
on three important axes of complexity: the cardinality or amount of information, the spa-
tial and temporal patterns by which the information is exposed, and the propagation of
both of these concerns down through the hierarchy.
The Integrated Hierarchical Clonal Selection Framework was concerned with the inte-
gration and interaction between the hierarchical clonal selection and antigenic exposure
frameworks, with a full elaboration of the system and environmental concerns. Impor-
tantly, the framework highlighted the fact that the maximum in bracketable complexity
of a given tier represents the minimum in bracketable complexity for the following tier
in the hierarchy (the maximum-minimum principle). The non-contigious and strongly-
bracketed perspectives of the framework provide opportunities for further investigation,
as does the consideration of tiers above and beyond those considered in the framework
and the elaboration of those perspectives already considered. The adopted Methodology
provided a general cellular focused approach for the systematic distillation of an interest-
ing information processing metaphor, and controlled steps of elaboration, reduction, and
further elaboration. Broadly, the artefact of the applied methodology (the IHCSF) may
be a useful tool for elaborating existing AIS frameworks beyond the constraints of a cellu-
lar perspective. Generally, the application of the methodology to existing Computational
Intelligence sub-fields provides both a context for comparison with superficially similar
approaches toward integration (hybridisation) and racing.
7.6.2 Integration
The proposition and analysis of the hierarchical frameworks provided both insights into
the design decisions, boundaries of concern, and limitations of the investigated clonal
selection models and algorithms from across the three tiers. Further, the analysis also
provided motivation for the investigation of the myriad of variations facilitated by the
strongly applied complexity bracketing. The application of the adopted methodology and
framework to related Computational Intelligence fields suggested at the potential insights
that may be provided in assessing proposed approaches against state of the art. This
suggestion motivates the next chapter the focus of which is dominated by the applicability
of the clonal selection models and algorithms from across the three tiers in the specific
context of Function Optimisation and Function Approximation problem domains.
289
Chapter 8
Suitability and Application

The previous chapters proposed and investigated models and algorithms inspired by clonal
selection with little regard for their applicability as Computational Intelligence tools. Be-
fore applicability is considered, it is important to outline a systematic methodology to
define suitability not just for the application of a technique to a problem, but also for
comparison of a technique to other similar techniques as well as the transferability of re-
sults on one problem to other similar problems. Section 8.2 considers the suitability of
application of the clonal selection algorithms. A simple feature-based suitability method-
ology is defined and is applied to the base algorithms from across the three paradigms
both eliciting general features and projecting those features onto general attributes of
computational problems. Two problem formalisms that embody many of the most diffi-
cult problems faced by Artificial and Computational Intelligence are reviewed: Function
Optimisation in Section 8.3 and Function Approximation in Section 8.4. Each formalism
is described in terms of (1) its general properties, (2) a set of specialised sub-problems,
and (3) select approaches from the fields of AI and CI that both address problem in-
stances that fit into the formalism and exhibit some superficially similar features with
clonal selection algorithms from across the hierarchical framework. These problem exam-
ples provide a tangible framing for the features of the clonal selection algorithms, which
are systematically mapped onto specific and suitable sub-problems from the respective
problem domains.
8.2 Suitability of Application

From a problem-solving perspective, the tools that emerge from the field of Computational
Intelligence are generally assessed with regard to their utility as efficiently or effectively
solving problems. An important lesson from the ‘no-free-lunch theorem’ was to bound
claims of applicability (Section 2.7.2). An approach toward this end is to consider the
suitability of a given strategy with regard to the feature overlap with the attributes of a
given problem domain. From a Computational Intelligence perspective, one may consider
the architecture, processes, and constraints of a given strategy as the features of an ap-
290
proach. This section lists such features of the approaches from the Cellular, Tissue, and
Host Paradigms and projects their features onto general capability with regard to prob-
lem features to which the approaches may be suitable. These so-called general features
provide the foundation for further consideration of the application of the clonal selection
algorithms from across the hierarchical framework. System features are discussed in the
context of the general intention and observed behaviours. Paradigms are discussed in a
compressed form, not taking into consideration the subsumed properties of the hierarchical
framework.
8.2.1 Considering Suitability

In this work, the suitability of the application of an ‘approach’ to a ‘problem’ takes into
considerations concerns such as the appropriateness (can the approach address the prob-
lem), the feasibility (available resources and related efficiency concerns), and the flexibility
(ability to address unexpected or unintended effects). This section summarises a general
methodology toward addressing the problem of suitability in the context of Computational
Intelligence tools. This methodology involves (1) the systematic elicitation of system and
problem features, and (2) the consideration of the overlap of problem-problem, algorithm-
algorithm, and problem-algorithm overlap of feature sets.
Systematic Feature Elicitation
A feature of a system (tool, strategy, model) or a problem is a distinctive element or

property that may be used to differentiate it from similar and/or related cases. Examples
may include functional concerns such as: processes, data structures, architectures, and
constraints, as well as emergent concerns that may have a more subjective quality such
as general behaviours, organisations, and higher-order structures. The process of the
elicitation of features may be taken from a system or problem perspective as follows:
• System Perspective: This requires a strong focus on the lower level functional ele-
ments and investigations that work toward correlating specific controlled organisa-
tions towards predictable emergent behaviours. An example is the central focus of
this work in Chapters 4 through 7.
• Problem Perspective: May require both a generalisation of the specific case to the
general problem case, as well as a functional or logical decomposition into constituent
parts.
Problem generalisation and functional decomposition are important and well used pat-
terns for problem solving in the broader fields of Artificial Intelligence and Machine Learn-
ing as the promotion of simplification and modularity can reduce the cost and complexity
of achieving solutions [342, 58].
291
Feature Overlap
Overlap in elicited features may be considered from three important perspectives: be-
tween systems, between problems, and between a system and a problem (depicted in Fig-
ure 8.1). Further, such overlap may be considered at different levels of detail with regard
to generalised problem solving strategies and problem definitions. These overlap cases are
considered as follows:
• System Overlap: Defines the suitability of comparing one system to another, referred
to as comparability. For example, systems may be considered for the same general
problems and compared in terms of theoretical or empirical capability, the results of
which may only be meaningful if the systems are significantly similar to each other
as assessed in terms of feature overlap.
• Problem Overlap: Defines the suitability of comparing one problem to another, re-
ferred to as transferability. From a systems focus for example, transferability refers
to the capability of a technique on a given problem to be transferred to another
problem, the result of which is only meaningful if there is a strong overlap between
the problems under consideration.
• System-Problem Overlap: Defines the suitability of a system on a given problem,

referred to as applicability. For example, a system is considered suitable for a given
problem if it has a significant overlap in capabilities with the requirements of the
problem definition.
Suitability of Application
Strategy Features Problem Features
Strategy Problem
Comparability Transferability
Strategy Problem
Figure 8.1: Depiction that the suitability of application of a strategy to a problem is

defined by the overlap in their features, as well as the comparability with other strategies,
and transferability of results to other problems.
Such mappings are expected to have noise given the subjective assessment and/or
complexity required in both the elicitation and consideration overlap of the of features,
292
the noisiest of which is expected to be the mapping between systems and problems. The
mapping of salient features of algorithms and problems was proposed as an important
reconciliation of the ‘no-free-lunch theorem’ by Wolpert and Macready [436], although the
important difference of this approach is that the system and algorithm are given prior
to the assessment. In [435], Wolpert and Macready specifically propose the elicitation
of the features from a problem first perspective, for which specialised algorithms can be
defined. Therefore, this methodology of suitability may be considered a generalisation
of this reconciliation suitable for the altered ‘Computational Intelligence’ (strategy first)
perspective on Artificial Intelligence.
8.2.2 Cellular Algorithms as Adaptive Strategies

This section reviews the general properties of cellular algorithms as adaptive strategies,
and considers the approach that such algorithms may be used for inductive modelling with
different perspectives of a given problem. Finally, the general cellular algorithm features
are listed, and projected onto suitable general problem features.
Cellular Algorithms Overview
The cellular algorithms are adaptive which is interpreted as their general capability of ob-
taining characteristics that improve the systems relative performance in an environment.
This adaptive behaviour is achieved through a selectionist process of iterative selection
and descent with modification. The discrete cell-based architecture is inherently par-
allel allowing for concurrent selection processes, and is robust providing redundancy of
information and flexibility in terms of resource allocation. The method of acquiring infor-
mation is called inductive learning (learning from example), where the approach uses the
implicit assumption that specific examples are representative of the broader information
content of the environment, specifically with regard to anticipated need. Generally, cel-
lular approaches maintain a population of samples that provide both a representation for
acquired information, and the basis for further induction. This method of simultaneously
improving information and optimising decisions is called the k-armed bandit (two-armed
and multi-armed bandit) problem from the field of statistical decision making [338] (for
a contemporary treatment see [39]). This class of problem has had a long tradition of
as a formalism for considering genetic algorithms and niching variants with regard to the
adaptive processes capability of the ‘automatic’ allocation of resources proportional to
expected payoff [173].
The acquired information is generally approximate, and is done so using a stochastic
method. The general method is called Monte Carlo in which randomness is exploited to
provide good average performance, quickly, and with a low chance of the worst case per-
formance. Such approaches are suited to problems with many coupled degrees of freedom,
for example large high-dimensional spaces. The selection method by which induction oc-
curs may be modelled as a series of parallel random walks that exploit gradients in the
underlying cost surface (directly, without derivatives), and as such is known as a Markov
chain Monte Carlo method [25, 76] (sampling from a target distribution function using a
293
Markov chain mechanism). This highlights that the robustness of the approach extends
to the induction process itself with regard to an improving approximation in the face of
potentially incomplete, incorrect and inconsistent sampled data. Finally, the induction
is performed in a piece-wise or online manner satisficing the concerns of real-time data,
with potentially dynamic changes that may be benign or adversarial with regard to the
internal model.
Inductive Model Generation
This section considers the scope of the inductive learning and resultant model in the
context of a given problem, and more specifically the information made available by the
problem. Three general classes of problem modelling are considered, as follows:
• Independent Models: The system models the entire scope of the problem irrespective
of any functional decompositions that may be available. This approach may be
referred to as holistic or naı̈ve models.
• Independent Sub-Models: The system models the problem in terms of clearly defined
sub-problems that are logically and/or functionally independent with regard to the
desired solution. This approach may be referred to as partitioned, symbolic, or
disjunctive models.
• Dependent Partial-Models: The system models the problem in terms of unclear or

implicit sub-problems that are logically and/or functionally dependent with regard
to the solution to the problem. This approach may be referred to as sub-symbolic,
or conjunctive models.
These three model types are used to assess both the Cellular Algorithms, as well as the
Tissue and Host Algorithms in the sections that follow. Figure 8.2 depicts the relationship
between the cellular algorithms and the three modelling types in the context of problem
instances that support such modelling. The figure clearly shows that cellular algorithms
model all three induction problems using a single repertoire of cells. The implication of
such an approach is that specific mechanisms may be required to facilitate the integration
of independent information in the case of sub-models, and inter-dependent information in
the case of partial models.
The three inductive modelling types may be related to the Antigen Colour Space Prob-
lems (ACSP) investigated in Chapter 4. The exposure of a single antigen (Nantigen = 1)
to the system may be considered an independent model, whereas the multiple antigen
(Nantigen = N ) and explicit aggregation of solutions may be considered the modelling
of multiple independent sub-problems. Importantly, given that the ACSP definition did
explicitly partition the antigenic set, although did not define the ordering of the exposure
of patterns to the system, the (Nantigen = N ) problem case may be considered a holistic
problem with unknown sub-problems and therefore given this unknown must be modelled
using partial-dependent sub-models (problem-side aggregation of responses). If the order-
ing of antigenic exposures was known to the system, then explicit sub-repertoires could be
294
Sub- Sub- Sub- Sub- Sub- Sub-
Problem Problem Problem Problem Problem Problem Problem
Model Model Model
(a) Independent Model. (b) Independent Sub-Models. (c) Dependent Partial-Models.
Figure 8.2: Depiction of the cellular algorithm approach to inductive learning with holistic,
sub-, and partial-models of a given problem.
defined and specialised for each pattern (system-side aggregation of responses). Finally,
the modelling of colour space patterns using colour components (in the case of DCCSA)
may be considered the modelling of the ACSP using independent cellular components
(one-cell equals one-component), although the modelling of sub-symbolic determinants
(such as partial strings) would have provided an example of dependent partial modelling
of single antigen.
Feature Summary
This section summarises the important descriptive features of Cellular Clonal Selection
Algorithms, and projects their features onto general features of problems to which this
class of algorithm may be suitable for application.
Salient System Features The following provides a list of the five features identified as
being distinctive in characterising the general class of Cellular Clonal Selection Algorithms.
• Selectionist Adaptation: Relative improvement of the system’s decision making ca-

pability in the context of a problem via a process of selection and decent with
modification to iteratively acquire and refine acquired information.
• Inductive Learning: Information is acquired and generalised from discrete and spe-
cific examples provided by the problem.
• Stochastic Process: The interaction with the problem and the resultant adaptation
have an inherent element of randomness that promotes approximation of acquired
information, general robustness of the system to noise, and flexibility of the system
to unexpected events.
• Population-based : Maintenance of a diverse although clustered set (inherent redun-

dancy) of pre-committed structures by the system, each of which may interact with
the problem potentially in a concurrent manner (inherent parallelism).
295
• Pattern Recognition: The acquisition of information and overt decision making are
coupled and triggered via a pattern recognition mechanism and response to infor-
mation from the problem.
Projected Problem Features The following provides a projection of the characterising

features of Cellular Clonal Selection Algorithms onto general problem features to which
such algorithms are expected to be suitable.
• Discrete: The problem is comprised of discrete packets information to which it is in

the systems best interest to respond.
• Online: Continuous decision making and improvement are desired given the variable
and potential dynamic or adversarial manner in which information is made available
to the system.
• Unknowns: Many of the parameters and/or constraints of the problem are known,
although important elements regarding the information content and/or ordering of
the exposure of information remain unknown to the system.
• Noise: The information signal may be disrupted, intermittent, and/or noisy requir-
ing a level of robustness of interpretation.
• Regularities: Problem information contains regularities or patterns that must be

exploited for effective progress to be made by a system.
8.2.3 Tissue Algorithms as Data Fusion

This section reviews the general properties of tissue algorithms as an adaptive data fusion
approach, and considers the approach that such algorithms are useful for the inductive
modelling using different perspectives of a given problem. Finally, the general tissue
algorithm features are listed, and projected onto suitable general problem features.
Tissue Algorithms Overview
Tissue algorithms are explicitly decentralised, with each tissue providing an adaptive pro-
cess making local decisions and improvements, the effects of which are implicitly inte-
grated into holistic understanding of the information environment. The induction of the
information content and anticipated exposure is extended across each point of exposure
providing a spatial acquisition and anticipation, that trades-off concerns of localisation
and dissemination. The tissue architecture is pre-determined, with tightly coupled and
fixed neighbourhood relationships between tissues. Tissues themselves are not redundant,
nor flexible in their topology (additions, modifications, deletions), although each tissues
cellular information content is redundant, allowing general or specific loss of random sets
of cells. The information composition across tissues is generally heterogeneous given the
specialisation of each tissue (given an asymmetric information exposure pattern), although
296
the general capability at any tissue in the system is equivalent (heterogeneous composi-
tion, homogeneous capability). This equivalence in capability is provided by the continuous
dissemination (high frequency, generally low amplitude) of the underlying system informa-
tion, and the localised as well as opportunistic way in which such information is exploited.
The general method is called multisensor data fusion (information fusion or distributed
sensing) whereby the modularity of information acquisition is exploited and combined for
holistic application providing an improved model over that of a series of independent mod-
els [195]. The modularity may be imposed requiring explicit decomposition or partitioning
of an information environment using prior information. Alternatively, the modularity of
the information and its integration with the system may be an inherent feature of the
environment. In this latter case, the interaction with the environment governed by the
tissue exposure regime may or may not be known, making information acquisition and ap-
plication opportunistic. Integration is achieved by the continuous recirculation and mixing
of acquired information, suggesting that the application of the integrated information may
also be piece-wise or modular like the acquisition of the information (so-called piecewise
acquisition and application from the system perspective, or exposure and expectation from
the environment perspective).
The tissue algorithms may be considered in the context of the three generative inductive
models, specifically independent models and sub-models, and dependent partial models.
Figure 8.3 depicts the general relationship between the tissue algorithms and the three
modelling types in the context of problem instances that support such modelling. The
architecture of tissue algorithms forces an an explicit partitioning of a given problem
information space whether such partitioning exists or not. The partitioning of the Infection
Antigenic Exposure Problem (IAEP) via the Tissue Exposure Regimes provides a basis for
considering the suitability of the tissue model against each modelling type (Chapter 5). In
the case of a holistic modelling, one may consider a symmetric or point exposure regime in
which case the persistent unbiased recirculation generally provides a disruptive influence to
such modelling. Similarly, one may consider an asymmetric exposure regime representative
of an explicit decomposition of the problem, where persistent unbiased recirculation has
the same disruptive effects. Specifically, the results suggest, that although the tissue
algorithms can construct such inductive models (innate capability), tissue algorithms may
not be suitable for constructing holistic or sub-models unless there is an expectation that
dissemination will be beneficial, for example in the problem definition, or if the intra-
tissue CCSA is capable of exploiting such information1 . This finding is further supported
by analogous findings from related approaches in Section 8.3.2 and Section 8.4.2.
Importantly, the functional decomposition was not included in the IAEP definition,
therefore, the tissue algorithms may be considered to have constructed partial models for
the unknown sub-problems in the problem definition (TER’s). As such, the aggregation
1
This latter example was demonstrated to not be the case for in the use of RCCSA as an intra-tissue
algorithm.
297
Sub-
Problem Sub- Sub- Sub-
Problem Problem Problem
Sub- Sub-
Model Model Model Sub- Sub- Sub- Partial Partial Partial

Model Model Model Model Model Model
(a) Independent Models. (b) Independent Sub-Models. (c) Dependent Partial-Models.
Figure 8.3: Depiction of the tissue algorithm approach to inductive learning with holistic,
of results for a specific TCSA across the TER’s provides a general indication of algorithm
capability. The aggregation of results clearly highlight that MTCSA is more suitable for
per-exposure responses, and that the recirculation algorithms may be more suitable for
holistic solution aggregation at any given single tissue under an unknown exposure regime,
likely at any given time under such a regime. More specifically, this suggests the suitability
of recirculation tissue approaches are with the redundancy of acquired information to
data loss and impromptu localised solution aggregation. Beyond this general suitability,
the results also suggest the specific suitability of recirculation methods under a random
exposure regime compared to the non-recirculation approach. Although the exploratory
results from Chapter 5 are not conclusive, they suggest strong constraints on the suitability
of applicability of recirculation-based Tissue Clonal Selection Algorithms.
Feature Summary
This section summarise the important descriptive features of Tissue Clonal Selection Al-
gorithms, and projects the features onto general features of problems to which this class
of algorithm may be suitable for application.
being distinctive in characterising the general class of (recirculation-based) Tissue Clonal
Selection Algorithms.
• Distributed Model : The population-based model is distributed across a series of

discrete and communicating sub-populations.
• Decentralised Adaptation: Acquisition of information and decision making occurs

locally at each distributed point of the system providing a natural mapping for
problems that can be functionally or logically decomposed.
• Tight Coupling: Sub-populations are organised in a spatial topology with fixed neigh-
bourhood relationships that generally communicate a low amplitude of information
with a high frequency.
298
• Data Fusion: Information acquired in localised regions of the structure is dissemi-
nated and made available across the system, providing an integrated although dis-
tributed holistic model.
• Opportunistic: The general availability of acquired information and decision making

capability facilitates the opportunistic exploitation of variably accessible problem
information or information processing capability.

features of Tissue Clonal Selection Algorithms onto general problem features to which such
algorithms are expected to be suitable.
• Discrete Decomposition: In addition to facilitating discrete interaction with the

system, the problem should be able to be functionally or logically decomposed into
sub- or partial problems which are naturally or explicitly partitioned across points
of contact with the system.
• Online: It is desirable for a holistic system to achieve continuous improvement

against the problem, although less sensitive with regard to behaviour across specific
spatial interactions.
• Spatial-Temporal Regularities: The notions of noisy information signals and regular-

ities extend beyond sub-problems, and extend across the spatially discrete interac-
tions with the system (notions of Tissue Exposure Regimes).
• On Demand : Information and/or decision making capability acquired by the system

holistically is required at a single given point of the system in an as needed manner
(impromptu and point-wise solution aggregation).
• Hazardous Environment: Addressing the problem requires some level of system ro-
bustness with regard to the loss, corruption, or failure of general (random sample
from all tissues) or specifically-localised (content of random tissue or information in
transit) sub-sets of information acquired by the system.
8.2.4 Host Algorithms as Mixtures of Experts

This section reviews the general properties of host algorithms as an adaptive mixture of
experts approach, and considers the application of such algorithms for the use in inductive
modelling, expliting different perspectives of a given problem. Finally, the general host
algorithm features are listed, and projected onto suitable general problem features.
Host Algorithms Overview
Like tissue algorithms, host algorithms are comprised of a set of decentralised processes
making localised decisions, that in aggregation represent the systems information com-
position and general capability. Unlike the tissue paradigm, hosts are loosely coupled,
299
without fixed neighbourhood relationships. The localised adaptive processes are quasi-
independent and their interactions are impromptu, opportunistic, and generally peer-to-
peer. Their quasi-independence is reflected in their perspectives of the information en-
vironment, where interactions do not promote an integration of perspectives, rather are
low-bandwidth dissemination of information (low frequency, generally high amplitude of
cells) from which, in aggregation the population is expected to benefit. As such infor-
mation dissemination is selective and organised although impromptu given the generally
unstructured nature of the population of hosts. The general method is called mixture of
experts (ensemble methods) where systems exploit the modularity of the perspectives of
the information environment by specialising in separate although potentially overlapping
perspectives [310, 327]. As with the data fusion perspective of the tissue paradigm, the
modularity of the environment may be imposed through explicit decomposition or parti-
tioning, or may be an inherent feature of the environment. Each host generates a model of
its perspective of the information environment through the cellular and tissues process of
induction via adaptation. The generational host algorithms relate to the general method
called multiple restarts (iterative restarts or multiple runs) as this generative process is pe-
riodically reinitialised for a different although potentially biased set of starting conditions
that influence the inductive model generation process [225, 69].
The host algorithms may be considered in the context of the three generative inductive
models, specifically independent models and sub-models, and dependent partial models.
Figure 8.4 provides a depiction of general host algorithms with regard to the three in-
ductive model types. The figures show the explicit partitioning of the model, as was the
case with the tissue algorithms. Also as with the tissue algorithms, spatial exposures
regimes were the focus of the capabilities of the host algorithms embodied in the Habitat
Antigenic Exposure Problem (HAEP). Model generation can be considered with regard
to the population-based and generation-based approaches separately. In the case of a
holistic and sub-model generation the non-interaction MP-HCSA demonstrated its suit-
ability with regard to exposure-wise aggregation of results (system error), and its lack
of suitability for impromptu point-wise aggregation of results in the same manner as the
MTCSA. Interestingly, both the Random-Pairing Pathogen Transmission HCSA and the
Small Sample SI-HCAS demonstrated a trend of decreased point-wise aggregation error
and in the latter case system error. This highlights the suitability of population-based
HCSA with small unstructured dissemination of acquired information or triggers not only
for holistic (SHER) and sub-models (AHER), but also for partial models (across unknown
HER’s), in particular, SI-HCSA-SS was competitive with MP-HCSA under the AHER
and SHER exposure regimes.
The generational approach of host populations provides capability for restarting the
model generation process that may remove specific training biases by using varied starting
conditions. Further, the adaptation of the model generation process itself in the EI-HCSA
provides capability for tuning the process toward improved resultant models. Although
300
Problem Sub- Sub- Sub-
Sub- Sub- Sub-
Sub- Partial
Model
Model Model
Model Model Sub- Sub- Partial Partial
Model Model Model Model
(a) Independent Models. (b) Independent Sub-Models. (c) Dependent Partial-Models.
Figure 8.4: Depiction of the host algorithm approach to inductive learning with holistic,
capable, neither approach was demonstrated suitable for such application from the results
of the limited exploratory experimentation, likely given the specifics of the configurations
used.
Feature Summary
This section summarise the important descriptive features of Host Clonal Selection Algo-
rithms, and projects their features onto general problem features to which this class of
algorithm may be suitable for application.
being distinctive in characterising the general class of Host Clonal Selection Algorithms.
• Distributed Model : The generation, storage and maintenance of the inductive model
is distributed across a set of discrete and communicating sub-systems, where the
acquisition of information and decision making is decentralised across the discrete
contacts with the problem promoting a natural mapping for partitioned or parti-
tionable problems.
• Loose Coupling: Sub-systems are not organised into any explicit spatial topology,
therefore neighbourhood interactions are impromptu, generally involving a small
number of sub-systems communicating a moderate amplitude of information with a
low frequency.
• Mixture of Experts: Sub-systems become specialists in potentially overlapping par-

titions of the problem space, from which infrequent intra-population communication
benefits the quasi-independent partial and/or sub-models.
• Multiple Restarts: The generational concerns of the model provide an iterative

restart mechanism providing multiple model generational trials that may be selec-
tively biased.
301
• Adaptation of Process: The evolutionary generational concerns of the model provide
an iterative adaptation mechanism for improving the model generation process with
regard to a specific problem ‘environment’.

features of Host Clonal Selection Algorithms onto general problem features to which such
algorithms are expected to be suitable.
• Discrete Partitions: The problem must facilitate discrete interaction with the sub-
systems, and likely be comprised of or partitioned into multiple sub- or partial
problems to which each sub-system may specialise.
• Online Improvement: It is desirable for the system to promote continuous iterative

improvement holistically (across sub-systems) likely in response to varying informa-
tion or exposure patterns, although there is less of a concern for per sub-system
improvement.
• On Demand : Access to the acquired information and/or acquired decision making

capability from across all sub-systems may be required holistically at a single point
of interaction with the problem in an impromptu manner.
• Unclear Constraints: The constraints of inductive model generation may be unknown

on unclear requiring multiple trials for generating models, and/or a meta-process to
improve the model generation process.
• Hazardous Environment: The computational environment in which the sub-systems

must operate to address the problem may be hazardous, resulting in the in the loss
of whole sub-systems and/or sub-sets of acquired information.
8.3 Function Optimisation

Real-world optimisation problems and generalisations thereof can be drawn from most
fields of science, engineering, and information technology (for a sample see [16, 402]).
Importantly, optimisation problems have had a long tradition in the fields of Artificial In-
telligence and Computational Intelligence in motivating basic research into new problem
solving techniques, and for investigating and verifying systemic behaviour against bench-
mark problem instances. This section considers the Function Optimisation Formalism and
related specialisations as a general motivating problem for demonstrating the suitability
of the applicability of clonal selection algorithms from across the hierarchical framework.
This is achieved firstly with a review of the problem formalism, nomenclature, and related
sub-fields. Selected Artificial and Computational Intelligence research is reviewed that
demonstrate feature overlap with concerns from at least one of the three clonal selection
paradigms, highlighting relevant problem features that such algorithms may exploit. Fi-
nally, the Cellular, Tissue, and Host Clonal Sselection Paradigms are mapped onto general
cases of Function Optimisation problems to which they are suited.
302
8.3.1 Function Optimisation Overview
Problem Description
Optimisation (in the mathematical sense), is defined as the search for a combination of
parameters commonly referred to as decision variables (x = {x1 , x2 , x3 , . . . xn }) which
minimise or maximise some ordinal quantity (c) (typically a scalar called a score or cost)
assigned by an objective function or cost function (f ), under a set of constraints (g =
{g1 , g2 , g3 , . . . gn }). For example, a general minimisation case would be as follows: f (x&) ≤
f (x), ∀xi ∈ x. Constraints may provide boundaries on decision variables (for example in
a real-value hypercube 5n ), or may generally define regions of feasibility and infeasibility
in the decision variable space or cost space. In applied mathematics the field may be
referred to as Mathematical Programming. More generally the field may be referred to
as Global or Function Optimisation given the focus on the objective function (for more
general information on optimisation, see [222, 330, 53]).
Sub-Fields of Study
The study of optimisation is comprised of many specialised sub-fields, generally based on

an overlapping taxonomy that focuses on the principle concerns in the general formalism.
For example, with regard to the decision variables, one may consider univariate and mul-
tivariate optimisation problems. The type of decision variables promotes the specialities
for continuous, discrete, and permutations of variables. Dependencies between decision
variables under a cost function defines the fields of Linear Programming, Quadratic Pro-
gramming, and Nonlinear Programming. A large class of optimisation problems can be
reduced to discrete sets, which are considered in the field of Combinatorial Optimisation,
to which many theoretical properties are known, most importantly that many interesting
and relevant problems cannot be solved by an approach with polynomial time complexity
(so-called NP-complete, for example see [311]).
The topography of the response surface for the decision variables under the cost func-
tion may be convex, which is an important class of functions to which many important
theoretical findings have been made, not limited to the fact that location of the local op-
timal configuration also means the global optimal configuration of decisional variables has
been located [53]. Many interesting and real-world optimisation problems produce cost
surfaces that are non-convex or so called multi-modal2 (rather than uni-modal) suggesting
that there are multiple peaks and valleys. Further, many real-world optimisation problems
with continuous decision variables cannot be differentiated given their complexity or lim-
ited information availability meaning that derivative-based gradient decent methods that
are well understood are not applicable, requiring the use of so-called ‘direct search’ (sample
or pattern-based) methods [272]. Further, real-world objective function evaluation may
be noisy, non-continuous, and dynamic, and the constraints of real-world problem solving
2
Taken from statistics referring to the centres of mass in distributions, although in optimisation it refers
to ‘regions of interest’ in the search space, in particular valleys in minimisation, and peaks in maximisation
cost surfaces.
303
may require a viable or approximate solution in limited time or resources, motivating the
need for inductive model-generation based approaches.
Selected Examples
This section reviews a select set of approaches toward addressing optimisation problems
from the field of Artificial and Computational Intelligence to both provide general insight
into the state of the interaction between stochastic algorithms and the field of optimisation,
and to provide a context for relating the clonal selection algorithms to optimisation prob-
lems by analogy. This section draws heavily from the field of Evolutionary Computation,
Swarm Intelligence, and related Computational Intelligence sub-fields (see Section 3.4).
• Global and Local Optimisation: Global Optimisation refers to seeking a globally opti-
mal structure or approximation thereof, typically in an unconstrained n-dimensional
real valued space. Global is differentiated from Local Optimisation in that the latter
focuses on locating an optimal structure within a constrained region of the decision
variable search space, such as a single peak or valley (basin of attraction). In AI and
CI literature, global optimisation problems refers to the class of optimisation prob-
lems that generally cannot be addressed through more conventional approaches such
as gradient decent methods (that require derivatives), and pattern search (that can
get ‘stuck’ in local optima and/or may never converge) [329, 402]. Such problems are
typically addressed through a directed search strategy with stochastic elements, typ-
ically involving a population-based model in the case of global search (for example
a GA, see Section 3.4.1), and a mutation hill climbing-based model for local search
(see Section 3.4.2). The general strategy involves sampling broadly initially, and
converging on those areas of the search space that are most likely to payoff toward
the end of the search. It is common to apply a local search method to the solutions
of a global search procedure as an refinement strategy.
• Parallel Optimisation: A natural first step in addressing difficult (large and rugged
cost landscapes) is to exploit parallel and distributed hardware, to get an improved
result in the same amount of time, or the same result in less time, or both [85].
Towards unifying the myriad of approaches and hardware configurations, a general
consensus and taxonomy has been defined by the Parallel Evolutionary Algorithms
(PEA) and Parallel Metaheuristics fields that considers the ratio of communication
to computation called granularity [68, 13]. This taxonomy is presented concisely by
Alba and Tomassini as a plot or trade-off of three concerns: (1) the number of sub-
populations (models or parallel strategies working on the problem), (2) the coupling
between the sub-populations (frequency and amplitude of communication), and (3)
the size of the sub-populations (size or extent of the sub-models) [14]. Two important
and relevant findings from the narrower field of Parallel Evolutionary Algorithms
include (1) that tightly coupled (frequent migration) between coarse-grained models
typically results in worse performance than a non-distributed approach [15], and (2)
that loose coupling (infrequent migration) between coarse-grained models has been
304
consistently shown to provide a super-linear increase in performance [12, 35, 68].
• Cooperative Search: This is a more general approach that considers the use of multi-
ple models that work together to address a difficult optimisation problems. Durfee,
et al. consider so-called Cooperative Distributed Problem Solving (CDPS) in which
a network of loosely coupled solvers are employed to address complex distributed
problems. Specifically, such systems are desirable to match the processing capa-
bilities of the solver with regard to the attributes of the problem. For example,
a given problem may have spatially distributed, functionally distributed, or tempo-
rally distributed sub-problems to which a centralised and monolithic system may not
be suitable. Lesser considers CDPS and proposes such models perform distributed
search on dependent or independent and potentially overlapping sub-problems as
a motivating perspective for conducting research into Distributed Artificial Intelli-
gence (DAI)3 [271]. Lesser points out that in real world applications, it is hard to
get a optimal mapping between the allocated resources to the needs or availability
of information for a given problem, suggesting that such problems may be caused by
a mismatch in processing times and/or number of sub-problems, interdependencies
between sub-problems, and local experts whose expertise cannot be effectively com-
municated. For a more detail on the relationships between parallel and cooperative
search, El-Abd and Kamel provide a rigorous taxonomy [134].
• Hybrid Search: Hybrid Search is a perspective on optimisation that focuses on the use
of multiple and likely different approaches either sequentially (as in the canonical
global and local search case), or in parallel (such as in Cooperative Search). For
example in this latter case, it is common in the field of PEA to encourage different
levels of exploration and and exploitation across island populations by varying the
operators or operator configurations used [382, 3]. Talbi proposed a detailed 4-
level taxonomy of Hybrid Metaheuristics that encompassed concerns of parallel and
cooperating approaches [379]. The taxonomy encompasses parallel and cooperative
considerations for optimisation and focuses on the discriminating features in the
lowest level such as heterogeneity, and specialisation of approaches.
• Functional Decomposition: Three examples of a functional decomposition of opti-

misation include (1) multiple objectives, (2) multiple constraints, and (3) partitions
of the decision variable search space. Multi-Objective Optimisation (MOO) is a
sub-field that is concerned with the optimisation of decision variables under a set
of objective functions. A solution to a MOO conventionally involves locating and
returning a set configurations under the objective functions called the optimal or
non-dominated Pareto set of solutions [124]. The complexity with MOO problems is
in the typically unknown dependencies between decision variables across objectives,
that in the case of conflicts, must be traded off (Purshouse and Fleming provide a
taxonomy of such complexity [331]). Constraint Satisfaction Problem’s (CSP) in-
volve the optimisation of decision variables under a set of constraints. The principle
3
This perspective provided the basis for what became the field of Multi-Agent Systems (MAS).
305
complexity in such problems is in locating structures that are feasible or violate the
least number of constraints, potentially optimising such feasibility [399, 263]. Search
Space Partitioning involves partitioning of the decision variable search space (for
example see Multispace Search by Gu, et al [131, 193, 192]). This is a critical con-
sideration given that for equal-sized dimensions bounds on parameters, the increase
in decision variables results in an exponential increase in the volume of the space to
search.
• Availability Decomposition: Optimisation problems may be partitioned by the con-

cerns of temporal and spatial distribution of (1) information availability, and (2)
computation availability. An interesting area of research regarding variable informa-
tion availability for optimisation problems is called Interactive Evolutionary Com-
putation, in which one or a collection of human operators dynamically interact with
an optimisation process [378]. Example problem domains include but are not lim-
ited to computer graphics, industrial design, image processing, and drug design.
There is an increasing demand to exploit clusters of heterogeneous workstations to
complete large-scale distributed computation tasks like optimisation, typically in an
opportunistic manner such as when individual machines are under utilised. The ef-
fect is that optimisation strategies, such as random partitioning of the search space
(independent non-interacting processing) are required to take advantage of such en-
vironments for optimisation problems [351, 274].
• Meta Optimisation: One may optimisation at a level above that considered in pre-
vious sections. Specifically, (1) the iterative generation of an inductive model called
multiple restart optimisation, and (2) the optimisation of the parameters of the
process that generates an inductive model of an optimisation problem. Multiple or
iterative restarts involves multiple independent algorithm executions from different
(random) starting conditions. It is generally considered as an method for achieving
an improved result in difficult optimisation problems in which a given strategy is
deceived by local or false optima [300, 225], typically requiring a restart schedule
[158]. A second and well studied form of meta optimisation involves the optimisa-
tion of the search process itself. Classical examples include the self-adaptation of
mutation parameters (step sizes) in the Evolutionary Strategy [333, 352] and Evolu-
tionary Programming [144, 143] approaches. Smith and Fogarty provided a review
of genetic algorithms with adaptive strategies providing a taxonomy in which the
meta-adaptations are applied at one of three levels: (1) the population (adapting the
overall sampling strategy), (2) the individual (adapting the creation of new samples
in the decision variable space), and (3) components (modifying component contri-
butions and/or individual step sizes as in ES and EP) [366].
8.3.2 Clonal Selection Applicability

This section considers the suitability of application of the clonal selection algorithms from
across the Cellular, Tissue, and Host Paradigms to the general problem of Function Opti-
306
misation. Applicability is considered in the context of the algorithm features reviewed in
Section 8.2 against the selected examples reviewed in the previous section.
Cellular Algorithms
A functional mapping of clonal selection onto the Function Optimisation formalism is

as follows: cells represent the decision variables (or a mapping thereof) that define a
structure, antigen is abstracted as the cost function and/or constraints imposed on the
structure, the repertoire represents a clonal set of cells specialised (converged) toward ad-
dressing the antigen in response to the affinity assigned to structures by the cost function.
This single-antigen mapping of clonal selection to the optimisation formalism was shown
in in Section 3.3.2 to be the basis of the development and investigation of a variety of
clonal selection algorithms not limited to CLONALG, BCA, and the IA family of algo-
rithms where in each case a repertoire of cells was adapted to address a cost function.
This mapping embodies the affinity maturation principle reviewed in Section 3.3.1 as a
constrained clonal selection metaphor suitable for application to optimisation domains,
specifically as a process directed toward the improvement of affinity (maturation) of cells
in the context of an antigen.
The general cellular algorithms (CCSA and RCCSA) are suited to global and local
optimisation in the case where the objective function is mapped to an affinity interaction
with a single antigen. One may consider the procedure a local search when a single cell
is used (Ncells = 1) and a multi-point global search procedure otherwise (Ncells > 1).
As such, the procedure is generally less sophisticated than a an Evolutionary or Swarm
Intelligence approach, and belongs to the general class of directed random search, and
(point and/or population-based) stochastic-based hill climbing and steady-state genetic
algorithm without crossover (see Section 3.4). The point-wise basis to exposures across all
cellular algorithms promotes considerations of parallelism within the repertoire, perhaps
exemplified in the SCCSA which strongly resembles the so-called fine-grained parallisation
of genetic algorithms for multi-processor hardware (such as diffuse and cellular genetic al-
gorithms [323]). The mediated (ECCSA) and network (NCCSA) algorithms do not have
clear analogies in the field of optimisation. Although these approaches are immature,
further development may reveal their capability in learning complex inter-relationships
between decision variables or other functional decompositions such as objectives, con-
straints, and regions in the search space. Finally, one may consider the functional de-
compositions of optimisation problems and their mapping onto the foundational CCSA
as either explicit sub-models or inferred partial models. The generality of these founda-
tional approaches means that augmentation may be required to effectively manage explicit
sub-problems such as MOO and CSP. A summary of the suitability of application for the
Cellular Clonal Selection Algorithms to Function Optimisation is provided in Table 8.1.
Tissue Algorithms
The functional mapping of tissue algorithms to Function Optimisation is as follows: tissues

are mapped onto different (logically or physically) global or local optimisation processes,
307
Algorithm Nantigen Ncells Optimisation Strategy Inductive Model
MCCSA 1 1 Local Search Holistic
R/CCSA 1 N Global Search Holistic/Partial
R/CCSA N N Decomposition Search Sub/Partial
SCCSA 1 N Parallel Global Holistic/Partial
E/NCCSA 1 N Decomposition Search Partial
E/NCCSA N N Decomposition Search Sub/Partial
Table 8.1: Summary of the projected applicability of Cellular Clonal Selection Algorithms
to Function Optimisation Problems.
recirculation is mapped onto inter-process communication and any overheads associated

with such communication (for example network latency and/or memory requirements),
infections may be mapped as single objective functions (single antigen), or comprised
of a set of objective functions, constraints or other functional decompositions (multiple
antigen). The tissue exposure regime may be influenced by a pre-designed arrangement
of tissues (architecture and network topology), or may represent the interaction of the
system with a known or unknown functional or logical decomposition of a given problem.
As considered in with cellular algorithms, a single tissue may be exploited for a global
or local optimisation process. Multiple non-interacting tissues in the MTCSA may be
considered parallel global search or parallel multiple restarts, whereas those algorithms
that use recirculation (RTCSA, HTCSA, ITCSA) may be considered cooperative search
parallel strategies. Given the high-frequency nature of recirculation, the RTCSA fits in
somewhere between coarse and fine-grained parallel algorithms, with a moderate popu-
lation size, moderate number of tissues, and fixed (directed cyclic graph) topology. Im-
portantly the reviewed findings from parallel EA’s suggest that such a configuration may
perform as well as or worse than a non-distributed model (single population) for global op-
timisation. Interestingly, such behaviour (strong competition and system-wide takeover)
may be desirable if parallel tissues cooperate using hybrid global optimisation strategies.
For example, variations on cellular algorithms (such as a global and local variation of a
given CCSA), or the holistic replacement of CCSA with other similar approaches (such as
evolutionary or swarm intelligence algorithms). Finally, the tissue architecture provides
a natural mapping for functional and availability decompositions of a given optimisation
problem. For example, objectives, constraints, and search space partitions may be ex-
plicitly or randomly assigned to tissues of the algorithm, where recirculation promotes
the cooperative concerns such as trade-off’s between the sub-problems. The tissue archi-
tecture is suited to such problem descriptions, although this suitability may be further
improved if the functional decomposition is motivated by an availability decomposition
such as distributed information availability, event-driven (such as human interaction), or
opportunistic allocation of computation time. Table 8.2 summarises the applicability of
TCSA to Function Optimisation problems.
308
Algorithm Ninf ection Ntissues Optimisation Strategy Inductive Model
MTCSA 1 1 Global Search Holistic
MTCSA 1 N Parallel Global Holistic
MTCSA N N Parallel/Decomposition Sub
R/H/ITCSA 1 N Cooperative/Hybrid Holistic
R/H/ITCSA N N Cooperative/Decomposition Sub/Partial
Table 8.2: Summary of the projected applicability of Tissue Clonal Selection Algorithms
to Function Optimisation Problems.
Host Algorithms
The functional mapping of host algorithms to Function Optimisation is as follows: hosts

may map onto different (logically or functionally) global or local search processes (Ntissues =
1 per H), or more interestingly onto groups of tightly-coupled clusters of global or local
search processes (Ntissues > 1 per H). Similarly, a habitat may be a single objective func-
tion (Ninf ection = 1 per B), a set of objectives functions or constraints (Ninf ection > 1 per
B), or a set of holistic although different function optimisation problems (Nhabitats > 1).
As with the mapping of the tissue paradigm, the complexity of the system-problem inter-
action is relegated to the (in this case host) exposure regime.
Given the bracketing of tissues, one may consider the extent of capabilities of the tissue
algorithm subsumed by a host or a population of hosts, as was the case between the tissue
and cellular paradigms. As such, if we ignore (constrain to Ntissue = 1) tissue cardinality,
the MP-HCSA provides a set of independent global or local strategies. The SI-HCSA may
be exploited for cooperative parallel global optimisation, given their strong resemblance in
configuration to parallel evolutionary algorithms. Specifically, the small to moderate num-
ber of hosts, moderate to large population sizes, and low frequency, low amplitude sharing
of information (decrease in frequency compared to experimentation in Chapter 6). Fur-
ther, the sharing-based population host algorithms encourage an unstructured topology of
impromptu host interactions, more akin to a P2P network topology than a master-server,
star or other fixed topology commonly exploited by PEA’s. The cooperative relationship
of SI-HCSA may also be exploited for hybrid optimisation, perhaps with an increased fre-
quency of migration, as was the case in the exploratory experimentation. As with tissue
paradigm, the population-based host algorithms may exploit the natural modularity of
optimisation problems in terms of their functional and information availability decompo-
sitions, although it is unclear whether the varied information sharing scheme of SI-HCSA
would promote the sufficient level of integration of partial-models where required (depen-
dent decompositions). Interesting, the THCSA, specifically the Pathogen and Vaccination
variations promote the consideration of sharing the triggers of adaptation, in this case an
explicitly partitioned or limited availability resources between nodes in the system. Where
this mechanism is supported under the constraints of a specific problem. Such behaviour
may promote an adaptive-self organisation of responsibility within the host population.
The generational host algorithms provide an clear relationship to the meta-optimisation
techniques. MG-HCSA with a single host provides a model for a sequential restart strat-
309
Algorithm Nhabitats Nhosts Optimisation Strategy Inductive Model
MP-HCSA 1 1 Global Search Holistic
MP-HCSA 1 N Parallel Global Holistic
SI-HCSA 1 N Parallel/Cooperative/Hybrid Holistic
SI/THCSA N N Cooperative/Decomposition Sub/Partial
MG-HCSA 1/N 1 Sequential Iterative-Restarts Holistic
MG-HCSA 1 1/N Parallel Iterative-Restarts Holistic
MI-HCSA 1 1/N Multiple Restart/Meta Holistic
EI-HCSA 1 N Meta-Optimisation Holistic
Table 8.3: Summary of the projected applicability of Host Clonal Selection Algorithms to
Function Optimisation Problems.
egy, and a parallel restart strategy when configured with multiple hosts. The MI-HCSA
algorithm provides a restart strategy in which the initialisation of naı̈ve systems is bi-
ased with acquired information. This strategy may be effective in environments where
the availability is subject to frequent and unexpected change to seed new instances of
host processes (such as in the exploitation of idle computation time in a network envi-
ronment). Finally, the suggestion of optimising the initial conditions in the EI-HCSA
provides a link between the restart strategy in GHCSA and meta optimisation strategies
that optimise the processes that operate on a given problem instance. Importantly the
meta optimisation strategies (restarts and optimisation of process) apply not only to all
cellular and tissue approaches to which the hosts subsume, but also to the capabilities of
the population-based host algorithms. Table 8.3 summarises the applicability of HCSA to
function optimisation problems.
8.3.3 Summary
This section summarises the insight provided into the applicability of the clonal selection
algorithms from across the three paradigms generally with regard to Function Optimisa-
tion, and specifically with regard to some specialisations of the problem and techniques
for addressing the problem.
1. General Suitability
(a) Mapping: The mapping between the general clonal selection adaptive strategy
and function optimisation is natural for the specific case of blindly adapting a
clonal set for a single antigen.
(b) Systems: Beyond the adaptive cellular level, the tissue and host paradigms
provide a basis for decentralised and distributed optimisation, given a suitable
problem mapping.
(c) Environments: Global optimisation alone is insufficient to motivate application
to tissue and host paradigms, requiring functional and availability decomposi-
tions of a given optimisation problem to provide an effective mapping of the
problem to the antigenic exposure paradigm.
2. Algorithm Applicability
310
(a) Cellular : R/CCSA are suitable for local and global optimisation tasks, poten-
tially for functional decompositions of optimisation problems, and SCCSA is
suitable for parallisation.
(b) Tissue: MTCSA is suitable for parallel global optimisation, R/H/ITCSA are
suitable for cooperative hybrid and potentially functional decomposition, and
likely not suitable for cooperative global search.
(c) Host: MP-HCSA is suitable for parallel global search, SI-HCSA is suitable for
parallel global search and likely cooperative hybrid and functional decomposi-
tion, THCSA variants are suitable for functional decompositions and potentially
cooperative parallel optimisation. MG-HCSA is suitable for multiple restarts,
MI-HCSA may be suitable for biased multiple restarts, and EI-HCSA suggests
at meta optimisation with further extension.
8.4 Function Approximation

The phrasing of real-world problems in the Function Approximation formalism are among
the most computationally difficult considered in the broader field of Artificial Intelligence
for reasons including: incomplete information, high-dimensionality, noise in the sample
observations, and non-linearities in the target function. This section considers the Function
Approximation Formalism and related specialisations as a general motivating problem for
demonstrating the suitability of the applicability of clonal selection algorithms from across
the hierarchical framework. Firstly the general problem is reviewed including standard
nomenclature and a summary of sub- and related fields of study. Selected Artificial and
Computational Intelligence research is reviewed that demonstrate feature overlap with
concerns from at least one of the three clonal selection paradigms, highlighting relevant
problem features that such algorithms may exploit. Finally, the Cellular, Tissue, and Host
Clonal Selection Paradigms are mapped onto general cases of Function Approximation
problems to which that are proposed to be well suited.
8.4.1 Function Approximation Overview

Problem Description
Function Approximation (in the mathematical sense) is an inductive problem of finding a

function (f ) that approximates a target function (g), where typically the approximated
function is selected based on a sample of observations (x, also referred to as the training
set) taken from the unknown target function. In machine learning, the function approx-
imation formalism is used to describe general problem types commonly referred to as
pattern recognition, such as classification, clustering, and curve fitting (so-called decision
or discrimination function). Specifically, such general problem types are described in terms
of approximating an unknown Probability Density Function (PDF), which underlies the
relationships in the problem space, and represented to some degree in the sample data.
This function approximation perspective of such problems is commonly referred to as
statistical machine learning and/or density estimation [159, 46].
311
Sub-Fields of Study
The function approximation formalism can be used to phrase some of the hardest prob-
lems faced by Computer Science, and Artificial Intelligence in particular such as natural
language processing and computer vision. The general process focuses on (1) the collec-
tion and preparation of the observations from the target function, (2) the selection and/or
preparation of a model of the target function, and (3) the application and ongoing refine-
ment of the prepared model. Some important problem-based sub-fields include: Feature
Selection where a feature is considered an aggregation of attributes, where only those fea-
tures that have meaning in the context of the target function are necessary to the modelling
process [262, 194], Classification where observations are inherently organised into labelled
groups (classes) and a supervised process models an underlying discrimination function
to classify unobserved samples, Clustering where observations may be organised into in-
herent groups based on common features although the groups are unlabelled requiring a
process to model an underlying discrimination function without corrective feedback, and
Curve or Surface Fitting where a model is prepared that provides a ‘best-fit’ (or regres-
sion) for a set of observations that may be used for interpolation over known observations
and extrapolation for observations outside what has been observed.
The field of Function Optimisation is related to Function Approximation, as many-sub-
problems of Function Approximation may be defined as optimisation problems. As such
many of the inductive modelling paradigms are differentiated based on the representation
used and/or the optimisation process used to minimise error or maximise effectiveness
on a given approximation problem. The difficulty of Function Approximation problems
centre around (1) the nature of the unknown relationships between attributes and features,
(2) the number (dimensionality) of of attributes and features, and (3) general concerns
of noise in such relationships and the dynamic availability of samples from the target
function. Additional difficulties include the incorporation of prior knowledge (such as
imbalance in samples, incomplete information and the variable reliability of data), and
problems of invariant features (such as transformation, translation, rotation, scaling and
skewing of features).
Selected Examples
This section reviews a select set of approaches toward addressing Function Approximation
problems from the field of Artificial and Computational Intelligence to both provide general
insight into the state of the interaction between stochastic algorithms and the field, and to
provide a context for relating the clonal selection algorithms to Function Approximation
problems by analogy. The review draws heavily from the fields of artificial neural networks,
specifically competitive learning, as well as related inductive machine learning fields such
as instance based learning.
• Vector Quantisation: Vector Quantisation (VQ) refers to a method of approximat-

ing a target function using a set of exemplar (prototype or codebook) vectors. The
exemplars represent a discrete sub-set of the problem, generally restricted to the
312
features of interest using the natural representation of the observations in the prob-
lem space, typically an an unconstrained n-dimensional real valued space. The VQ
method provides the advantages of a non-parametric model of a target function (like
instance-based and lazy learning such as kNN, see Section 3.4) using a symbolic
representation that is meaningful in the domain (like tree-based approaches). The
promotion of compression addresses the storage and retrieval concerns of kNN, al-
though the selection of codebook vectors (the so-called quantisation problem) is a
hard problem that is known to be NP-complete [162]. More recently Kuncheva and
Bezdek have worked towards unifying quantisation methods in the application to
classification problems, referring to the approaches as Nearest Prototype Classifiers
(NPC) and proposing a generalised nearest prototype classifier [264, 265].
• Parallelisation: Instance-based approaches are inherently parallel given the generally

discrete independent nature in which they are used, specifically in a case or per-
query manner. As such, parallel hardware can be exploited in the preparation of the
corpus of prototypes (parallel preparation), and more so in the application of the
corpus given its read-only usage [1, 301, 326]. With regard to vector quantisation
specifically, there is an industry centred around the design and development of VQ
and WTA algorithms and circuits given their usage to compress digital audio and
video data [302, 312].
• Cooperative Methods: Classical cooperative methods in the broader field of statisti-

cal machine learning are referred to as Ensemble Methods [310, 327] or more recently
Multiclassifier Systems [171]. Boosting is based on the principle of combining a set
of quasi-independent weak learners that collectively are as effective as a single strong
learner [244, 349]. The seminal approach is called Adaptive Boosting (AdaBoost)
that involves the preparation of a series of classifiers, where subsequent classifiers
are prepared for the observations that are misclassified by the proceeding classifier
models (creation of specialists) [350]. Bootstrap Aggregation (bagging) involves par-
titioning the observations into N randomly chosen subsets (with reselection), and
training a different model on each [55]. Although robust to noisy datasets, the ap-
proach requires careful consideration as to the consensus mechanism between the
independent models for decision making. Stacked Generalisation (stacking) involves
creating a sequence of models of generally different types arranged into a stack,
where subsequently added models generalise the behaviour (success or failure) of the
model before it with the intent of correcting erroneous decision making [434, 394].
• Functional Decomposition: As demonstrated, it is common in ensemble methods to

partition the dataset either explicitly or implicitly to improve the approximation
of the underlying target function. A first important decomposition involves parti-
tioning the problem space into sub-spaces based on the attributes, regular groups
of attributes called features, and decision attributes such as class labels. A popu-
lar method for attribute-based partitioning is called the Random Subspace Method,
involving the random partitioning of attributes to which specialised model is pre-
313
pared for each (commonly used on tree-based approaches) [209]. A related approach
involves a hierarchical partitioning of attributes space into sub-vectors (sub-spaces)
used to improve VQ-based compression [170]. Another important functional decom-
position involves the partitioning of the set of observations. The are many ways
in which observations may be divided, although common approaches include pre-
processing using clustering techniques to divide the set into natural groups, addi-
tional statistical approaches that partition based on central tendency and outliers,
and re-sampling methods that are required to reduce the volume of observations.
• Availability Decomposition: The availability observations required to address func-

tion approximation in real-world problem domains motivate the current state of the
art in Distributed Data Mining (DDM or so-called Collective Data Mining), Parallel
Data Mining (PDM), and Distributed Knowledge Discovery in Database (DKDD)
[239]. The general information availability concerns include (1) the intractable vol-
ume of observations, and (2) the spatial (geographical) and temporal distribution of
information [444]. It is common in many large real-world problems for it to be
infeasible to centralise relevant observations for modelling, requiring scalable, load
balancing, and incremental acquisition of information [362].
• Meta Approximation: The so-called ensemble or multiple-classifier methods may be

considered meta approximation approaches as they are not specific to a given mod-
elling technique. As with function optimisation, meta-approaches may be divided
into restart methods and meta-learning algorithms. The use of restart methods is
a standard practice for connectionist approaches, and more generally in approaches
that use random starting conditions and a gradient or local search method of re-
finement. The method provides an opportunity for over-coming local optima in the
error-responds surface, when there is an unknown time remaining until convergence
[278], and can exploit parallel hardware to provide a speed advantage [47]. Ensemble
methods and variants are examples of meta approximation approaches, as well as
the use of consensus classifiers (gate networks in mixtures of experts) to integrate
and weight the decision making properties from ensembles.
8.4.2 Clonal Selection Applicability

This section considers the suitability of the application of the clonal selection algorithms
from across the Cellular, Tissue, and Host Paradigms to the general problem of Func-
tion Approximation. Applicability is considered in the context of the algorithm features
reviewed in Section 8.2 against the selected examples reviewed in the previous section.
Cellular Algorithms
A generalised mapping of clonal selection onto the Function Approximation formalism is

as follows: cells are exemplars (a.k.a. codebooks or prototypes) in the parameter space
(or a mapping thereof) of the target function, antigen are the sample of observations from
314
the target function (therefore there are always N ), the repertoire represents an approx-
imation of the target function adapted in the presence of the sample observations, and
affinity is calculated based on a distance metric between observations and exemplars. As
considered in the review of the state of clonal selection algorithm research (Section 3.3.2),
this was the perspective used in the development and application of AIRS to classification
where exemplars quantised observations. This mapping was also used in the application of
CLONALG to binary pattern recognition, and was the basis for the colour space problems
investigated across the three clonal selection paradigms in this work. As considered in
paradigms related to the field of clonal selection algorithms in Section 3.4, the algorithms
competition-based approach to preparing the exemplars in response to the sample obser-
vations is related to the field of competitive learning, whereas the interpretation of the
resulting exemplars is related to the field of instance-based learning, specifically k-Nearest
Neighbours (kNN).
The cellular clonal selection algorithms provide a series of global optimisers suitable for
vector quantisation, not unlike the use of EC, SA, and hill climbing for the same task. The
general nature of VQ makes the cellular algorithms suitable for classification (Nearest Pro-
totype Classifiers), clustering, and curve fitting. Given the competition-based interaction
between exemplars and the unknown context-based relationships that such competition
represents, it is reasonable to categorise the cellular clonal selection algorithms as the
induction of partial-models. Although it is important to point out that it is the degree to
which such relationships have meaning in the problem space that defines whether the ex-
emplars may be considered partial (strongly related) or sub (weakly or unrelated) models
of the broader function approximation problem. The iterative (per-exposure) competition
based learning and decision making suggest at the general suitability of the approach to
incremental and dynamic specialisations of the general problem. The spatial properties
of SCCSA show similarities with projection-based clustering methods such as SOM and
Neural Gas, although unlike those approaches SCCAS in its present form does not demon-
strate the preservation of topological features. As with function optimisation, the role of
ECCAS and NCCAS is not clear given the approaches immature although demonstrated
similarities with sub-symbolic feature-based approaches such as Learning Classifier Sys-
tem. The VQ approach to function approximation highlighted an important perspective on
negative selection algorithms and their relationship to clonal selection algorithms. Specif-
ically, that such approaches may be considered in the context of a function approximation
formalism (the standard domains of anomaly and change detection as discriminating func-
tions), where the sub-string approaches used (hamming distance, r-chunks, and others)
may be considered the simultaneous identification and extraction of features and instance-
based modelling of the underlying PDF. Table 8.4 summarises the suitability of the Clonal
Selection Algorithms for general variations of the Function Approximation formalism.
Tissue Algorithms
The mapping of tissue algorithms onto the Function Approximation formalism is as fol-
lows: tissues represent individual vector quantisation models each with a set of exemplars,
315
Algorithm Nantigen Ncells Approximation Strategy Inductive Model
R/CCSA N N Vector Quantisation Sub/Partial
SCCSA N N Vector Quantisation Partial
E/NCCSA N N Feature Extraction Partial
Table 8.4: Summary of the projected applicability of Cellular Clonal Selection Algorithms
to Function Approximation Problems.
recirculation represents inter-process communication (parallel and/or distributed) between

the quantisation models, infections represent logical or functional groups of observations
from the target function, and the tissue exposure regime represents the mapping of the
logical or function partitions of the observations to the individual models where the map-
ping may be known: designed allowing tissue-specialisation on sub-problems, or unknown:
based on availability decompositions forcing tissue-specialisation on partial-problems.
The MTCSA provides a candidate for ensemble methods given the required indepen-
dence between such models, although the sharing of exemplars between the ensemble in
the case of recirculation algorithms likely makes a mapping for the R/H/ITCSA unsuitable
for these specific cooperate approaches. Specifically, the tissue topology and the MTCSA
has a strong resemblance to the ‘bagging’ and potentially the ‘boosting’ ensemble meth-
ods. More generally, the TCSA approaches may be used for functional decompositions of
the problem. The persistent communication and information sharing in the recirculation
algorithms strongly suggests suitability of the approaches as a distributed vector quan-
tisation method beyond efficiency concerns of parallelisation (such as Parallel-AIRS), in
particular for those real-world problems with information and communication availability
constraints (considered in DDM and DKDD). This suitability is likely bounded to cases
where adaptive localisation and dissemination are required, for example where there is an
unknown partitioning of the observations (availability of information) and the use and/or
exploitation of the decision making capability (such as use of the approximated function
for discrimination) also occurs in an unknown and partitioned (localised) manner. This
suitability scenario may be considered a general application-case of Hofmeyr’s proposed
use of his negative selection system (later ARTIS and LISYS) in distributed anomaly
detection (see Section 2.5.2), although constrained to a localised ring network topology
of machines. Table 8.5 provides a summary of the suitability of tissue clonal selection
algorithms to function approximation.
Algorithm Ninf ection Ntissues Approximation Strategy Inductive Model

MTCSA N N Cooperative/Decomposition Sub
R/H/ITCSA N N Cooperative/Decomposition Partial
Table 8.5: Summary of the projected applicability of Tissue Clonal Selection Algorithms
to Function Approximation Problems.
316
Host Algorithms
The functional mapping of host algorithms is as follows: hosts may be considered col-
lections of highly integrated vector quantisation models or large (bracketed) models, host
interaction may be considered the inter process, inter system, or inter-network communi-
cation involving either the sharing of model information or information to which models
are exposed, habitats represent different perspectives of the observations in terms of func-
tional or availability decompositions, where the host exposure regime defines the designed
or opportunistic mapping between such decompositions and the system.
The population of hosts provide the similar capabilities as the tissue algorithms, al-
though scaled up. In particular the MP-HCSA provides a broader interpretation of en-
semble methods considering the hierarchical formulation of such approaches. Importantly
the SI-HCSA provides a less rigid communication mechanism between VQ models, that
permits a more directed and/or rapid replication of acquired information more akin to
ARTIS in a broadcast environment, and peer-to-peer propagation. This is also related
to epidemic and gossip based replication of information in distributed databases which
host models could adopt toward predictable efficiency gains [127, 413]. The generational
algorithms exploit the restart property inherent in the global optimisation algorithm that
underlies the adaptive capability of the approach. As such MI-HCSA provides general bias
of ensembles and EI-HCSA provides general adaptation of the process within ensembles.
Table 8.6 provides a summary of the suitability of the Host Clonal Selection Algorithms
to Function Approximation.
Algorithm Nhabitats Nhosts Approximation Strategy Inductive Model

MP-HCSA N N Cooperative/Decomposition Sub
SI/THCSA N N Cooperative/Decomposition Partial
MG-HCSA N 1 Sequential Iterative-Restarts Sub
MG-HCSA N N Parallel Restarts Sub
MI-HCSA N N Multiple Restart/Meta Holistic
EI-HCSA N N Meta-Optimisation Holistic
Table 8.6: Summary of the projected applicability of Host Clonal Selection Algorithms to
Function Approximation Problems.
8.4.3 Summary
This section summarises the insights provided into the suitability of the applicability of
the clonal selection algorithms from across the three paradigms, generally with regard
to Function Approximation, and specifically with regard to some specialisations of the
problem and techniques for addressing the problem.
1. General Suitability
(a) Mapping: The mapping of clonal selection to vector quantisation based func-
tion approximation provides a natural fit, that may only be further improved
through the parallel task of feature selection.
317
(b) Systems: Clonal selection provides an optimisation process over the scope of
managed exemplars, that provides a natural parallel realisation of the mapping
between the process and function optimisation. As such, the system is always
comprised of > 1 cell.
(c) Environments: The set of observations from the target function and the iter-
ative exposure and adaptation or application in response to such observations
embodies the principles and provide a natural mapping for onto the antigenic
exposure paradigm.
2. Algorithm Applicability
(a) Cellular : R/CCSA provides a competitive learning realisation of vector quan-

tisation applicable for compression, clustering, and nearest prototype classifi-
cation.
(b) Tissue: MTCSA provides a natural suitability for ensemble methods, bagging
in particular, whereas the recirculation algorithms (R/H/ITCSA) are suited
to availability decompositions of problems, in particular information availabil-
ity concerning observations from the target function, and the unknown and
point-wise application of the approximated function across the tightly-coupled
deployment (DDM and DKDD).
(c) Host: MP-HCAS provides a broader scoped suitability for parallel and hierar-
chical ensemble methods, where as the SI/THCSA are suited to less constrained
topologies for DDM and KDD with both model and observation information
sharing. Generation HCSA are suitable for restart and iterative tuning of the
model and process respectively, both of which are concerns of the underlying
stochastic-optimisation process.
8.5 Chapter Summary

This section reviews the findings from this chapter regarding the investigated clonal selec-
tion algorithms suitability and applicability, specifically in the context of Function Opti-
misation and Function Approximation.
8.5.1 Suitability and Applicability

General Suitability Section 8.2 considered the general suitability of the clonal se-
lection algorithms, specifically with regard to the systematic elicitation of features from
the general algorithms (cellular, tissue, host) from across the hierarchical clonal selection
framework. The general suitability and projection onto problem attributes is summarised
as follows:
1. Cellular Algorithms: Stochastic-adaptive algorithms that use a population-based

pattern recognition as a method for induction and modelling of a problems. Suited to
318
problems that facilitate discrete interaction, that provide noisy although somewhat
regular signals with unknown levels of complexity likely in an online manner.
2. Tissue Algorithms: Distributed model of a domain using a decentralised adapta-

tion across multiple tightly coupled populations with high-frequency communication
promoting integration of acquired information. Suited to noisy complex problems
with regularities that are both spatial as well as temporal, operating in a potentially
hazardous and/or opportunistic computational environment.
3. Host Algorithms: Distributed model of a domain using a set of loosely-coupled pop-

ulations each specialising in an aspect of the problem with low frequency communi-
cation and long term strategies including model restarts and adaptation of internal
processes. Suited to large and complex problems requiring online continuous im-
provement, likely with unclear constraints and hazards.
Problem Applicability Sections 8.3 and 8.4 considered the applicability of the clonal
selection algorithms in the context of the Function Optimisation and Function Approxi-
mation problem formalisms respectively, highlighting sub-problems to which specific algo-
rithms are suggested to be suited. The general mapping of the clonal selection paradigm
and applicability to the respective problems is summarised as follows:
1. Optimisation: The adaptive qualities of clonal selection, specifically a clone (group)

of cells in response to an antigenic stimulus provides a natural mapping to an adap-
tive strategy for function optimisation. This is a reduction of the capability and
complexity of the general strategy, which may be further exploited through func-
tional and/or information availability decompositions of a given problem instance,
both improving the mapping and employing more available problem specific infor-
mation.
2. Approximation: The adaptive qualities of clonal selection considered in a parallel

and decentralised where a system is presented with a large set of stimuli to which
to model, provides a natural mapping to a decentralised adaptation strategy for
exemplar and pattern recognition based function approximation.
8.5.2 Integration
This final chapter provides closure on the adaptive and distributed conceptual models
and clonal selection algorithms proposed and investigated in this work by grounding their
salient features to specific problem formalisms and sub-problems. Although the claims of
suitability and applicability are reasoned from the qualitative observations from previous
chapters, such claims require verification on specific standard benchmark and real-world
problem instances. This clear and important extension of this investigation is considered
in conjunction with further extensions, limitations, and criticisms in the following Chapter
that concludes this work.
319
Chapter 9
Conclusions
9.1 Research Goals

This work has demonstrated that the interpreted structure and function of the immune
system do provide an effective motivating metaphor for the realisation of distributed clonal
selection algorithms. This section reviews the contributions of this dissertation in the
context of the research hypothesis, defined by five goals outlined in Section 1.3.
1. Identify a systematic methodology for realising a novel biological inspired

computational framework and models. A systematic methodology was outlined in
Section 2.7, comprised of (1) an immunological and information processing centric frame-
work for progressing from metaphor, model, strategy, and algorithm, (2) an experimental
framework for empirical verification and demonstration of computational models and algo-
rithms, and (3) an adaptive algorithm centric framework for the integration of qualitative
and quantitative information regarding so-called ‘conceptual machines’. This methodology
was applied at three different and inter-related scales: (1) General Metaphor : The broad
consideration of the information processing properties of the clonal selection theory, inves-
tigated in the context of three specific and distinct interpretations called paradigms. (2)
Specific Metaphors: The consideration of computational paradigms inspired by a specific
immunological metaphors that constrain the clonal selection process, which are investi-
gated through instantiations of specific models and algorithms. (3) Specific Algorithms:
The consideration of a specific algorithm in the context of a broader paradigm which is
described in terms of its further constrained metaphor, information processing strategy,
general model, algorithmic procedure instantiation, and empirical verification and demon-
stration of expected behaviours.
2. Identify limitations with and elaborate upon the base clonal selection
paradigm. The Clonal Selection Theory was considered in depth in Section 3.2 pro-
viding a context for effectively interpreting the motivations for the state of the art in
clonal selection algorithms. More importantly, the depth of immunological background
provided insight into the constraints that were imposed in realising the classical and state
of the art algorithms. The review of clonal selection algorithms in Section 3.3 highlighted
320
three main limitations, specifically: (1) Lack of explicit consideration of adaptive quali-
ties: Information processing inspired by structures and functions of the clonal selection
theory were traditional considered adaptive implicitly given the theories relationship with
evolutionary theory and resultant evolutionary algorithms. (2) Lack of positioning in the
broader field : Consideration of inspired algorithms was traditionally limited to related al-
gorithms in the field of Artificial Immune Systems with some consideration of the related
approaches in the field of Evolutionary Computation, broader consideration in computa-
tional and machine intelligence was rarely considered. (3) Lack of consideration beyond
the cellular : Clonal selection algorithms, as well as the majority of the broader Artificial
Immune System algorithms are concerned primarily with the structure and function of the
immune system at the cellular level, specifically given the cellular (lymphocyte immune
cell) and molecular (antibody protein) focus of the principle inspiring theories.
Each of these three limitations were address in turn, specifically: Section 3.4 consid-
ered clonal selection in the broader context of Computational and Machine Intelligence,
strengthening the relationship with Evolutionary Computation and Lazy Learning, and
highlighting the relationship with Hill Climbing algorithms and Competitive Learning.
Section 3.5 considered clonal selection as an adaptive strategy, phrasing the approach in
Holland’s adaptive systems formalism. Section 3.6 outlined an agenda for the investi-
gation and elaboration of the broader clonal selection paradigm, including: the Cellular
Clonal Selection perspective involving the investigation of the classical paradigm pursued
in Chapter 4, the new Tissue Clonal Selection perspective investigated in Chapter 5, and
the new Host Clonal Selection perspective investigated in Chapter 6.
3. Identify immunological structures and/or functions which clearly exhibit

distributed information processing. An important motivation for the project was
the identified promise of decentralised, autonomous, and distributed information process-
ing from in the broader field of Artificial Immune Systems reviewed in Section 2.5. The
review demonstrated that although such approaches had been proposed, few investigated
the intersection of clonal selection and distributed information processing explicitly. Sec-
tion 3.6 identified two distinct perspectives of clonal selection that exhibited distributed
information processing properties: clonal selection across the tissues of a host organism,
and clonal selection across the hosts in a population of organisms. The Tissue Clonal
Selection Paradigm was investigated in Chapter 5, involving a review of the motivating
metaphor in the physiology of the lymphatic system and the behaviour of migrating lym-
phocytes. The underlying strategy of Tissue Clonal Selection Algorithms was defined
as the localised although decentralised acquisition and application of information with
continuous dissemination toward general system capability. Toward this end, investiga-
tion of tissue algorithms focused on the localisation and dissemination of information in
unknown distributed exposure scenarios. The Host Clonal Selection Paradigm was inves-
tigated in Chapter 6, providing a review of the motivating metaphor in the immunisation
and evolutionary pressures on the immune system. The underlying trend of Host Clonal
Selection algorithms was defined as similar to that of Tissue Algorithms involving the
localisation and dissemination of information, with the increase in localisation providing
321
host-specialisation, and the variation of disseminated giving hosts control over what in-
formation is disseminated, and when. Toward this end, investigation of host algorithms
also focused on the information acquisition and application capabilities under a range of
unknown distributed exposure scenarios.
4. Study the behaviours, capabilities, and limitations of devised computa-

tional models. A model problem domain and problem instances were defined in Sec-
tion 4.2.3 called the Colour Space Domain suitable for testing and assessing clonal se-
lection algorithms with properties of optimisation and pattern recognition. Each clonal
selection paradigm was phrased in terms of (1) abstract models that outlined the motivat-
ing structural and/or functional metaphor and general computational concerns (Sections
4.2, 5.3, and 6.3 for the cellular, tissue, and host paradigms respectively), and (2) re-
alised algorithms, problems, and measures that outlined the specific system behavioural
and capability expectations (Sections 4.3, 5.4, and 6.4 for the cellular, tissue, and host
paradigms respectively). Clonal Selection Algorithm behavioural capabilities and expec-
tations were systematically demonstrated and verified empirically with specific algorithm
instantiations on specific problem instance scenarios across the three paradigms. Impor-
tantly, the verification process bounded the claims of expected capability, limiting them
to specific algorithms and/or problem scenarios. The suitability of application of the gen-
eral paradigms was assessed in Chapter 8. This involved a distillation of each approach
into distinctive information processing features based on observed and verified behaviours
(Section 8.2). The information processing features were then systematically mapped onto
specialised cases of Function Optimisation (Section 8.3.2) and Function Approximation
(Section 8.4.2) problem domains.
5. Propose an integrated clonal selection framework that unites the base and
distributed variants. The clonal selection paradigm was partitioned in section Sec-
tion 3.6 into that of ‘systems’ and ‘environments’, where clonal selection governed immune
systems were mapped to ‘systems’, and the scope of antigen and pathogen were mapped to
‘environments’ (elaborated initially in Section 4.2.2). The clonal selection (systems) parti-
tion was reduced into the three distinct paradigms, each of which was investigated. Chap-
ter 7 considered the reconstruction of the three paradigms and the system-environment
partition into a hierarchical framework. A Hierarchical Clonal Selection Framework was
defined in Section 7.2 that aggregated the clonal selection algorithms from across the
three paradigms, highlighting the specific constraints imposed on the quintessential in-
formation processing procedure. A Hierarchical Antigenic Exposure Framework was de-
fined in Section 7.3 that aggregated the contrived antigenic environments from across the
three paradigms, highlighting the commonality of piece-wise information exposure, and
the common source of complexity in the cardinality and spatial temporal patterns of ex-
posure, which are compounded across the paradigms. The frameworks were combined
into an Integrated Hierarchical Clonal Selection Framework in Section 7.4, uniting the
system and environmental concerns, suggesting at additional perspectives, and account-
ing for the bracketing of complexity in each specific clonal selection paradigm. Finally,
322
the adopted partition-reduction methodology was reviewed and its application considered
in the broader context in the field of Artificial Immune Systems, and as a general tool for
investigating related Computational Intelligence fields of study (Section 7.5).
9.2 Contributions
This section reviews the contributions as standalone conceptual and computational arte-
facts. Artefacts are defined as discrete and differentiable objects produced and qualified
within this investigation that may drawn from and used beyond the work.
9.2.1 Computational Artefacts

A computational artefact contribution is a defined formalism that may be exploited beyond
this investigation for computation, including tools such as metaheuristics or solvers and
benchmark problems.
1. Cellular Clonal Selection: A generic paradigm for unifying existing cellular-based

clonal selection algorithms (Chapter 4), and promoting the development of novel
algorithms based on inter-cellular interactions, specifically the Spatial Clonal Selec-
tion Algorithm (Section 4.5), Mediated Clonal Selection Algorithm (Section 4.6), and
the integration of the network metaphor embodied in the Network Clonal Selection
Algorithm (Section 4.7). These algorithms are generally suited to global function
optimisation and vector quantisation based function approximation using a holistic
inductive model of a given problem (Sections 8.3.2 and 8.4.2).
2. Tissue Clonal Selection: A distributed clonal selection paradigm inspired by the

lymphatic system and lymphocyte recirculation (Chapter 5), including the follow-
ing specific decentralised algorithms: the Recirculation (Section 5.5), Homing (Sec-
tion 5.6), and Inflammation (Section 5.7) Tissue Clonal Selection Algorithms. These
algorithms are generally suited to cooperative function optimisation and approxima-
tion with hybrid approaches and/or partial inductive models based on functional
decompositions of the problem (Sections 8.3.2 and 8.4.2).
3. Host Clonal Selection: A distributed clonal selection paradigm inspired by popu-

lations of organisms with immune systems undergoing immunisation and evolution
(Chapter 6), including the following specific decentralised algorithms: Transmission
(Section 6.5) and Shared Immunity (Section 6.6) Population Host Clonal Selection
Algorithms generally suited to parallel and cooperative function optimisation and
ensemble function approximation, and the Maternal (Section 6.7) and Evolved Im-
munity (Section 6.8) Generational Host Clonal Selection Algorithms generally suited
to multiple restart and adaptation of process for both optimisation and function ap-
proximation problems (Sections 8.3.2 and 8.4.2).
323
9.2.2 Conceptual Artefacts
A conceptual artefact contribution is a defined formalism that may be exploited beyond
this investigation for structuring practices, including methodologies and frameworks.
1. Systematic Methodology: An integrated set of practices derived from best practices

for realising biologically inspired computational intelligence models from metaphor
to computational artefact, including an immunological-centric conceptual framework
(Section 2.7.1), an experimental framework (Section 2.7.2), an engineering-centric
integration framework (Section 2.7.3), a partition and reduction framework for in-
vestigation a specific information processing strategy (Sections 3.6 and 7.5.1), and a
feature-centric suitability framework for assessing applicability, comparability, and
transferability (Section 8.2.1).
2. Hierarchical Framework : A conceptual scaffold were the capabilities and complexity

of a given constrained perspective of the information processing principle is sub-
sumed by the successive level (Chapter 7), used to integrate clonal selection algo-
rithms (Section 7.2), antigenic exposure problems (Section 7.3), integration of these
two frameworks into a unified hierarchical model of clonal selection (Section 7.4),
with suggested broader applicability in for other immunological-based computational
intelligence approaches (Sections 7.5.2).
9.3 Limitations
Limitations are observed as decision points with alternative options, each of which repre-
sents a trade-off in terms of the contributions made from the investigation in the context
of the hypothesis and research goals. This section critically assesses the limitations of
this work, specifically in the context of design and methodological constraints imposed on
the investigation, as well as the alternatives, benefits, and restrictions they provide. The
limitations of three specific decision points are considered: (1) the method of biological
interpretation and transition to computational intelligence algorithm, (2) the focus in the
chosen method of algorithm development, and (3) the methodology adopted to verify and
demonstrate expected behaviours and capabilities of developed algorithms.
9.3.1 Biological Interpretation

The motivating immunological metaphor for the Cellular, Tissue, and Host Clonal Selec-
tion Paradigms was based on a qualitative interpretation of the structure and function
of the acquired immune system at different scales resulting in generalised models and al-
gorithms. A problem with this approach is that the behavioural expectations of realised
models was also qualitative, and therefore difficult to assess and verify. An alternative
approach which is common in the field of Artificial Immune Systems is to use mathemati-
cal models of immunological function as a basis for computational intelligence models and
algorithms, for example the shape-space and affinity landscape geometric formalisms that
324
underlie the principles of much of the field (Section 3.2.2). This alternative reduces up-
front effort by exploiting interpretations made by theoretical immunologists and produces
specific rather than general approaches with a clear mathematical foundation. The strong
specialisation of such approaches likely limits the broader applicability of the approaches,
and the inspiring models are likely bounded by immunological dogma (for example self-
nonself discrimination) which may or may not be a useful perspective in computational
intelligence modelling and algorithm development. A specific and relevant example from
this work is the tissue paradigm which suggests at multiple tissues communicating in se-
ries (general), whereas the alternative may be to base the organisation of tissues on a
specific model of the lymphatic system [19], or a communication mechanism based on
mathematical models of lymphocyte trafficking [371, 370, 369]. The qualitative interpre-
tation approach taken in this work was chosen as it facilitates the conception of broader
fields of study (paradigms) which may be useful in and of themselves, and which may be
specialised at a later time with restrictive mathematical models.
9.3.2 Approach Development

The development of the clonal selection algorithms was focused on the strategy and the
constraints to its information processing properties resulting in the broadly conceived
algorithms and paradigms for grouping algorithms without an explicit problem domain
application. A standard alternative is to take a problem centric perspective of the devel-
opment of approaches, and specialise an existing strategy or paradigm for a given problem
domain or instance. The benefit of this alternative is that it results in specific and com-
parable findings for a specialised algorithm on a known difficult problem class, directly
contributing to an assessment of the relative utility of the approach. The strong problem
focus of this alternative in turn imposed strong limitations on the broader applicability
of an approach given its specialisation to the features and available information regarding
the problem. A specific example from his work is the foundational cellular clonal selec-
tion algorithms developed in the context of global function optimisation (such as BCA,
the IA family, and to a lesser extent CLONALG). Although this class of algorithm is
well suited for this application, the single-antigen perspective of the strategy ignores the
potential application of the approach to function approximation, and has the compound
effect of bounding the applicability of algorithms that subsume the cellular level1 . The
strategy focus in the development of approaches was chosen in the spirit of the goals of
Computational Intelligence toward investigating generalised unconventional methods of
computation, specifically bound to the information processing qualities of clonal selection.
The generality of the developed approaches may be bound at a later time through spe-
cialisations of the algorithms in the context of known problem instances to which they are
highly suited.
1
The field of Evolutionary Computation is generally bound by the first application of the approach
to function optimisation [121], that in turn has generally limited subsumed approaches such as Parallel
Evolutionary Algorithms.
325
9.3.3 Verification and Demonstration
The verification of the behaviours and capabilities of developed modes as achieved em-
pirically using experimentation with general algorithm instances on a generalised and
trivial pattern recognition problem domain. A problem with this approach of verifica-
tion and demonstration is that it is qualitative and inaccurate, specifically in the context
of Goldberg’s economy of modelling in Section 2.7.3. An alternative is to demonstrate
the capability and verify behaviour using mathematical modelling tools. Analytical tools
would provide generalised descriptions of relationships and behaviours that would be more
accurate although a more restricted interpretation of the motivating information process-
ing properties. An alternative to the development of costly mathematical models is the
approach outlined by Goldberg’s is to use a spectrum of modelling types in addition to em-
pirical assessment and statistical analysis, specifically facet-wise and dimensional models.
The qualitative empirical approach was chosen primarily given the speed both in imple-
mentation and in obtaining results. Empirical demonstration required implementation
providing a grounding of conceptual models to the specific concerns of a functioning infor-
mation processing system. The broader inaccuracy of qualitative observations was address
through the use of multiple problem scenarios providing calibration and cross-correlation
of behaviour, and through the use of statistical tools.
9.4 Further Research

This section proposes recommendations for further research and improvements on the
work from the investigation. Extensions are partitioned as follows: (1) framework in
which additional related metaphors are proposed as a basis for new paradigms within the
integrated hierarchical clonal selection framework, (2) paradigms in which additional and
related constrained metaphors are considered within the context of each clonal selection
paradigm as the basis for new algorithms and architectures, (3) and algorithms in which
elaborations and improvements on algorithms proposed in this work are considered. Many
of the proposed extensions, particularly the alternate and additional motivating metaphors
were preliminary investigated as apart of this work, although were not integrated into the
framework given the constraints imposed on the project.
9.4.1 Framework Extensions

Integrated frameworks were considered in Chapter 7 which proposed a hierarchical clonal
selection, antigenic exposure, and integrated clonal selection frameworks. The integration
of the three clonal selection paradigms highlighted the following extensions that may be
pursued beyond this work:
1. Additional Configurations: The investigation of clonal selection systems beyond the

simplest configuration (such as minimum component or interaction cardinality) in
what was referred to as varied interpretation of clonal selection models, as high-
lighted by the bracketing analysis of the integrated framework (Section 7.4.2). This
326
investigation would immediately provide new interpretations of clonal selection al-
gorithms from across the three paradigms, and reveal insights into the behaviour
and application of clonal selection algorithms covered by the framework.
2. Additional Tiers: The elicitation and investigation of systems motivated by con-

straints on the information processing properties of clonal selection below the cellular
level (such as the molecular concerns of epitopes and paratopes) and above the host
level (such as interacting species of organisms) (Section 7.4.2). These investigations
would reveal insights into the scope of sub-cellular and super-host bracketing already
imposed by the algorithms covered by the framework, as well as provide a basis for
new constraints and interpretations of clonal selection as a basis for Computational
Intelligence algorithms.
3. Additional Mappings: The investigation of the clonal selection centric acquired im-
mune system models across the hierarchies biased by one or more of the related
computational paradigms (such as negative selection, immune network, or danger
theory) as remarked in the bracketing analysis (Section 7.4.3), and explicitly con-
sidered with regard to the integrated framework in Section 7.5.2. This investigation
would not only provide insights into the approaches covered by the framework, it
would explicitly merge related immunological computational paradigms immediately
providing a firm basis of models, architectures, and algorithms to be exploited by
the integrated paradigm.
In addition to improving and augmenting the integrated clonal selection framework,

additional holistic metaphors may be incorporated toward providing further general in-
sights into the computational strategy, as well as additional metaphors for information
processing, as follows:
1. Asexual Evolution: Clonal Selection Theory describes a process of asexual evolution,

akin to the evolution of micro-organisms not limited to bacteria, virus, unicellular
organisms, as well as multi-cellular asexual reproduces. This obvious connection was
made by Burnet in his monograph on the theory where he described the relation-
ship of the principles of clonal selection with bacteria and cancer cells [64]. Much is
known regarding the genetic and adaptive models of asexual evolution, for example
advantagious and deleterious mutation rates in populations [169, 340], which may
be interpreted and integrated into an improved computational model of B-cell re-
ceptor evolution in the cellular paradigm. This fundamental work would not only
provide insights into the configuration and application of cellular algorithms, given
the cellular basis of all three paradigms (and the high-level constraints imposed
on the cellular mechanisms) it would provide insights for all clonal selection based
approaches.
2. Pathogen Models: The antigenic exposure paradigm was defined as an abstract prob-
lem for clonal selection systems encapsulating the properties of all self and foreign
327
antigen (defined in Section 4.2.2, elaborated in Sections 5.3.2 and 6.3.2, and inte-
grated in Section 7.3). As such, the paradigm received little elaboration beyond
definition providing a large opportunity for the investigation of pathogenic models.
Pathogen models may be elicited and investigated as either models for problems to
be solved by clonal selection systems, or as distinct computational models for prob-
lem solving in and of themselves. Three clear motivating metaphors for the latter
case are as follows: (1) Pathogen Evolution: Suggested in Section 6.3.2 as providing
a model for dynamic problems in which a system and problem co-adapt in response
to each other based on reviewed parasitism in Section 6.2.2, (2) Intra-Host Dynam-
ics: That describe the strategies and behaviours of a pathogen after it enters a host
organism including evasive and adaptive qualities [26, 161] (3) Intra-Population Dy-
namics: Also suggested in Section 6.3.2 and providing part of the motivation for
system-controlled transmission in the PT-HCSA (Section 6.5)2 .
3. Cancer : Uniting asexual somatic evolution and an elaborated self-antigen exposure

paradigm, cancer provides a motivation for a computational paradigm that may be
investigated independently or as a complement clonal selection. Cancer provides
a co-evolutionary problem model for clonal selection in which the accumulation of
mutations is exploited for survival of tumour cells. The acquired immune system
must discriminate the endogenous antigen and (at the very least) counter the adap-
tations to surface structures. More interestingly, cancer also provides an intra-host
Darwinian microcosm [409, 56, 410], that can aggressively recruit resources such as
cells and nutrient sources, disseminate throughout the host (such as the metastasis
of malignant tumour cells), and in some cases can be transmitted between hosts
[299].
9.4.2 Paradigm Extensions

This section enumerates the three clonal selection paradigms and highlights elaborations
of proposed models as additional motivating metaphors for new computational models,
not specific to any given algorithm within the paradigms.
1. Cellular Paradigm: Likely the most important of the three paradigms, the cellular
paradigm may be elaborated in three specific directions: (1) lymphocyte life cycle
and types, and (2) regulatory mechanisms. The model of lymphocytes as atomic
data points is a dramatic simplification of both the complexity of the life cycle of
immune cells, as well as the diverse variety of cells that are involved in the immune
response. The computational properties of cellular life cycles introduces task-division
and developmental mechanism (Section 5.2.2), whereas the exploitation of the di-
verse set of immune cell types further promotes task division and more importantly
interaction and collaboration between cells. The regulation of lymphocytes under
clonal selection in which the cellular repertoire is continually acquiring information
2
There is much work on epidemiology, specifically adaptive transmission [50], and adaptive virulence
[335], to motivate this paradigm.
328
and expanding, presents problems of cell population size regulation whilst maintain-
ing an effective immunological memory. More elaborate homoeostasis mechanisms
should be investigated toward providing improved flexibility in cellular algorithms, in
particular with regard to promoting decentralised rather than centralised resource
allocation in response to antigenic exposures within a given repertoire. Such ap-
proaches may be motivated by the variety of mathematical models that describe
lymphocyte homoeostasis [290, 87].
2. Tissue Paradigm: Two general areas for elaboration of the tissue paradigm include
(1) the investigation of the variety of lymphocyte trafficking behaviours and the de-
centralised receptor centric mechanisms that influence them (Section 5.2.2) and (2)
the tissues of the lymphatic system and the behavioural restrictions and antigenic-
interactions promoted by the different tissue structures and arrangements (reviewed
in Section 5.2.1, with models proposed in Section 5.3.3). A final additional motivat-
ing metaphor is the emerging field of Lymphocyte Population Biology, in which the
behaviours and interactions of immune cells are studied using the tools of population
genetics and ecology, driven more recently by Freitas Antonio and colleagues at the
Pasteur Institute in France [154, 165, 4]. This aggregated perspective of lympho-
cyte behaviours introduces interesting models of lymphocyte niche exploitation and
regulatory concerns of density dependent competition.
3. Host Paradigm: The two principle areas of the host paradigm present a large ca-
pacity for elaboration given the generality of base motivations, specifically (1) host
population immunisation (reviewed in Section 6.2.1), and (2) host immune system
evolution (reviewed in Section 6.2.2). Host vaccination provides a promising area
for elaboration, specificly the computational concerns of vaccine identification and
transmission regimes in the context of the benefits of population coverage called
‘herd immunity’. Such investigations are expected to provide models for informa-
tion dissemination in distributed databases such as database replication [127], and
peer-to-peer content dissemination networks [413] suggested in Section 8.4.2. The
evolution of the immune system is a broader motivating metaphor, although fur-
ther research may consider the following two promising cases: (1) the adaptation
of innate immunity where a selective process motivates the integration of lifetime
acquired traits into the genome (the intersection of lifetime and evolutionary learn-
ing considered in Section 6.3.1), and (2) the adaptation of the affinity maturation
process beyond the base repertoire where the mechanisms used in lifetime are also
evolved.
9.4.3 Algorithm Extensions

The investigation of algorithms in this work was broad in coverage although shallow in
depth, providing plenty of opportunity for further ‘basic science’. Two general classes of
fundamental research are highlighted regarding the proposed algorithms: (1) application
of algorithms as outlined in Chapter 8, and (2) investigation into the predictable elicita-
329
tion of information processing capabilities not limited to parameter studies, specialisation,
and additional empirical benchmarking. This section highlights specific cases of the sec-
ond class of fundamental research for specific algorithms across the three clonal selection
paradigms.
1. Cellular Algorithms: Three specific cellular algorithms are in need of elaboration to

capitalise on their expected potential: (1) degenerate, (2) mediated, and (3) network.
Cohen’s notions of emergent specificity through degenerate pattern recognition cells
in different contexts (reviewed in Section 4.4.3) provides the potential for a pow-
erful sub-symbolic population-based adaptive paradigm, requiring investigation into
viable representation and aggregation mechanisms. The two-signal model of lympho-
cyte activation provides an important capability for the system to decide to which
signals to respond and not respond. The mediated algorithm requires further investi-
gation into mapping mechanisms for affinity-based activation potentials between cell
populations (see Section 4.6). Finally, the the immune network paradigm has moti-
vated many immuno-cognitive models, although the specific approach suggested in
this work was an explicit pattern recognition method akin to Learning Classifier Sys-
tems, requiring much work into representation and mapping schemes (Section 4.7).
An opportunistic improvement on the cell-cell interaction models would be to adopt
a LCS architecture, representation, and credit assignment mechanism and bias the
approaches by immune network behaviours.
2. Tissue Algorithms: The tissue algorithms, specifically the use of recirculation and
mechanisms to impose localisation biases on recirculation (such as homing and in-
flammation) would benefit from analytical analysis. Specifically, (1) probabilistic
models of information availability across a given structure and how such models are
effected by localisation biases, and (2) probabilistic models of information dissemi-
nation based on varied exposure models. These models would motivate predictable
configurations (such as migration or imprinting rates) for eliciting well defined system
capability under different exposure scenarios supporting empirical findings. Further,
an investigation into the field of Spatial-Temporal Learning and the relationship
of Tissue (and Host) Algorithms would provide a deeper theoretical understanding
of the fundamental problem and recirculation-based strategy solution of systemic
immunity within a host [354].
3. Host Algorithms: Mathematical models of vaccination and virus transmission are

available with which to bias the configuration of shared immunity and transmis-
sion host clonal selection algorithms. Such models would provide predictable intra-
population interactions allowing focus of these algorithms to shift from base inter-
action mechanism toward more interesting population-based information acquisition
and application scenarios (mixtures of experts and ensembles). A similar situation
exists with the generational host algorithms that must overcome the fundamentals
predictable inter-generational interaction before investigating the interesting and
desirable effects of cross-generational information acquisition. This highlights that
330
the compounded complexity effect of the host level of algorithms that are both de-
pendent on predictable algorithms which they subsume, as well as the predictable
foundational behaviours (such as interaction) of the paradigm.
9.5 Final Words

This work has demonstrated that a holistic perspective of the immune system as a compu-
tational metaphor provides a fertile ground for adaptive and distributed Artificial Immune
Systems, and Clonal Selection Algorithms specifically. The culmination of this work ad-
vocates a broader perspective of the acquired immune system beyond a classical cellular
focus as a viable path toward a so-called second generation of Artificial Immune Systems.
This path not only facilitates, but requires the currently accepted approach of turning
back to the motivating metaphor as a strategy for developing systems that incorporate
more detailed information processing models of immunology. The computational focus of
immunophysiology and the integrated framework presented in this work provide a general
scaffold for such work, the ongoing investigation of which promotes a multifaceted Compu-
tational Intelligence research agenda of (1) basic (foundational) research, (2) exploratory
research, and (3) application-focused research.
331
Appendix A
Tables of Empirical Results
A.1 Tissue Clonal Selection

Tables of results from the empirical investigations undertaken in Chapter 5.
A.1.1 Recirculation Empirical Study

Table A.1 summarises the results from the empirical study in Section 5.5.3 investigating
the Recirculation Tissue Clonal Selection Algorithm (RTCSA). The non-parametric Mann-
Whitney U statistical test was calculated pair-wise for all algorithms on each TER group.
332
Problem System HD HE ATD ATE
TER TCSA x̄ σ x̄ σ x̄ σ x̄ σ
ATER MTCSA 1610.68 78.407 0.013 0.006 85.903 0.92 0.155 0.011
ATER RTCSA-SML 733.841 57.614 0.131 0.021 78.82 2.812 0.148 0.017
ATER RTCSA-MED 608.013 45.143 0.149 0.024 76.792 3.723 0.153 0.018
ATER RTCSA-LRG 562.136 54.663 0.152 0.025 76.969 4.213 0.152 0.017
Significant True Truea Trueb False
OTER MTCSA 1708.99 89.685 0.033 0.01 84.394 1.151 0.152 0.01
OTER RTCSA-SML 690.006 51.553 0.122 0.022 78.947 4.497 0.137 0.021
OTER RTCSA-MED 602.637 38.677 0.141 0.024 77.86 4.731 0.145 0.018
OTER RTCSA-LRG 561.076 47.781 0.15 0.021 78.593 4.583 0.145 0.02
Significant True Truec Trued Falsee
PTER MTCSA 876.922 40.518 0.093 0.026 92.366 0.442 0.158 0.01
PTER RTCSA-SML 690.785 22.011 0.148 0.019 93.192 0.182 0.172 0.018
PTER RTCSA-MED 635.74 21.472 0.162 0.024 93.515 0.129 0.173 0.018
PTER RTCSA-LRG 572.749 14.306 0.167 0.021 93.826 0.118 0.169 0.017
Significant True Truef True Trueg
RTER MTCSA 2313.88 100.596 0.121 0.021 74.7 1.777 0.123 0.01
RTER RTCSA-SML 602.977 71.514 0.11 0.018 76.113 4.058 0.11 0.016
RTER RTCSA-MED 561.37 55.816 0.109 0.015 76.277 4.166 0.115 0.017
RTER RTCSA-LRG 503.919 65.54 0.114 0.024 75.347 4.988 0.111 0.019
Significant True False Falseh Falsei
STER MTCSA 2142.95 160.562 0.026 0.009 76.972 2.421 0.093 0.017
STER RTCSA-SML 936.657 125.483 0.046 0.013 79.098 2.812 0.087 0.021
STER RTCSA-MED 672.651 72.745 0.065 0.019 76.659 3.648 0.098 0.024
STER RTCSA-LRG 579.055 62.934 0.07 0.019 76.445 3.952 0.094 0.023
Significant True Truej Truek Falsel
a
False for RTCSA-MED and RTCSA-LRG
b
c
d
False for RTCSA-SML and RTCSA-MED, False for RTCSA-SML and RTCSA-LRG, False for RTCSA-
MED and RTCSA-LRG
e
True for MTCSA and RTCSA-SML
f
g
False for RTCSA-SML and RTCSA-MED, False for RTCSA-SML and RTCSA-LRG, False for RTCSA-
MED and RTCSA-LRG
h
True for MTCSA and RTCSA-SML, True for MTCSA and RTCSA-MED
i
True for MTCSA and RTCSA-SML, True for MTCSA and RTCSA-LRG
j
k
False for MTCSA and RTCSA-MED, False for MTCSA and RTCSA-LRG, False for RTCSA-MED
and RTCSA-LRG
l
True for RTCSA-SML and RTCSA-MED
Table A.1: Summary of results from an empirical study into the RTCSA compared to the
MTCSA.
Table A.2 summarises the results of the RTCSA for the analysis of the approach in
Section 5.5.3.
333
System Problem HD HE ATD ATE
TCSA TER x̄ σ x̄ σ x̄ σ x̄ σ
MTCSA ATER 1610.68 78.407 0.013 0.006 85.903 0.92 0.155 0.011
MTCSA OTER 1708.99 89.685 0.033 0.01 84.394 1.151 0.152 0.01
MTCSA PTER 876.922 40.518 0.093 0.026 92.366 0.442 0.158 0.01
MTCSA RTER 2313.88 100.596 0.121 0.021 74.7 1.777 0.123 0.01
MTCSA STER 2142.95 160.562 0.026 0.009 76.972 2.421 0.093 0.017
Significant True True True Truea
RTCSA-LRG ATER 562.136 54.663 0.152 0.025 76.969 4.213 0.152 0.017
RTCSA-LRG OTER 561.076 47.781 0.15 0.021 78.593 4.583 0.145 0.02
RTCSA-LRG PTER 572.749 14.306 0.167 0.021 93.826 0.118 0.169 0.017
RTCSA-LRG RTER 503.919 65.54 0.114 0.024 75.347 4.988 0.111 0.019
RTCSA-LRG STER 579.055 62.934 0.07 0.019 76.445 3.952 0.094 0.023
Significant Falseb Truec Trued Truee
a
False for OTER and ATER, PTER and ATER
b
True for RTER and ATER, RTER and OTER, RTER and PTER, STER and RTER
c
False for ATER and OTER
d
False for ATER and OTER, RTER and ATER, STER and ATER, STER and OTER, STER and
RTER
e
False for OTER and ATER
Table A.2: Summary of results from an empirical study into the RTCSA comparing TER’s
with the MTCSA local and RTCSA-LRG global information management strategies.
A.1.2 Homing Empirical Study

the Homing Tissue Clonal Selection Algorithm (HTCSA). The non-parametric Mann-
334
ATER MTCSA 1610.68 78.407 0.013 0.006 85.903 0.92 0.155 0.011
ATER RTCSA-S 733.841 57.614 0.131 0.021 78.82 2.812 0.148 0.017
ATER HTCSA-S 747.071 52.578 0.099 0.019 81.631 3.123 0.142 0.016
ATER HTCSA-M 816.023 53.374 0.084 0.019 83.046 2.228 0.15 0.018
ATER HTCSA-L 816.023 53.374 0.084 0.019 83.046 2.228 0.15 0.018
Significant Truea Trueb Truec Falsed
OTER MTCSA 1708.99 89.685 0.033 0.01 84.394 1.151 0.152 0.01
OTER RTCSA-S 690.006 51.553 0.122 0.022 78.947 4.497 0.137 0.021
OTER HTCSA-S 737.646 81.211 0.099 0.02 79.893 3.451 0.136 0.018
OTER HTCSA-M 756.301 64.802 0.1 0.019 81.077 3.568 0.146 0.015
OTER HTCSA-L 756.301 64.802 0.1 0.019 81.077 3.568 0.146 0.015
Significant Truee Truef Trueg Falseh
PTER MTCSA 876.922 40.518 0.093 0.026 92.366 0.442 0.158 0.01
PTER RTCSA-S 690.785 22.011 0.148 0.019 93.192 0.182 0.172 0.018
PTER HTCSA-S 678.451 20.88 0.136 0.019 93.232 0.137 0.169 0.016
PTER HTCSA-M 685.395 27.472 0.145 0.017 93.206 0.161 0.17 0.018
PTER HTCSA-L 685.395 27.472 0.145 0.017 93.206 0.161 0.17 0.018
Significant True True Falsei Falsej
RTER MTCSA 2313.88 100.596 0.121 0.021 74.7 1.777 0.123 0.01
RTER RTCSA-S 602.977 71.514 0.11 0.018 76.113 4.058 0.11 0.016
RTER HTCSA-S 635.615 52.025 0.109 0.021 76.828 3.359 0.113 0.017
RTER HTCSA-M 673.063 58.672 0.114 0.022 75.595 3.745 0.114 0.018
RTER HTCSA-L 673.063 58.672 0.114 0.022 75.595 3.745 0.114 0.018
Significant Truek Falsel Falsem Falsen
STER MTCSA 2142.95 160.562 0.026 0.009 76.972 2.421 0.093 0.017
STER RTCSA-S 936.657 125.483 0.046 0.013 79.098 2.812 0.087 0.021
STER HTCSA-S 984.195 130.37 0.045 0.014 79.29 2.245 0.086 0.023
STER HTCSA-M 995.975 114.778 0.046 0.013 79.731 2.071 0.09 0.022
STER HTCSA-L 995.975 114.778 0.046 0.013 79.731 2.071 0.09 0.022
Significant Falseo Falsep Falseq False
a
False for RTCSA-S and HTCSA-S, HTCSA-M and HTCSA-L
b
False for HTCSA-M and HTCSA-L
c
False for HTCSA-S and HTCSA-M, HTCSA-S and HTCSA-L, HTCSA-M and HTCSA-L
d
True for MTCSA and HTCSA-S
e
f
g
False for RTCSA-S and HTCSA-S, HTCSA-S and HTCSA-M, HTCSA-S and HTCSA-L, HTCSA-M
and HTCSA-L
h
True for MTCSA and RTCSA-S, MTCSA and HTCSA-S, HTCSA-S and HTCSA-M, HTCSA-S and
HTCSA-L
i
True for MTCSA and RTCSA-S, MTCSA and HTCSA-S, MTCSA and HTCSA-M, MTCSA and
HTCSA-L
j
HTCSA-L
k
False for RTCSA-S and HTCSA-S, HTCSA-M and HTCSA-L
l
True for MTCSA and HTCSA-S
m
True for MTCSA and RTCSA-S, MTCSA and HTCSA-S
n
HTCSA-L
o
HTCSA-L
p
HTCSA-L
q
HTCSA-L
Table A.3: Summary of results from an empirical study into the HTCSA.
335
A.1.3 Inflammation Empirical Study
the Inflammation Tissue Clonal Selection Algorithm (ITCSA). The non-parametric Mann-

ATER MTCSA 1610.68 78.407 0.013 0.006 85.903 0.92 0.155 0.011
ATER RTCSA-S 733.841 57.614 0.131 0.021 78.82 2.812 0.148 0.017
ATER ITCSA-N 1885.64 81.789 0.014 0.007 86.034 0.846 0.149 0.01
ATER ITCSA-S 800.373 70.917 0.122 0.027 80.34 3.392 0.138 0.019
Significant True True Truea True
OTER MTCSA 1708.99 89.685 0.033 0.01 84.394 1.151 0.152 0.01
OTER RTCSA-S 690.006 51.553 0.122 0.022 78.947 4.497 0.137 0.021
OTER ITCSA-N 4298.85 145.475 0.061 0.026 36.523 5.666 0.223 0.015
OTER ITCSA-S 977.768 151.773 0.125 0.022 58.817 11.181 0.16 0.027
Significant True Trueb True Truec
PTER MTCSA 876.922 40.518 0.093 0.026 92.366 0.442 0.158 0.01
PTER RTCSA-S 690.785 22.011 0.148 0.019 93.192 0.182 0.172 0.018
PTER ITCSA-N 1186.91 88.006 0.151 0.032 89.802 0.958 0.164 0.011
PTER ITCSA-S 769.419 37.061 0.13 0.023 93.063 0.24 0.169 0.017
Significant True Trued True Truee
RTER MTCSA 2313.88 100.596 0.121 0.021 74.7 1.777 0.123 0.01
RTER RTCSA-S 602.977 71.514 0.11 0.018 76.113 4.058 0.11 0.016
RTER ITCSA-N 3300.08 188.667 0.132 0.02 64.262 3.485 0.128 0.009
RTER ITCSA-S 742.331 84.362 0.091 0.016 73.855 4.428 0.093 0.012
Significant True Truef Trueg True
STER MTCSA 2142.95 160.562 0.026 0.009 76.972 2.421 0.093 0.017
STER RTCSA-S 936.657 125.483 0.046 0.013 79.098 2.812 0.087 0.021
STER ITCSA-N 3711.46 130.433 0.066 0.013 51.85 3.256 0.149 0.022
STER ITCSA-S 1331.72 189.05 0.066 0.014 64.593 4.886 0.123 0.02
Significant True Trueh True Truei
a
False for MTCSA and ITCSA-N
b
False for RTCSA-S and ITCSA-S
c
False for MTCSA and ITCSA-S
d
False for RTCSA-S and ITCSA-N
e
False for RTCSA-S and ITCSA-N, False for RTCSA-S and ITCSA-S, False for ITCSA-N and ITCSA-S
f
False for MTCSA and RTCSA-S
g
False for MTCSA and ITCSA-S
h
False for ITCSA-N and ITCSA-S
i
False for MTCSA and RTCSA-S
Table A.4: Summary of results from an empirical study into the ITCSA.
A.2 Host Clonal Selection

Tables of results from the empirical investigations undertaken in Chapter 6.
A.2.1 Transmission Empirical Study

the Transmission Host Clonal Selection Algorithm (THCSA). The non-parametric Mann-
336
Whitney U statistical test was calculated pair-wise for all algorithms on each HER group.
Problem System PH PE AHD AHE

HER HCSA x̄ σ x̄ σ x̄ σ x̄ σ
AHER MHCSA 1606.24 57.906 0.013 0.007 85.97 0.702 0.157 0.01
AHER PT-RP 2257.49 92.704 0.028 0.01 75.558 1.552 0.116 0.011
AHER PT-SP 2007.71 68.019 0.018 0.008 79.905 1.037 0.138 0.011
AHER VT-L 2218.74 103.049 0.023 0.01 76.433 1.72 0.127 0.011
AHER VT-S 1952.49 45.303 0.015 0.008 80.94 0.692 0.141 0.011
Significant Truea Trueb True Truec
OHER MHCSA 1792.98 99.787 0.031 0.011 83.271 1.398 0.152 0.009
OHER PT-RP 2269.86 87.319 0.042 0.015 75.496 1.48 0.122 0.011
OHER PT-SP 2098.97 74.909 0.043 0.015 78.482 1.156 0.134 0.01
OHER VT-L 2254.70 88.167 0.037 0.013 75.797 1.508 0.126 0.012
OHER VT-S 2015.66 95.136 0.038 0.011 79.928 1.476 0.138 0.009
Significant Trued Falsee Truef Trueg
PHER MHCSA 894.093 40.821 0.102 0.03 92.315 0.493 0.159 0.009
PHER PT-RP 1716.58 88.816 0.102 0.03 84.24 1.175 0.147 0.01
PHER PT-SP 1251.29 50.971 0.102 0.03 88.887 0.64 0.152 0.009
PHER VT-L 2193.93 94.627 0.102 0.03 76.966 1.621 0.131 0.011
PHER VT-S 1716.78 67.258 0.102 0.03 84.268 0.955 0.147 0.011
Significant Trueh False Truei Truej
RHER MHCSA 2341.24 72.031 0.122 0.023 74.238 1.264 0.123 0.009
RHER PT-RP 2353.58 106.255 0.118 0.023 73.869 1.878 0.117 0.011
RHER PT-SP 2368.57 141.087 0.112 0.019 73.619 2.647 0.116 0.011
RHER VT-L 2369.23 118.525 0.123 0.022 73.59 2.095 0.118 0.012
RHER VT-S 2364.53 115.166 0.125 0.018 73.853 2.108 0.124 0.009
Significant False Falsek False Falsel
SHER MHCSA 2144.72 172.247 0.028 0.012 76.929 2.717 0.094 0.019
SHER PT-RP 2150.92 187.501 0.029 0.009 76.78 2.886 0.093 0.019
SHER PT-SP 2161.80 194.783 0.029 0.011 76.603 3.084 0.093 0.018
SHER VT-L 2164.46 178.009 0.026 0.009 76.583 2.847 0.094 0.018
SHER VT-S 2162.90 173.729 0.026 0.011 76.697 2.77 0.093 0.018
Significant False False False False
a
False for PT-RP and VT-L
b
False for MHCSA and VT-S, PT-RP and VT-L, PT-SP and VT-S
c
False for PT-SP and VT-S
d
e
True for MHCSA and PT-RP, MHCSA and PT-SP, MHCSA and VT-S
f
g
False for PT-RP and VT-L, PT-SP and VT-S
h
False for PT-RP and VT-S
i
False for PT-RP and VT-S
j
False for PT-RP and PT-SP, PT-RP and VT-S, PT-SP and VT-S
k
True for PT-SP and VT-L, PT-SP and VT-S
l
True for MHCSA and PT-SP, PT-RP and VT-S, PT-SP and VT-S
Table A.5: Summary of results from an empirical study into the THCSA compared to
MP-HCSA.
Table A.6 summarises the pathogen transmission and vaccination transmission results
from an empirical study into the THCSA from Section 6.5.2.
337
AHER PT-RP 2257.49 92.704 0.028 0.01 75.558 1.552 0.116 0.011
AHER VT-L 2218.74 103.049 0.023 0.01 76.433 1.72 0.127 0.011
Significant False False True True
OHER PT-RP 2269.86 87.319 0.042 0.015 75.496 1.48 0.122 0.011
OHER VT-L 2254.70 88.167 0.037 0.013 75.797 1.508 0.126 0.012
PHER PT-RP 1716.58 88.816 0.102 0.03 84.24 1.175 0.147 0.01
PHER VT-L 2193.93 94.627 0.102 0.03 76.966 1.621 0.131 0.011
Significant True False True True
RHER PT-RP 2353.58 106.255 0.118 0.023 73.869 1.878 0.117 0.011
RHER VT-L 2369.23 118.525 0.123 0.022 73.59 2.095 0.118 0.012
SHER PT-RP 2150.92 187.501 0.029 0.009 76.78 2.886 0.093 0.019
SHER VT-L 2164.46 178.009 0.026 0.009 76.583 2.847 0.094 0.018
Table A.6: Summary of the pathogen transmission and vaccination transmission results
from an empirical study into the THCSA.
A.2.2 Shared Immunity Empirical Study

the Shared Immunity Host Clonal Selection Algorithm (SI-HCSA). The non-parametric
Mann-Whitney U statistical test was calculated pair-wise for all algorithms on each HER
group.
338
AHER MHCSA 1606.24 57.906 0.013 0.007 85.97 0.702 0.157 0.01
AHER SI-SS 1021.50 53.326 0.012 0.005 88.691 0.593 0.082 0.011
AHER SI-SL 660.944 45.926 0.03 0.01 88.719 0.893 0.086 0.011
AHER SI-LS 679.249 48.4 0.028 0.012 88.249 1.101 0.086 0.017
AHER SI-LL 418.271 35.337 0.076 0.013 84.116 2.518 0.104 0.012
Significant Truea Trueb Truec Trued
OHER MHCSA 1792.98 99.787 0.031 0.011 83.271 1.398 0.152 0.009
OHER SI-SS 944.546 65.987 0.027 0.01 88.584 0.739 0.093 0.013
OHER SI-SL 634.095 50.014 0.053 0.014 87.804 0.986 0.092 0.016
OHER SI-LS 611.977 43.081 0.041 0.014 87.712 1.252 0.086 0.013
OHER SI-LL 426.655 44.262 0.095 0.019 82.11 2.352 0.116 0.016
Significant Truee Truef Trueg Trueh
PHER MHCSA 894.093 40.821 0.102 0.03 92.315 0.493 0.159 0.009
PHER SI-SS 718.195 51.235 0.105 0.023 91.069 0.647 0.124 0.01
PHER SI-SL 500.736 34.383 0.112 0.016 88.74 1.034 0.131 0.014
PHER SI-LS 522.83 37.182 0.108 0.017 88.611 1.09 0.125 0.01
PHER SI-LL 402.229 39.746 0.118 0.015 82.375 2.624 0.131 0.012
Significant True Falsei Truej Truek
RHER MHCSA 2341.24 72.031 0.122 0.023 74.238 1.264 0.123 0.009
RHER SI-SS 792.363 65.438 0.094 0.014 84.998 1.899 0.094 0.012
RHER SI-SL 530.103 49.515 0.107 0.018 83.07 2.296 0.105 0.014
RHER SI-LS 556.441 57.438 0.106 0.016 83.359 2.092 0.105 0.012
RHER SI-LL 420.333 43.22 0.121 0.017 79.294 3.281 0.126 0.013
Significant Truel Truem Truen Trueo
SHER MHCSA 2144.72 172.247 0.028 0.012 76.929 2.717 0.094 0.019
SHER SI-SS 1393.50 158.49 0.027 0.01 83.148 1.765 0.072 0.019
SHER SI-SL 852.855 107.04 0.034 0.013 83.743 1.163 0.068 0.016
SHER SI-LS 911.251 119.976 0.033 0.012 84.085 1.771 0.07 0.018
SHER SI-LL 564.659 52.838 0.055 0.011 81.828 2.587 0.087 0.014
Significant Truep Trueq Truer Trues
a
False for SI-SL and SI-LS
b
False for MHCSA and SI-SS, SI-SL and SI-LS
c
False for SI-SS and SI-SL, SI-SS and SI-LS, SI-SL and SI-LS
d
e
f
False for MHCSA and SI-SS
g
h
i
True for MHCSA and SI-LL, SI-SS and SI-SL, SI-SS and SI-LL, SI-LS and SI-LL
j
k
False for SI-SS and SI-LS, SI-SL and SI-LL, SI-LS and SI-LL
l
m
False for MHCSA and SI-LL, SI-SL and SI-LS
n
o
False for MHCSA and SI-LL, SI-SL and SI-LS
p
q
False for MHCSA and SI-SS, MHCSA and SI-LS, SI-SS and SI-LS, SI-SL and SI-LS
r
False for SI-SS and SI-SL, SI-SS and SI-LL, SI-SL and SI-LS
s
False for MHCSA and SI-LL, SI-SS and SI-SL, SI-SS and SI-LS, SI-SL and SI-LS
Table A.7: Summary of results from an empirical study into the SI-HCSA compared to
MP-HCSA.
339
A.2.3 Maternal Immunity Empirical Study
the Maternal Immunity Host Clonal Selection Algorithm (MI-HCSA). The non-parametric
group.

AHER MG-HCSA 1270.97 78.43 0.106 0.021 89.685 0.78 0.162 0.009
AHER MIHCSA-S 1216.84 38.157 0.106 0.02 90.254 0.351 0.159 0.009
AHER MIHCSA-M 1233.45 57.015 0.104 0.019 90.036 0.531 0.162 0.009
AHER MIHCSA-L 1334.18 54.323 0.058 0.016 89.042 0.533 0.157 0.007
Significant Truea Trueb Truec Falsed
OHER MG-HCSA 1260.92 92.225 0.121 0.02 89.618 0.877 0.159 0.009
OHER MIHCSA-S 1253.80 59.516 0.118 0.019 89.694 0.606 0.162 0.009
OHER MIHCSA-M 1330.90 60.777 0.107 0.017 88.938 0.644 0.16 0.009
OHER MIHCSA-L 1392.42 70.266 0.087 0.018 88.271 0.726 0.157 0.009
Significant Truee True Truef Falseg
PHER MG-HCSA 882.648 43.358 0.125 0.016 92.239 0.492 0.161 0.008
PHER MIHCSA-S 890.847 26.554 0.122 0.022 92.24 0.303 0.16 0.008
PHER MIHCSA-M 890.237 42.345 0.122 0.019 92.299 0.451 0.161 0.008
PHER MIHCSA-L 979.421 33.188 0.112 0.019 91.631 0.363 0.161 0.009
Significant Trueh Truei Truej False
RHER MG-HCSA 1590.41 66.83 0.157 0.02 85.945 0.811 0.153 0.01
RHER MIHCSA-S 1592.31 72.526 0.153 0.019 85.877 0.882 0.151 0.01
RHER MIHCSA-M 1618.38 69.943 0.151 0.016 85.521 0.876 0.146 0.009
RHER MIHCSA-L 1717.02 70.612 0.145 0.019 84.245 0.911 0.144 0.01
Significant Truek Falsel Truem True
SHER MG-HCSA 2321.94 100.304 0.047 0.01 74.558 1.826 0.123 0.011
SHER MIHCSA-S 2314.24 132.046 0.045 0.009 74.572 2.239 0.121 0.01
SHER MIHCSA-M 2328.53 122.946 0.043 0.007 74.419 2.161 0.12 0.011
SHER MIHCSA-L 2275.90 102.101 0.044 0.01 75.279 1.804 0.115 0.012
Significant False Falsen False Falseo
a
False for MIHCSA-S and MIHCSA-M
b
False for MG-HCSA and MIHCSA-S, MG-HCSA and MIHCSA-M, MIHCSA-S and MIHCSA-M
c
False for MIHCSA-S and MIHCSA-M
d
True for MG-HCSA and MIHCSA-L
e
False for MG-HCSA and MIHCSA-S
f
False for MG-HCSA and MIHCSA-S
g
True for MIHCSA-S and MIHCSA-L
h
i
j
k
l
True for MG-HCSA and MIHCSA-L
m
n
True for MG-HCSA and MIHCSA-M
o
True for MG-HCSA and MIHCSA-L, MIHCSA-S and MIHCSA-L
Table A.8: Summary of results from an empirical study into the MI-HCSA.
340
A.2.4 Evolved Immunity Empirical Study
the Evolved Immunity Host Clonal Selection Algorithm (EI-HCSA). The non-parametric
group.

AHER MGHCSA 1272.81 62.233 0.111 0.016 89.697 0.601 0.163 0.01
AHER EI-HCSA 1048.32 71.445 0.112 0.019 90.22 0.561 0.165 0.014
Significant True False True False
OHER MGHCSA 1319.73 75.51 0.123 0.017 89.064 0.77 0.161 0.01
OHER EI-HCSA 1142.77 72.429 0.112 0.016 88.975 0.626 0.163 0.009
Significant True True False False
PHER MGHCSA 902.305 35.86 0.118 0.016 92.149 0.399 0.161 0.009
PHER EI-HCSA 588.025 55.818 0.126 0.021 91.98 0.678 0.164 0.012
Significant True False False False
RHER MGHCSA 1605.65 65.629 0.146 0.019 85.69 0.789 0.149 0.01
RHER EI-HCSA 1454.08 67.685 0.158 0.018 85.77 0.853 0.153 0.012
Significant True True False False
SHER MGHCSA 2328.98 117.362 0.048 0.01 74.437 2.041 0.123 0.011
SHER EI-HCSA 2272.29 105.348 0.049 0.009 74.247 1.812 0.125 0.012
Table A.9: Summary of results from an empirical study into the EI-HCSA.
341
Appendix B
Works Arising from the Study
This appendix lists all works derived from this research project in the interest of full
disclosure.
B.1 Peer-Reviewed Publications

This section lists peer-reviewed publications derived from this research project.
1. J. Brownlee IIDLE: An Immunological Inspired Distributed Learning Environment for Mul-

tiple Objective and Hybrid Optimisation. Proceedings of the IEEE Congress in Evolutionary
Computation (CEC’06); Sheraton Vancouver Wall Centre Hotel, Vancouver, BC, Canada.
USA: IEEE Press; 2006: 507-513.
B.2 Software
This section lists software created as a result of this research project. All experimentation
in this dissertation were achieved using the OAT platform with customised extensions.
1. IIDLE: Immunologically Inspired Distributed Learning Environment: IIDLE is based

in a holistic abstraction of the acquired immune system as a spatially distributed, cir-
culating and heterogeneous population of specialised discrete units that provide a ho-
mogeneous defence against external pathogenic material. It is these discrete, atomic,
immutable and disposable units or information packets that represent the knowledge
captured by the system, and are the substrate for the maintenance and learning
processes. Includes Java implementation, with applets, user interface, and optimisa-
tion experiments. See [60] and download the software from http://www.ict.swin.
edu.au/personal/jbrownlee/iidle/index.html or http://jason.brownlee05.
googlepages.com.
2. WEKA Classification Algorithms: A collection of plug-in algorithms for the WEKA

machine learning workbench including Artificial Neural Network (ANN) algorithms
and Artificial Immune System (AIS) algorithms. Algorithms include Learning Vector
Quantisation (LVQ), Perceptron, Back-Propagation, Self-Organising Map (SOM),
Artificial Immune Recognition System (AIRS), and the Clonal Selection Algorithm
(CLONALG). Download the software from: http://wekaclassalgos.sourceforge.
net.
3. Optimisation Algorithm Toolkit (OAT): A workbench and toolkit for developing,

evaluating, and playing with classical and state-of-the-art optimisation algorithms
342
on standard benchmark problem domains; including reference algorithm implemen-
tations (Java), graphing, visualisations and much more. Algorithms include Evo-
lutionary, Swarm, Immune System and others. Domains include Continuous Func-
tion, Travelling Salesman, Protein Folding and others. Download the software from:
http://optalgtoolkit.sourceforge.net.
B.3 Unpublished Technical Reports

This section lists unpublished technical reports derived from this research project, many of
which were reviewed internally (Complex Intelligent Systems Laboratory). These technical
reports are available on the Internet at the following URL: http://jason.brownlee05.
googlepages.com.
B.3.1 Year 2005

1. J. Brownlee Artificial Immune Recognition System (AIRS) - A Review and Analysis [Tech-
nical Report]. Victoria, Australia: Centre for Intelligent Systems and Complex Processes
(CISCP), Faculty of Information and Communication Technologies (ICT), Swinburne Uni-
versity of Technology; 2005 Jan; Technical Report ID: 1-01.
2. J. Brownlee Clonal Selection Theory and CLONALG - The Clonal Selection Classification
Algorithm (CSCA) [Technical Report]. Victoria, Australia: Centre for Intelligent Systems
and Complex Processes (CISCP), Faculty of Information and Communication Technologies
(ICT), Swinburne University of Technology; 2005 Jan; Technical Report ID: 2-01.
3. J. Brownlee Immunos-81 - The Misunderstood Artificial Immune System [Technical Report].

Victoria, Australia: Centre for Intelligent Systems and Complex Processes (CISCP), Faculty
of Information and Communication Technologies (ICT), Swinburne University of Technology;
2005 Feb; Technical Report ID: 3-01.
4. J. Brownlee Immunological Inspired Novelty Detection [Technical Report]. Victoria, Aus-

tralia: Centre for Intelligent Systems and Complex Processes (CISCP), Faculty of Infor-
mation and Communication Technologies (ICT), Swinburne University of Technology; 2005
Apr; Technical Report ID: 4-01.
5. J. Brownlee On Biologically Inspired Computation [Technical Report]. Victoria, Australia:

Centre for Intelligent Systems and Complex Processes (CISCP), Faculty of Information
and Communication Technologies (ICT), Swinburne University of Technology; 2005 Jul;
Technical Report ID: 5-01.
6. J. Brownlee Protein Folding - A Benchmark Combinatorial Optimisation Problem [Tech-

nical Report]. Victoria, Australia: Centre for Intelligent Systems and Complex Processes
(CISCP), Faculty of Information and Communication Technologies (ICT), Swinburne Uni-
versity of Technology; 2005 Jul; Technical Report ID: 6-01.
7. J. Brownlee The Pole Balancing Problem - A Review of a Benchmark Control Theory Prob-
lem [Technical Report]. Victoria, Australia: Centre for Intelligent Systems and Complex
Processes (CISCP), Faculty of Information and Communication Technologies (ICT), Swin-
burne University of Technology; 2005 Jul; Technical Report ID: 7-01.
8. J. Brownlee Introduction to IIDLE - The Immunological Inspired Distributed Learning En-

vironment [Technical Report]. Victoria, Australia: Centre for Intelligent Systems and Com-
plex Processes (CISCP), Faculty of Information and Communication Technologies (ICT),
Swinburne University of Technology; 2005 Sep; Technical Report ID: 8-01.
9. J. Brownlee Preliminary Experiments with IIDLE [Technical Report]. Victoria, Australia:

and Communication Technologies (ICT), Swinburne University of Technology; 2005 Sep;
343
10. J. Brownlee Implementation Specification for IIDLE [Technical Report]. Victoria, Australia:
and Communication Technologies (ICT), Swinburne University of Technology; 2005 Sep;
11. J. Brownlee Further Preliminary Experiments with IIDLE [Technical Report]. Victoria,
Australia: Centre for Intelligent Systems and Complex Processes (CISCP), Faculty of Infor-
Oct; Technical Report ID: 11-01.
B.3.2 Year 2006

1. D. Angus, and J. Brownlee Discrete History Ant Systems [Unpublished Manuscript]. Vic-
toria, Australia: Centre for Intelligent Systems and Complex Processes (CISCP), Faculty of
Information and Communication Technologies (ICT), Swinburne University of Technology;
2006.
2. J. Brownlee A Coarse Taxonomy of Artificial Immune Systems [Technical Report]. Victo-

ria, Australia: Complex Intelligent Systems Laboratory (CIS), Centre for Information Tech-
nology Research (CITR), Faculty of Information and Communication Technologies (ICT),
Swinburne University of Technology; 2006 May; Technical Report ID: TR2006-01.
B.3.3 Year 2007

1. J. Brownlee A Note on Research Methodology and Benchmarking Optimization Algorithms
[Technical Report]. Victoria, Australia: Complex Intelligent Systems Laboratory (CIS),
Centre for Information Technology Research (CITR), Faculty of Information and Communi-
cation Technologies (ICT), Swinburne University of Technology; 2007 Jan; Technical Report
ID: 070125.
2. J. Brownlee Clonal Selection Algorithms [Technical Report]. Victoria, Australia: Complex

Intelligent Systems Laboratory (CIS), Centre for Information Technology Research (CITR),
Faculty of Information and Communication Technologies (ICT), Swinburne University of
Technology; 2007 Feb; Technical Report ID: 070209A.
3. J. Brownlee A Review of the Clonal Selection Theory of Acquired Immunity [Technical Re-
port]. Victoria, Australia: Complex Intelligent Systems Laboratory (CIS), Centre for Infor-
mation Technology Research (CITR), Faculty of Information and Communication Technolo-
gies (ICT), Swinburne University of Technology; 2007 Feb; Technical Report ID: 070223A.
4. J. Brownlee A Series of Adaptive Models Inspired by the Acquired Immune System [Tech-
nical Report]. Victoria, Australia: Complex Intelligent Systems Laboratory (CIS), Centre
for Information Technology Research (CITR), Faculty of Information and Communication
Technologies (ICT), Swinburne University of Technology; 2007 Feb; Technical Report ID:
070227A.
5. J. Brownlee Complex Adaptive Systems [Technical Report]. Victoria, Australia: Complex

Technology; 2007 Mar; Technical Report ID: 070302A.
6. J. Brownlee Satisficing, Optimization, and Adaptive Systems [Technical Report]. Victoria,

Australia: Complex Intelligent Systems Laboratory (CIS), Centre for Information Tech-
Swinburne University of Technology; 2007 Mar; Technical Report ID: 070305A.
7. J. Brownlee The ‘shape-space’ and ‘affinity landscape’ Immunological Paradigms [Techni-

cal Report]. Victoria, Australia: Complex Intelligent Systems Laboratory (CIS), Centre
Technologies (ICT), Swinburne University of Technology; 2007 Mar; Technical Report ID:
070310A.
344
8. J. Brownlee The Physiology of Lymphocyte Migration [Technical Report]. Victoria, Aus-
tralia: Complex Intelligent Systems Laboratory (CIS), Centre for Information Technology
Research (CITR), Faculty of Information and Communication Technologies (ICT), Swin-
burne University of Technology; 2007 Mar; Technical Report ID: 070316A.
9. J. Brownlee A Series of Discrete Repertoire Models Inspired by Lymphocyte Migration [Tech-

Technologies (ICT), Swinburne University of Technology; 2007 Mar; Technical Report ID:
070320A.
10. J. Brownlee An Adaptive Systems Formalism [Technical Report]. Victoria, Australia: Com-
plex Intelligent Systems Laboratory (CIS), Centre for Information Technology Research
(CITR), Faculty of Information and Communication Technologies (ICT), Swinburne Uni-
versity of Technology; 2007 Mar; Technical Report ID: 070320A
11. J. Brownlee ‘Colour Space’: A Contrived Problem Domain for Investigating Adaptive Mod-
els [Technical Report]. Victoria, Australia: Complex Intelligent Systems Laboratory (CIS),
cation Technologies (ICT), Swinburne University of Technology; 2007 Mar; Technical Report
ID: 070322A.
12. J. Brownlee ‘Small Models’: A Methodology for Designing and Investigating Adaptive Sys-
tems [Technical Report]. Victoria, Australia: Complex Intelligent Systems Laboratory
(CIS), Centre for Information Technology Research (CITR), Faculty of Information and
Communication Technologies (ICT), Swinburne University of Technology; 2007 Mar; Tech-
nical Report ID: 070326A.
13. J. Brownlee Autonomous Distributed Control in the Immune System Using Diffuse Feedback
cation Technologies (ICT), Swinburne University of Technology; 2007 Mar; Technical Report
ID: 070329A.
14. J. Brownlee Artificial Immune System Thesis Bibliography [Technical Report]. Victoria,
Swinburne University of Technology; 2007 Mar; Technical Report ID: 070330A.
15. J. Brownlee Experimental Scenarios in Colour Space [Technical Report]. Victoria, Australia:
Complex Intelligent Systems Laboratory (CIS), Centre for Information Technology Research
versity of Technology; 2007 Apr; Technical Report ID: 070402A
16. J. Brownlee Populations of Interacting Immune Systems: Evolution and Immunization [Tech-
Technologies (ICT), Swinburne University of Technology; 2007 Apr; Technical Report ID:
070410A.
17. J. Brownlee A Series of Multiple Immune System Adaptive Models Inspired by Evolution and
Immunization [Technical Report]. Victoria, Australia: Complex Intelligent Systems Labo-
ratory (CIS), Centre for Information Technology Research (CITR), Faculty of Information
and Communication Technologies (ICT), Swinburne University of Technology; 2007 Apr;
Technical Report ID: 070413A
18. J. Brownlee Realizing Elementary Clonal Selection Algorithms [Technical Report]. Victoria,
Swinburne University of Technology; 2007 Apr; Technical Report: 070418A
345
19. J. Brownlee The ‘Pathogenic Exposure’ Paradigm [Technical Report]. Victoria, Australia:
versity of Technology; 2007 Apr; Technical Report ID: 070422A
20. J. Brownlee Antigen-Antibody Interaction [Technical Report]. Victoria, Australia: Complex

Technology; 2007 Apr; Technical Report ID: 070427A
21. J. Brownlee Realizing Elementary Discrete Repertoire Clonal Selection Algorithms [Tech-
Technologies (ICT), Swinburne University of Technology; 2007 Apr; Technical Report ID:
070430A
22. J. Brownlee Cognition, Immunology, and the Cognitive Paradigm [Technical Report]. Victo-
ria, Australia: Complex Intelligent Systems Laboratory (CIS), Centre for Information Tech-
Swinburne University of Technology; 2007 May; Technical Report: 070504A.
23. J. Brownlee Lazy and Competitive Learning [Technical Report]. Victoria, Australia: Com-
versity of Technology; 2007 May; Technical Report: 070508A.
24. J. Brownlee Learning Classifier Systems [Technical Report]. Victoria, Australia: Complex
Technology; 2007 May; Technical Report: 070514A.
25. J. Brownlee Towards Unification... [Technical Report]. Victoria, Australia: Complex In-
telligent Systems Laboratory (CIS), Centre for Information Technology Research (CITR),
Technology; 2007 May; Technical Report: 070516A
26. J. Brownlee A Clonal Selection Algorithm and Extensions [Technical Report]. Victoria, Aus-
burne University of Technology; 2007 May; Technical Report: 070521A.
27. J. Brownlee A T-Cell Mediated Clonal Selection Algorithm [Technical Report]. Victoria,
Swinburne University of Technology; 2007 May; Technical Report: 070523A
28. J. Brownlee An Intra-Repertoire Recognition Algorithm [Technical Report]. Victoria, Aus-

burne University of Technology; 2007 May; Technical Report: 070525A
29. J. Brownlee A Spatial Clonal Selection Algorithm [Technical Report]. Victoria, Australia:
versity of Technology; 2007 May; Technical Report: 070531A
30. J. Brownlee Optimizing and Matching Bitstrings [Technical Report]. Victoria, Australia:
versity of Technology; 2007 Jun; Technical Report: 070604A.
346
31. J. Brownlee A Lymphocyte Homing Algorithm [Technical Report]. Victoria, Australia: Com-
32. J. Brownlee A Lymphoid Tissue Model and Lymphocyte Recruitment Algorithms [Techni-
cal Report]. Victoria, Australia: Complex Intelligent Systems Laboratory (CIS), Centre for
Information Technology Research (CITR), Faculty of Information and Communication Tech-
nologies (ICT), Swinburne University of Technology; 2007 Jun; Technical Report: 070609A.
33. J. Brownlee A Hierarchical Framework of the Acquired Immune System [Technical Report].
Victoria, Australia: Complex Intelligent Systems Laboratory (CIS), Centre for Informa-
tion Technology Research (CITR), Faculty of Information and Communication Technologies
(ICT), Swinburne University of Technology; 2007 Jun; Technical Report: 070613A.
34. J. Brownlee Darwinism and Selectionist Theories [Technical Report]. Victoria, Australia:
35. J. Brownlee Information Processing with a Lymphoid Tissue Architecture [Technical Report].
Victoria, Australia: Complex Intelligent Systems Laboratory (CIS), Centre for Informa-
tion Technology Research (CITR), Faculty of Information and Communication Technologies
(ICT), Swinburne University of Technology; 2007 Jun; Technical Report: 070619A.
36. J. Brownlee A Population-Based Clonal Selection Algorithm and Extensions [Technical Re-
gies (ICT), Swinburne University of Technology; 2007 Jun; Technical Report: 070621A.
37. J. Brownlee A Generational Population-Based Clonal Selection Algorithm and Extensions

cation Technologies (ICT), Swinburne University of Technology; 2007 Jun; Technical Report:
070623A.
38. J. Brownlee Models, Algorithms, and the Hierarchal Acquired Immune System Framework
070625A.
39. J. Brownlee Integration of the Pathogenic Exposure Paradigm and the Hierarchical Immune
System Framework [Technical Report]. Victoria, Australia: Complex Intelligent Systems
Laboratory (CIS), Centre for Information Technology Research (CITR), Faculty of Infor-
Jun; Technical Report: 070626A.
40. J. Brownlee Towards General Information Processing Models of the Acquired Immune Sys-
tem [Technical Report]. Victoria, Australia: Complex Intelligent Systems Laboratory (CIS),
070627A
41. J. Brownlee A Review of the Immunological Inspired Distributed Learning Environment

cation Technologies (ICT), Swinburne University of Technology; 2007 Jul; Technical Report:
070702A.
347
42. J. Brownlee Convergence and Dominance in the Minimal Clonal Selection Algorithm [Techni-
cal Report]. Victoria, Australia: Complex Intelligent Systems Laboratory (CIS), Centre for
Information Technology Research (CITR), Faculty of Information and Communication Tech-
nologies (ICT), Swinburne University of Technology; 2007 Jul; Technical Report: 070706A.
43. J. Brownlee Immunology as a Metaphor for Adaptive and Distributed Information Process-
ing [Technical Report]. Victoria, Australia: Complex Intelligent Systems Laboratory (CIS),
cation Technologies (ICT), Swinburne University of Technology; 2007 Jul; Technical Report:
070714A.
44. J. Brownlee Debrief of PhD Project Practices [Technical Report]. Victoria, Australia: Com-
versity of Technology; 2007 Dec; Technical Report: 20071204-A.
45. J. Brownlee Unexplored Territory: Seeds for Future Research Investigations [Technical Re-
gies (ICT), Swinburne University of Technology; 2007 Dec; Technical Report: 20071210-A.
46. J. Brownlee OAT HowTo: High-level Domain, Problem, and Algorithm Implementation
cation Technologies (ICT), Swinburne University of Technology; 2007 Dec; Technical Report:
20071218A.
47. J. Brownlee OAT: The Optimization Algorithm Toolkit [Technical Report]. Victoria, Aus-
burne University of Technology; 2007 Dec; Technical Report: 20071220A.
48. J. Brownlee OAT: A Quick Start Guide [Technical Report]. Victoria, Australia: Complex
Technology; 2007 Dec; Guide: 20071220A.
348
Bibliography
[1] A. Aamodt and E. Plaza. Case-based reasoning: Foundational issues, methodological

variations, and system approaches. Artificial Intelligence Communications, 7(1):39–
59–, 1994.
[2] S Aaronson. Np-complete problems and physical reality. ACM SIGACT News
(COLUMN: Complexity theory), 36(1):30–52–, 2005.
[3] P. Adamidis, P. Adamidis, and V. Petridis. Co-operating populations with different

evolution behaviours. In V. Petridis, editor, Proc. IEEE International Conference
on Evolutionary Computation, pages 188–191, 1996.
[4] F. Agens, M. M. Rosado, and A. A. Freitas. Peripheral b cell survival. Cell Mol Life
Sci, 57(8-9):1220–1228, Aug 2000.
[5] D. W. Aha. Case-based learning algorithms. In Proceedings of the 1991 DARPA

Case-Based Reasoning Workshop, pages 147–158–. Morgan Kaufmann, 1991.
[6] D. W. Aha. Artificial intelligence review. Editorial: Lazy Learning, 11(1-5):7–10–,

February 1997.
[7] D. W. Aha. Feature weighting for lazy learning algorithms. In H. Motoda, editor,
Feature Extraction, Construction and Selection: A Data Mining Perspective, pages –
. Kluwer, Norwell MA, USA, 1998.
[8] D. W. Aha, D. Kibler, and M. K. Albert. Instance-based learning algorithms. Ma-

chine Learning, 6(1):37–66–, January 1991.
[9] S. C. Ahalt, A. K. Krishnamurthy, P. Chen, and D. E. Melton. Competitive learning

algorithms for vector quantization. Neural Networks, 3(3):277–290–, 1990.
[10] E. Ahmed, A. S. Elgazzar, and A. S. Hegazi. An overview of complex adaptive

systems. ARXIV, eprint arXiv:nlin/0506059, pages –, July 2005.
[11] E. Ahmed and A. H. Hashish. On modelling the immune system as a complex

system. Theory in biosciences, 124(3-4):413–418–, 2006.
[12] E. Alba. Parallel evolutionary algorithms can achieve super-linear performance.

Information Processing Letters, 82:7–13, 2002.
[13] E. Alba. Parallel Metaheuristics: A New Class of Algorithms. John Wiley, 2005.
[14] E. Alba and M. Tomassini. Parallelism and evolutionary algorithms. IEEE Trans-
actions on Evolutionary Computation, 6(5):443–462, 2002.
[15] E. Alba and J. M. Troya. Influence of the migration policy in parallel distributed gas
with structured and panmictic populations. Applied Intelligence, 12:163–181, 2000.
349
[16] M. M. Ali, C. Storey, and A. Trn. Application of stochastic global optimization
algorithms to practical problems. Journal of Optimization Theory and Applications,
95(3):545–563, 1997.
[17] K. P. Anchor, P. D. Williams, G. H. Gunsch, and G. B. Lamont. The computer

defense immune system: current and future research in intrusion detection. In Pro-
ceedings of the 2002 Congress on Evolutionary Computation (CEC ’02), volume 2,
pages 1027–1032–. IEEE Press, USA, 2002.
[18] K. P. Anchor, J. B. Zydallis, G. H. Gunsch, and G. B. Lamont. Extending thecom-

puter defense immune system: Network intrusion detection with a multiobjective
evolutionary programming spproach. In P. J. Bentley, editor, First International
Conference on Artificial Immune Systems (ICARIS 2002), pages 1221–, UK, 2002.
[19] A. O. Anderson. Structure and organization of the lymphatic system. In Ethan M.

Shevach, editor, Immunophysiology : the role of cells and cytokines in immunity and
inflammation, pages 14–45–. Oxford University Press, New York, 1990.
[20] A. O. Anderson and S. Shaw. Chapter 3: Lymphocyte trafficking. In T.A. Fleisher,

B.D. Schwartz, W.T. Shearer, and W. Strober, editors, Clinical Immunology, Prin-
ciples and Practice, pages 39–49–. Mosby, London, 1995.
[21] P. W. Anderson, K. J. Arrow, and D. Pines. Proceedings of The Santa Fe Institute

Studies in the Sciences of Complexity - Economy As an Evolving Complex System.
Addison Wesley Publishing Company, USA, 1988.
[22] R. M. Anderson. Modern vaccines: immunisation and herd immunity. The Lancet,
335(8690):641–646–, March 1990.
[23] R. M Anderson and R. M. May. Vaccination and herd immunity to infectious dis-
eases. Nature, 318:323–329–, November 1985.
[24] P. S. Andrews and J. Timmis. Inspiration for the next generation of artificial immune
systems. In Marcin L. Pilat, Peter J. Bentley, and Jonathan Timmis, editors, Lecture
Notes in Computer Science, pages 126–138–. Springer-Verlag, 2005.
[25] C. Andrieu, N. de Freitas, A. Doucet, and M. I. Jordan. An introduction to mcmc

for machine learning. Machine Learning, 50:5–43, 2003.
[26] J. B. Andr, J. B Ferdy, and B. Godelle. Within-host parasite dynamics, emerging

trade-off, and evolution of virulence with immune system. Evolution, 57(7):1489–
1497, 2003.
[27] A.M. Anile, V. Cutello, G. Narzisi, G. Nicosia, and S. Spinella. Lipschitzian pat-
tern search and immunological algorithm with quasi-newton method for the protein
folding problem: An innovative multistage approach. In Lecture Notes in Computer
Science, pages 307–323–. Springer, 2006.
[28] W. Brian Arthur. Introduction: Process and emergence in the economy. In Steven
Durlauf and David A. Lane, editors, The Economy as an Evolving Complex System
II, volume Volume XXVII, pages 1–04–. Addison-Wesley Pub. Co, Reading, Mass,
USA, 1997.
[29] C. G. Atkeson, A. W. Moore, and S. Schaal. Locally weighted learning. Artificial

Intelligence Review, 11(1-5):11–73–, February 1997.
350
[30] T. Bäck. Optimal mutation rates in genetic search. In Proceedings of the Fifth
International Conference on Genetic Algorithms, pages 2–9–. Morgan Kaufman, San
Mateo, CA, USA, 1993.
[31] T. Back, D. B Fogel, and Z. Michalwicz. Evolutionary Computation 1 - Basic Algo-

rithms and Operators. Institute of Physics (IoP) Publishing, Bristol, UK, 2000.
[32] T. Back, D. B Fogel, and Z. Michalwicz. Evolutionary Computation 2 - Advanced

Algorithms and Operators. Institute of Physics (IoP) Publishing, Bristol, UK, 2000.
[33] R. Barr, B. Golden, J. Kelly, M. Rescende, and W. Stewart. Designing and reporting
on computational experiments with heuristic methods. Journal of Heuristics, 1:9–
32–, 1995.
[34] S. Behnke. Hierarchical Neural Networks for Image Interpretation. Springer, 2003.
[35] T. C. Belding. The distributed genetic algorithm revisited. In Proceedings of the 6th
International Conference on Genetic Algorithms, 1995.
[36] P. J. Bentley, J. Greensmith, and S. Ujjin. Two ways to grow tissue for artificial
immune systems. In Marcin L. Pilat, Peter J. Bentley, and Jonathan Timmis, edi-
tors, Lecture Notes in Computer Science, pages 139–. Springer, Berlin / Heidelberg,
Germany, 2005.
[37] C. Berek, A. Berger, and M. Apel. Maturation of the immune response in germinal
centers. Cell, 67(6):1121–1129–, 1991.
[38] C. Berek and M. Ziegner. The maturation of the immune response. Immunology
Today, 14(8):400–404–, 1993.
[39] D. Bergemann and J. Valimaki. Bandit problems. Cowles Foundation Discussion

Papers 1551, Cowles Foundation, Yale University, January 2006.
[40] H. Bersini and F. Varela. Hints for adaptive problem solving gleaned from immune
networks. In Lecture Notes In Computer Science, pages 343–354–. Springer-Verlag,
London, UK, 1990.
[41] K. Beyer, J. Goldstein, R. Ramakrishnan, and U. Shaft. When is ”nearest neighbor”

meaningful? In Lecture Notes in Computer Science, pages 217–, Springer Berlin /
Heidelberg, 1998.
[42] J. Bienenstock and A. D. Befus. Mucosal immunology. Immunology, 41(2):249–270–,

1980.
[43] M. Birattari, P. Balaprakash, and M.Dorigo. Aco/f-race: Ant colony optimization

and racing techniques for combinatorial optimization under uncertainty. In M. Gen-
dreau, P. Greistorfer, W. J. Gutjahr, R. F. Hartl, and M. Reimann, editors, MIC
2005: The Sixth Metaheuristics International Conference, pages 107–112–, 2005.
[44] M. Birattari and M. Dorigo. How to assess and report the performance of a stochastic
algorithm on a benchmark problem: Mean or best result on a number of runs?
Technical report, IRIDIA, Universit Libre de Bruxelles, Brussels, Belgium, 2005.
[45] M. Birattari, T. Sttzle, L. Paquete, and K. Varrentrapp. A racing algorithm for

configuring metaheuristics. In Proceedings of the Genetic and Evolutionary Compu-
tation Conference, pages 11–18–. Morgan Kaufmann Publishers Inc., San Francisco,
CA, USA, 2002.
351
[46] C. M. Bishop. Neural Networks for Pattern Recognition. Oxford University Press,
1995.
[47] A. Di Blas, A. Jagota, and R. Hughey. Optimizing neural networks on simd parallel
computers. Parallel Computing, 31:97–115, 2005.
[48] C. Blum and A. Roli. Metaheuristics in combinatorial optimization: Overview and

conceptual comparison. ACM Computing Surveys (CSUR), 35(3):268–308–, Septem-
ber 2003.
[49] L. Boggess and J. S. Hamaker. The effect of irrelevant features on airs, an artificial
immune-based classifier. In Intelligent Engineering Systems through Artificial Neural
Networks (ANNIE), pages 219–224–, 2003.
[50] S. Bonhoeffer and M. A. Nowak. Intra-host versus inter-host selection: viral strate-
gies of immune function impairment. Proc Natl Acad Sci U S A, 91(17):8062–8066,
Aug 1994.
[51] L. Bottou and V. Vapnik. Local learning algorithms. Neural Computation, 4(6):888–
900–, November 1992.
[52] J. Y. L. Boudec and S. Sarafijanovic. An artificial immune system approach to

misbehavior detection in mobile ad hoc networks. In Masayuki Murata and Naoki
Wakamiya, editors, Lecture Notes in Computer Science, pages 396–. Springer, Berlin
/ Heidelberg, Germany, 2004.
[53] S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge University Press,

2004.
[54] P. Brandtzaeg. Mucosal immunity: integration between mother and the breast-fed
infant. Vaccine, 21(24):3382–3388–, 2003.
[55] L. Breiman. Bagging predictors. Machine Learning, 24(2):123–140, 1996.
[56] J. Breivik. Commentary: cancer–evolution within. Int J Epidemiol, 35(5):1161–

1162, Oct 2006.
[57] P. Bretscher and M Cohn. A theory of self-nonself discrimination. Science,

169(950):1042–109–, 1970.
[58] R. Brooks. A robust layered control system for a mobile robot. IEEE Journal Of
Robotics And Automation, 2(1):14–23, 1986.
[59] D. S. Broomhead and David Lowe. Radial basis functions, multi-variable functional
interpolation and adaptive networks. Technical report, Royal Signals and Radar
Establishment Malvern, England, 1988.
[60] J. Brownlee. IIDLE: An immunological inspired distributed learning environment

for multiple objective and hybrid optimisation. In Proceedings of the IEEE Congress
in Evolutionary Computation (CEC’06), pages 507–513. IEEE Press, 2006.
[61] M. D. Buhmann. Radial basis functions. Acta Numerica, 9(00):1–38–, 2000.
[62] P. Bull, A. Knowles, G. Tedesco, and A. Hone. Diophantine benchmarks for the
b-cell algorithm. In Proceedings 5th International Conference on Artificial Immune
Systems (ICARIS 2006), pages –, 2006.
352
[63] F. M. Burnet. A modification of jernes theory of antibody production using the
concept of clonal selection. Australian Journal of Science, 20,:67–69–, 1957.
[64] F. M. Burnet. The clonal selection theory of acquired immunity. Vanderbilt Univer-
sity Press, Nashville, Tennessee, U.S.A., 1959.
[65] F. M. Burnet. Clonal selection and after. Theoretical Immunology, pages 63–85–,
1978.
[66] E. C. Butcher and L. J. Picker. Lymphocyte homing and homeostasis. Science,

272(5258):60–67–, April 1996.
[67] E. C. Butcher, M. Williams, K. Youngman, L. Rott, and M. Briskin. Lymphocyte

trafficking and regional immunity. Advances in Immunology, 72:209–253–, 1999.
[68] E. Cant-Paz. Efficient and Accurate Parallel Genetic Algorithms. Kluwer Academic
Publishers (Springer), 2000.
[69] E. Cant-Paz and D. E. Goldberg. Are multiple runs of genetic algorithms better
than one? In Genetic and Evolutionary Computation, GECCO 2003, volume 2723
of Lecture Notes in Computer Science. Springer Berlin / Heidelberg, 2003.
[70] J. Carneiro and J. Stewart. Rethinking ”shape space”: evidence from simulated
docking suggests that steric shape complementarity is not limiting for antibody-
antigen recognition and idiotypic interactions. Journal of theoretical biology,
169(4):391–402–, 1994.
[71] G. A. Carpenter and S. Grossberg. The art of adaptive pattern recognition by a

self-organizing neural network. Computer, 21(3):77–88–, March 1988.
[72] D. J. Cavicchio. Adaptive search using simulated evolution. PhD thesis, Computer
and Communication Sciences, University of Michigan, USA, 1970.
[73] D. J. Cavicchio. Reproductive adaptive plans. In Proceedings of the ACM annual

conference - Volume 1, volume 1, pages 60–70–. ACM Press, New York, NY, USA,
1972.
[74] K Chan, A Saltelli, and S Tarantola. Sensitivity analysis of model output: variance-
based methods make the difference. In Proceedings of the 29th conference on Winter
simulation (Winter Simulation Conference), pages 261–268–. ACM Press, New York,
NY, USA, 1997.
[75] M. Chiarandini, D. Basso, and T. Sttzle. Statistical methods for the comparison
of stochastic optimizers. In Michel Gendreau, Peter Greistorfer, Walter J. Gutjahr,
Richard F. Hartl, and Marc Reimann, editors, MIC2005: Proceedings of the 6th
Metaheuristics International Conference, pages 189–196–, 2005.
[76] E. Clark, A. Hone, and J. Timmis. A markov chain model of the b-cell algorithm. In
Lecture Notes in Computer Science, pages 318–330–. Springer, Berlin / Heidelberg,
2005.
[77] C. A. Coello Coello and N. C. Cortes. An approach to solve multiobjective optimiza-

tion problems based on an artificial immune system. In P. J. Bentley, editor, First
International Conference on Artificial Immune Systems (ICARIS), pages 212–221–,
2002.
353
[78] C. A. Coello Coello and N. C. Cortes. Hybridizing a genetic algorithm with an arti-
ficial immune system for global optimization. Engineering Optimization, 36(5):607–
634–, October 2004.
[79] C. A. Coello Coello and N. C. Corts. A parallel implementation of an artificial

immune system to handle constraints in genetic algorithms: Preliminary results.
In Congress on Evolutionary Computation (CEC’2002), volume 1, pages 819–824–,
USA, 2002. IEEE Computer Society.
[80] I. R. Cohen. Antigenic mimicry, clonal selection and autoimmunity. Journal of

Autoimmunity, 16(3):337–340–, 2001.
[81] I. R. Cohen. Tending Adam’s Garden: Evolving the Cognitive Immune Self. Aca-
demic Press, New York, 2001.
[82] I. R. Cohen, U. Hershberg, and S. Solomon. Antigen-receptor degeneracy and im-

munological paradigms. Molecular Immunology, 40(14-15):993–996–, 2004.
[83] N. C. Corts and C. A. Coello Coello. Multiobjective optimization using ideas from
the clonal selection principle. In Lecture Notes in Computer Science, pages 158–170–,
Berlin / Heidelberg, 2003. Springer.
[84] T. Cover and P. Hart. Nearest neighbor pattern classification. IEEE Transactions
on Information Theory, 13(1):21–27–, January 1967.
[85] T. G. Crainic and N. Hail. Parallel metaheuristics applications. In Parallel Meta-

heuristics. John Wiley & Sons, Inc., 2005.
[86] E. Crist and A. I. Tauber. Selfhood, immunity, and the biological imagination:
The thought of frank macfarlane burnet. Biology and Philosophy, 15(4):509–533–,
September 2000.
[87] J. Cunliffe. Tissue homeostasis and immunity–more on models. Scand J Immunol,

64(3):172–176, Sep 2006.
[88] V. Cutello, D. Lee, S. Leone, G. Nicosia, and M. Pavone. Clonal selection algo-
rithm with dynamic population size for bimodal search spaces. In Lecture Notes in
Computer Science, pages 949–958–. Springer, 2006.
[89] V. Cutello, D. Lee, G. Nicosia, M. Pavone, and I. Prizzi. Aligning multiple protein
sequences by hybrid clonal selection algorithm with insert-remove-gaps and block-
shuffling operators. In Lecture Notes in Computer Science, pages 321–334–. Springer,
Berlin / Heidelberg, 2006.
[90] V. Cutello, G. Morelli, G. Nicosia, and M. Pavone. Immune algorithms with ag-
ing operators for the string folding problem and the protein folding problem. In
Jens Gottlieb, editor, Lecture Notes in Computer Science, pages 80–90–, Berlin /
Heidelberg, Germany, 2005. Springer.
[91] V. Cutello, G. Narzisi, and G. Nicosia. A multi-objective evolutionary approach to

the protein structure prediction problem. Journal of the Royal Society Interface,
3(6):139–151–, 2006.
[92] V. Cutello, G. Narzisi, G. Nicosia, and M. Pavone. Clonal selection algorithms:

A comparative case study using effective mutation potentials. In Marcin L. Pilat,
Peter J. Bentley, and Jonathan Timmis, editors, LNCS, pages 13–28–, Germany,
2005. Springer.
354
[93] V. Cutello, G. Narzisi, G. Nicosia, and M. Pavone. An immunological algorithm for
global numerical optimization. In LNCS, pages 284–295–, 2005.
[94] V. Cutello, G. Narzisi, G. Nicosia, M. Pavone, and G. Sorace. How to escape

traps using clonal selection algorithms. In The First International Conference on
Informatics in Control, Automation and Robotics, ICINCO 2004, pages 322–326–.
INSTICC Press, 2004.
[95] V. Cutello and G. Nicosia. An immunological approach to combinatorial optimiza-

tion problems. In J.C. Riquelme and M. Toro, editors, Lecture Notes In Computer
Science, pages 361–370–, London, UK, 2002. Springer-Verlag.
[96] V. Cutello and G. Nicosia. Multiple learning using immune algorithms. In Proceed-
ings of 4th International Conference on Recent Advances in Soft Computing, RASC
2002, pages 102–107–, 2002.
[97] V. Cutello and G. Nicosia. Chapter vi. the clonal selection principle for in silico
and in vivo computing. In Fernando J. Von Zuben, editor, Recent Developments
in Biologically Inspired Computing, pages 104–146–, Hershey, London, Melbourne,
Singapore, 2005. Idea Group Publishing.
[98] V. Cutello and G. Nicosia. A clonal selection algorithm for coloring, hitting set
and satisfiability problems. In Lecture Notes in Computer Science, pages 324–337–.
Springer, 2006.
[99] V. Cutello, G. Nicosia, and E. Pavia. A parallel immune algorithm for global op-
timization. In Series on Advances in Soft Computing, pages 467–475–. Springer-
Verlag, 2006.
[100] V. Cutello, G. Nicosia, and M. Pavone. A hybrid immune algorithm with information
gain for the graph coloring problem. In Lecture Notes in Computer Science, pages
171–182–, Berlin / Heidelberg, 2003. Springer.
[101] V Cutello, G. Nicosia, and M Pavone. Exploring the capability of immune algo-
rithms: A characterization of hypermutation operators. In Vincenzo Cutello, Pe-
ter J. Bentley, and Jon Timmis, editors, Lecture Notes in Computer Science, pages
[102] V. Cutello, G. Nicosia, and M. Pavone. An immune algorithm with hyper-

macromutations for the dill’s 2d hydrophobic-hydrophilic model. In Congress on
Evolutionary Computation, (CEC2004), volume 1, pages 1074–1080–, USA, 2004.
IEEE Computer Society.
[103] V. Cutello, G. Nicosia, M. Pavone, and G. Narzisi. Real coded clonal selection al-
gorithm for unconstrained global optimization using a hybrid inversely proportional
hypermutation operator. In ACM, volume 2, pages 950–954–. ACM Press, New
York, NY, USA, 2006.
[104] V. Cutello, G. Nicosia, M. Pavone, and J. Timmis. An immune algorithm for pro-
tein structure prediction on lattice models. IEEE Transactions on Evolutionary
Computation, pages –, 2006.
[105] N. Danilova. The evolution of immune mechanisms. Journal of experimental zoology.

Part B. Molecular and developmental evolution, 306(6):496–520–, 2006.
355
[106] D. Dasgupta. Chapter ten: Information processing in the immune system. In
M. Dorigo and F. Glover, editors, New Ideas in Optimization, pages 203–216.
McGraw-Hill, England, 1999.
[107] D. Dasgupta and H. Brian. Mobile security agents for network traffic analy-
sis. DARPA Information Survivability Conference and Exposition (DISCEX II’01),
2:1332–, 2001.
[108] L. N. de Castro and J. Timmis. An artificial immune network for multimodal func-
tion optimization. In Proceedings of the 2002 Congress on Evolutionary Computation
(CEC ’02), volume 1, pages 699–704–. IEEE Computer Society, USA, 2002.
[109] L. N. de Castro and J. Timmis. Artificial Immune Systems: A new computational

intelligence approach. Springer-Verlag, Great Britain, 2002.
[110] L. N. de Castro and J. Timmis. Artificial immune systems: A novel paradigm to

pattern recognition. In L. Alonso and C. Fyfe, editors, Artificial Neural Networks
in Pattern Recognition, pages 67–84–. University of Paisley, UK, 2002.
[111] L. N. de Castro and J. Timmis. Artificial immune systems as a novel soft com-
puting paradigm. Soft Computing - A Fusion of Foundations, Methodologies and
Applications, 7(8):526–544–, August 2003.
[112] L. N. de Castro and F. J. Von Zuben. Artificial immune systems - part i: Basic
theory and applications. Technical report, Department of Computer Engineering
and Industrial Automation, School of Electrical and Computer Engineering, State
University of Campinas, Brazil, December 1999.
[113] L. N. de Castro and F. J. Von Zuben. Artificial immune systems: Part ii: A survey of
applications. Technical report, Department of Computer Engineering and Industrial
Automation, School of Electrical and Computer Engineering, State University of
Campinas, Brazil, February 2000.
[114] L. N. de Castro and F. J. Von Zuben. The clonal selection algorithm with engineering
applications. In David E. Goldberg, Erick Cant-Paz, Lee Spector, Ian C. Parmee,
and Hans-Georg Beyer, editors, Proceedings of the Genetic and Evolutionary Com-
putation Conference (GECCO ’00), Workshop on Artificial Immune Systems and
Their Applications, pages 36–37–. Morgan Kaufmann, 2000.
[115] L. N. de Castro and F. J. Von Zuben. An evolutionary immune network for data
clustering. In Proceedings Sixth Brazilian Symposium on Neural Networks, pages
84–89–. IEEE Computer Society, USA, 2000.
[116] L. N. de Castro and F. J. Von Zuben. Chapter xii: ainet: An artificial immune
network for data analysis. In R. A. Sarker and C. S. Newton, editors, Data Mining:
A Heuristic Approach, pages 231–259–. Idea Group Publishing, USA, 2001.
[117] L. N. de Castro and F. J. Von Zuben. Learning and optimization using the clonal
selection principle. IEEE Transactions on Evolutionary Computation, 6(3):239–251,
June 2002.
[118] L. N. de Castro and F. J. Von Zuben. Chapter 1: From biologically inspired comput-
ing to natural computing. In Fernando J. Von Zuben, editor, Recent developments
in biologically inspired computing, pages –. Idea Group, Hershey, PA, 2005.
356
[119] E. D. de Jong, D. Thierens, and R. A. Watson. Hierarchical genetic algorithms.
In Parallel Problem Solving from Nature - PPSN VIII, Lecture Notes in Computer
Science, pages 232–241. Springer, 2004.
[120] K. A. De Jong. An Analysis of the Behavior of a Class of Genetic Adaptive Systems.

PhD thesis, Computer and Communication Sciences, University of Michigan, USA,
1975.
[121] K. A. De Jong. Genetic algorithms are not function optimizers. In Proceedings of

the Second Workshop on Foundations of Genetic Algorithms (FOGA), pages 5–17–.
Morgan Kaufmann, 1992.
[122] S. Fazekas de St.Groth and R. G. Webster. Disquisitions on original antigenic sin.

i. evidence in man. The Journal of Experimental Medicine, 124:331–345–, 1966.
[123] K. Deb. Chapter 14 - speciation methods. In David B Fogel and Zbigniew Michal-
wicz, editors, Evolutionary Computation 2 - Advanced Algorithms and Operations,
pages 93–100–. Institute of Physics (IOP) Publishing Ltd., UK, 2000.
[124] K. Deb. Multi-Objective Optimization Using Evolutionary Algorithms. John Wiley

and Sons, 2001.
[125] K. Deb and S. Agrawal. Understanding interactions among genetic algorithm param-
eters. In Colin R. Reeves, editor, Proceedings of the Fifth Workshop on Foundations
of Genetic Algorithms (FOGA), pages 265–286–. Morgan Kaufmann, 1999.
[126] K. Deb and David E. Goldberg. An investigation of niche and species formation in
genetic function optimization. In Proceedings of the Third International Conference
on Genetic Algorithms, pages 42–50–. Morgan Kaufmann, 1989.
[127] A. Demers, D. Greene, C. Hauser, W. Irish, J. Larson, S. Shenker, H. Sturgis,

D. Swinehart, and D. Terry. Epidemic algorithms for replicated database mainte-
nance. In Proceedings of the sixth annual ACM Symposium on Principles of dis-
tributed computing, 1987.
[128] D. DeSieno. Adding a conscience to competitive learning. In IEEE International

Conference on Neural Networks, volume 1, pages 117–124–. IEEE Press, USA, 1988.
[129] P. D’haeseleer. An immunological approach to change detection: theoretical results.

In Proceedings of the 9th IEEE Computer Security Foundations Workshop, pages
18–26–. IEEE Computer Society, Washington, DC, USA, 1996.
[130] P. D’haeseleer, S. Forrest, and P. Helman. An immunological approach to change

detection: algorithms, analysis and implications. In IEEE Symposium on Security
and Privacy, pages 110–119–. IEEE Computer Society, Washington, DC, USA, 1996.
[131] B. Du, B. Du, J. Gu, W. Wang, and D.H.K. Tsang. Multispace search for minimizing
the maximum nodal degree. In Jun Gu, editor, Proc. Sixth International Conference
on Computer Communications and Networks, pages 364–367, 1997.
[132] A. E. Eiben and M. Jelasity. A critical note on experimental research methodology

in ec. In Proceedings of the 2002 Congress on Evolutionary Computation (CEC ’02),
volume 1, pages 582–587–. IEEE Press, USA, 2002.
[133] A. E. Eiben and C. A. Schippers. On evolutionary exploration and exploitation.

Fundamenta Informaticae, 35(1-4):35–50–, August 1998.
357
[134] M. El-Abd and M. Kamel. A taxonomy of cooperative search algorithms. In Hybrid
Metaheuristics, 2005.
[135] A. P. Engelbrecht. Computational Intelligence: An Introduction. J. Wiley & Sons,

2002.
[136] J. Englund, W. P. Glezen, and P. A. Piedra. Maternal immunization against viral

disease. Vaccine, 16(14-15):1456–1463–, 1998.
[137] J. D. Farmer, S. A. Kauffman, N. H. Packard, and A. S. Perelson. Adaptive dynamic

networks as models for the immune system and autocatalytic sets. Annals of the
New York Academy of Sciences, 504(1):118131–, 1986.
[138] J. D. Farmer, N. H. Packard, and Alan S. Perelson. The immune system, adaptation,
and machine learning. Physica D, 22:187–204–, 1986.
[139] D. Ferber. The education of t cells. Science: IMMUNOLOGY, 316(5822):191–193–,

April 2007.
[140] P. E. M. Fine. Herd immunity: History, theory, practice. Oxford Journals, Epidemi-
ologic Reviews, 15(2):265–302–, 1993.
[141] T. C. Fogarty. Varying the probability of mutation in the genetic algorithm. In

Proceedings of the third international conference on Genetic algorithms, pages 104–
109–. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 1989.
[142] D. B. Fogel. Evolutionary Computation : Toward a New Philosophy of Machine

Intelligence. John Wiley & Sons, Inc., New Jersey, USA, 1995.
[143] D.B. Fogel, L.J. Fogel, and J.W. Atmar. Meta-evolutionary programming. In Con-
ference Record of the Twenty-Fifth Asilomar Conference on Signals, Systems and
Computers, pages 540–545, 4–6 Nov. 1991.
[144] L. J. Fogel, A. J. Owens, and M. J. Walsh. Artificial Intelligence Through Simulated

Evolution. Wiley Publishing, 1966.
[145] N. Forbes. Biologically inspired computing. Computing in Science and Engineering,

2(6):83–87–, 2000.
[146] N. Forbes. Imitation of life: how biology is inspiring computing. Mit Press, Cam-
bridge, Massachusetts, 2004.
[147] S. Forrest and S. A. Hofmeyr. Immunology as information processing. In De-

sign Principles for the Immune System and Other Distributed Autonomous Systems,
pages 361–388–. Oxford University Press, New York, 2001.
[148] S. Forrest, S. A. Hofmeyr, and A. Somayaji. Computer immunology. Communica-

tions of the ACM, 40(10):88–96–, October 1997.
[149] S. Forrest, S. A. Hofmeyr, A. Somayaji, and T. A. Longstaff. A sense of self for unix
processes. In Proceedings of the 1996 IEEE Symposium on Security and Privacy,
pages 120–128–. IEEE Computer Society, 1996.
[150] S. Forrest, B. Javornik, R. E. Smith, and A. S. Perelson. Using genetic algorithms

to explore pattern recognition in the immune system. Evolutionary Computation,
1(3):191–211–, 1993.
358
[151] S. Forrest, A. S. Perelson, L. Allen, and R. Cherukuri. Self-nonself discrimination in
a computer. In Proceedings of the 1992 IEEE Symposium on Security and Privacy,
pages 202–212–. IEEE Computer Society Press, Washington, DC, USA, 1994.
[152] T. Francis Jr. On the doctrine of original antigenic sin. Proceedings of the American
Philosophical Society, 104(6):572–578–, December 1960.
[153] O. Francois and C. Lavergne. Design of evolutionary algorithms - a statistical per-

spective. IEEE Transactions on Evolutionary Computation, 5(2):129–148–, April
2001.
[154] A. A. Freitas and B. Rocha. Population biology of lymphocytes: the flight for
survival. Annu Rev Immunol, 18:83–111, 2000.
[155] A. A. Freitas and J. Timmis. Revisiting the foundations of artificial immune systems:
A problem-oriented perspective. In Peter Bentley and Emma Hart, editors, Lecture
Notes in Computer Science, pages 229–241–. Springer, Berlin, Heidelberg, 2003.
[156] Y. X. Fu and D. D. Chaplin. Development and maturation of secondary lymphoid

tissues. Annual review of immunology, 17:399–433–, 1999.
[157] T. Fukuda, K. Mori, and M. Tsukiyama. Immune networks using genetic algorithm
for adaptive production scheduling. 15th IFAC World Congress, 3:57–60–, 1993.
[158] A. S. Fukunaga. Restart scheduling for genetic algorithms. In Parallel Problem

Solving from Nature PPSN V, 1998.
[159] K. Fukunaga. Introduction to Statistical Pattern Recognition. Academic Press, 1990.
[160] N. Ganguly, G. Canright, and A. Deutsch. Design of an efficient search algorithm

for p2p networks using concepts from natural immune systems. In Proceedings of
the 8th International Conference on Parallel Problem Solving from Nature (PPSN
VIII), pages –, 2004.
[161] V. V. Ganusov, C. T. Bergstrom, and R. Antia. Within-host population dynamics

and the evolution of microparasites in a heterogeneous host population. Evolution,
56(2):213–223, 2002.
[162] M. Garey, D. Johnson, and H. Witsenhausen. The complexity of the generalized lloyd
- max problem (corresp.). IEEE Transactions on Information Theory, 28(2):255–
256, Mar 1982.
[163] S. M. Garrett. Parameter-free, adaptive clonal selection. In Congress on Evolu-

tionary Computing (CEC 2004), volume 1, pages 1052–1058–. IEEE Press, USA,
2004.
[164] S. M. Garrett. How do we evaluate artificial immune systems? Evolutionary Com-

putation, 13(2):145–177–, 2005.
[165] E. Gaudin, M. Rosado, F. Agenes, A. Mclean, and A. A. Freitas. B-cell homeosta-

sis, competition, resources, and positive selection by self-antigens. Immunological
Reviews, 197(1):102–, February 2004.
[166] M. Gell-Mann. Complex adaptive systems. In David Pines and David Meltzer,
editors, Complexity: metaphors, models, and reality, pages 17–45–. Addison-Wesley,
USA, 1994.
359
[167] A. J. T. George and D. Gray. Receptor editing during affinity maturation. Immunol-
ogy Today, 20(4):196–, April 1999.
[168] A. J. T. George and D. Gray. Jumping or walking: which is better? Immunology

Today, 21(1):55–, 2000.
[169] P. J. Gerrish and R. E. Lenski. The fate of competing beneficial mutations in an

asexual population. Genetica, 102-103(1-6):127–144, 1998.
[170] A. Gersho, A. Gersho, and Y. Shoham. Hierarchical vector quantization of speech

with dynamic codebook allocation. In Y. Shoham, editor, Proc. IEEE International
Conference on ICASSP ’84. Acoustics, Speech, and Signal Processing, volume 9,
pages 416–419, 1984.
[171] J. Ghosh. Multiclassifier systems: Back to the future. In Proceedings of the Third
International Workshop on Multiple Classifier Systems, 2002.
[172] W. P. Glezen and M. Alpers. Maternal immunization. Clinical infectious diseases,

International Conference on Acute Respiratory Infections, 28(2):188–239–, 1999.
[173] D. E. Goldberg. Genetic Algorithms in Search, Optimization and Machine Learning.

Addison Wesley Publishing Company, Inc., USA, Canada, 1989.
[174] D. E. Goldberg. From genetic and evolutionary optimization to the design of con-
ceptual machines. Evolutionary Optimization, 1(1):1–12–, 1999.
[175] D. E. Goldberg. The design of innovating machines: A fundamental discipline for a

postmodern systems engineering. In Engineering Systems Symposium, pages –. MIT
Engineering Systems Division, USA, 2004.
[176] D. E. Goldberg and J. R. Genetic algorithms with sharing for multimodal function
optimization. In Proceedings of the Second International Conference on Genetic
Algorithms on Genetic algorithms and their application, pages 41–49–. Lawrence
Erlbaum Associates, Inc., Mahwah, NJ, USA, 1987.
[177] D. Edward Goldberg. The Design of Innovation: Lessons from and for Competent
Genetic Algorithms. Kluwer Academic Publishers, USA, 2002.
[178] D. Goodman, L. Boggess, and A. Watkins. Artificial immune system classification

of multiple-class problems. In A. Buczak, J. Ghosh, M. Embrechts, O. Ersoy, and
S Kercel, editors, Fuzzy Logic, Evolutionary Programming, Complex Systems and
Artificial Life, volume 12, pages 179–184–, New York, 2002. ASME Press.
[179] D. Goodman, L. Boggess, and A. Watkins. An investigation into the source of

power for airs, an artificial immune classification system. In Proceedings of the
International Joint Conference on Neural Networks (IJCNN’03), pages 1678–1683–,
Portland, Oregon, USA, 2003.
[180] D. E. Goodman and L. C. Boggess. The role of hypothesis filtering in (airs), an artifi-
cial immune classifier. In Anna L. Buczak, Joydeep Ghosh, Mark J. Embrechts, and
Okan Ersoy, editors, Proceedings of the Artificial Neural Networks in Engineering
Conference, 2003., volume 13, pages 243–248–, New York, 2004. ASME Press.
[181] C. C. Goodnow. Chance encounters and organized rendezvous. Immunological Re-

views, 156:5–10–, 1997.
360
[182] H. J. Gould, P. Takhar, H. E. Harries, S. R. Durham, and C. J. Corrigan. Germinal-
centre reactions in allergic inflammation. Trends in Immunology, 27(10):446–452–,
October 2006.
[183] J. L. Gowans. The recirculation of lymphocytes from blood to lymph in the rat.
American journal of physiologic imaging, 146(1):54–69–, 1959.
[184] J. L. Gowans and E. J. Knight. The route of re-circulation of lymphocytes in
the rat. Proceedings of the Royal Society of London. Series B, Biological Sciences,
159(975):257–282–, 1964.
[185] J. Greensmith. New Frontiers For An Artificial Immune System. PhD thesis, Uni-
versity of Leeds, UK, 2003.
[186] J. Greensmith and S. Cayzer. An artificial immune system approach to semantic
document classification. In Peter Bentley and Emma Hart, editors, Lecture Notes
in Computer Science, pages 136–146–. Springer, Berlin / Heidelberg, 2003.
[187] A. Grilo, A. Caetano, and A. Rosa. Immune system simulation through a complex
adaptive system model. In M. Koppen, S. Ovaska, T. Furuhashi, and F. Homann,
editors, Springer Engineering Series, Soft Computing in Industrial Applications -
Recent Advances, pages 681–704–. Springer, 2002.
[188] S. Grossberg. Neural expectation: cerebellar and retinal analogs of cells fired by
learnable or unlearned pattern classes. Biological Cybernetics, 10(1):49–57–, January
1972.
[189] S. Grossberg. Adaptive pattern classification and universal recoding: I. parallel de-
velopment and coding of neural feature detectors. Biological Cybernetics, 23(3):121–
134–, September 1976.
[190] S. Grossberg. Adaptive pattern classification and universal recoding: Ii. feedback,
expectation, olfaction, illusions. Biological Cybernetics, 23(4):187–202–, December
1976.
[191] S. Grossberg. Competitive learning: from interactive activation to adaptive reso-
nance. Cognitive Science, 11:23–63–, 1988.
[192] J. Gu. Multispace search: A new optimization approach. In Algorithms and Com-
putation, 1994.
[193] J. Gu. Multispace search for satisfiability and np-hard problems. In Satisfiability
Problem: Theory and Applications : DIMACS Workshop, 1997.
[194] I Guyon and A Elisseeff. An introduction to variable and feature selection. Journal
of Machine Learning Research, 3:1157–1182, 2003.
[195] D. L. Hall and S. A. H. McMullen. Mathematical Techniques in Multisensor Data
Fusion. Artech House, 2004.
[196] L. A. Hanson, M. Korotkova, and E. Telemo. Breast-feeding, infant formulas, and
the immune system. Annals of allergy, asthma & immunology : official publication
of the American College of Allergy, Asthma, & Immunology, 90(6 Suppl 3):59–63–,
2003.
[197] P. K. Harmer, P. D. Williams, G. H. Gunsch, and G. B. Lamont. An artificial
immune system architecture for computer security applications. IEEE Transactions
on Evolutionary Computation, 6(3):252–280–, June 2002.
361
[198] W Harriman, H Volk, N Defranoux, and M Wabl. Immunoglobulin class switch
recombination. Annual Review of Immunology, 11:361–384–, April 1993.
[199] E. Hart and P. Ross. Clustering moving data with a modified immune algorithm.
In J. Gottlieb, P.L. Lanzi, R.E. Smith, S. Cagnoni, E. Hart, G.R. Raidl, and H. Ti-
jink, editors, Lecture Notes in Computer Science, pages 394–. Springer, Berlin /
Heidelberg, 2001.
[200] E. Hart and P. Ross. Exploiting the analogy between the immune system and sparse
distributed memories. Genetic Programming and Evolvable Machines, 4(4):333–358–
, December 2003.
[201] E. Hart and J. Timmis. Application areas of ais: The past, the present and the
future. In Marcin L. Pilat, Peter J. Bentley, and Jonathan Timmis, editors, Lecture
Notes in Computer Science, pages 483–. Springer, Berlin / Heidelberg, Germany,
2005.
[202] R. Hecht-Nielsen. Counterpropagation networks. In Proceedings of the IEEE Inter-

national Conference on Neural Networks, pages –. IEEE Press, USA, 1987.
[203] R. Hecht-Nielsen. Counterpropagation networks. Applied Optics, 26(23):4979–4984–,

1987.
[204] S. M. Hedrick. The acquired immune system: a vantage from beneath. Immunity,
21(5):607–615–, 2004.
[205] U. Hershberg, S. Solomon, and Irun R. Cohen. What is the basis of the immune
systems specificity? In The MIRIAM Project Series, pages 377–384–. ESCULAPIO
Pub. Co., Bologna, Italy, 2003.
[206] J. Hesser and R. Mnner. Towards an optimal mutation probability for genetic al-
gorithms. In Lecture Notes In Computer Science, pages 23–32–. Springer-Verlag,
London, UK, 1991.
[207] R. Hightower, S. Forrest, and A. S. Perelson. The baldwin effect in the immune
system: Learning by somatic hypermutation. In Adaptive Individuals in Evolving
Populations: Models and Algorithms, pages 159–167–. Addison Wesley, Reading,
MA. USA, 1995.
[208] R. R. Hightower, S. Forrest, and A. S. Perelson. The evolution of emergent orga-

nization in immune system gene libraries. In Proceedings of the 6th International
Conference on Genetic Algorithms, pages 344–350–, San Francisco, CA, USA, 1995.
Morgan Kaufmann Publishers Inc.
[209] T. K. Ho. The random subspace method for constructing decision forests. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 20:832–844, 1998.
[210] G. W. Hoffmann. A neural network model based on the analogy with the immune
system. Journal of Theorectical Immunology, 122(1):33–67–, September 1986.
[211] S. Hofmeyr and S. Forrest. Immunity by design: An artificial immune system. In

Jason Daida, Agoston E. Eiben, Max H. Garzon, Vasant Honavar, Mark Jakiela, and
Robert E. Smith, editors, Proceedings of the Genetic and Evolutionary Computation
Conference (GECCO), volume 2, pages 1289–1296–. Morgan-Kaufmann, Orlando,
Florida, USA, 1999.
362
[212] S. A. Hofmeyr. An Immunological Model of Distributed Detection and its Application
to Computer Security. PhD thesis, Department of Computer Sciences, University of
New Mexico, USA, 1999.
[213] S. A. Hofmeyr. An interpretative introduction to the immune system. In I. R. Cohen

L. A. Segel, editor, Design Principles for Immune Systems and Othe Distributed Au-
tonomous Systes, volume Santa Fe Institute, Studies in the Sciences of Complexity,
pages 3–26. Oxford University Press, 2001.
[214] S. A. Hofmeyr and S. Forrest. Architecture for an artificial immune system. Evolu-
tionary Computation, 8(4):443–473–, December 2000.
[215] S. A. Hofmeyr, S. Forrest, and A. Somayaji. Intrusion detection using sequences of

system calls. Journal of Computer Security, 6(3):151–180–, 1998.
[216] J. H. Holland. Adaptation in Natural and Artificial Systems: An Introductory Anal-

ysis with Applications to Biology, Control, and Artificial Intelligence, volume 2nd
Edition. MIT Press, USA, 1992.
[217] J. H. Holland. Complex adaptive systems. Daedalus, 121(1):17–30–, 1992.
[218] J. H. Holland. Hidden Order: How Adaptation Builds Complexity. Addison Wesley
Publishing Company, USA, 1995.
[219] A. Hone and J. Kelsey. Optima, extrema, and artificial immune systems. In Vincenzo
Cutello, Peter J. Bentley, and Jon Timmis, editors, Lecture Notes in Computer
Science, pages 80–90–. Springer, Berlin / Heidelberg, 2004.
[220] T. Honkela. Self-Organizing Maps in Natural Language Processing. PhD thesis,

Neural Networks Research Centre, Helsinki University of Technology, Finland, 1997.
[221] J. N. Hooker. Testing heuristics: We have it all wrong. Journal of Heuristics,

1(1):33–42–, September 1995.
[222] R. Horst, P. M. Pardalos, and N. V. Thoai. Introduction to Global Optimization.

Kluwer Academic Publishers, 2nd edition, 2000.
[223] N. Hozumi and S. Tonegawa. Evidence for somatic rearrangement of immunoglob-

ulin genes coding for variable and constant regions. National Academy of Sciences,
73(10):3628–3632–, 1976.
[224] T. Hrycej. Modular Learning in Neural Networks A Modularized Approach to Neural.

Wiley, 1992.
[225] X. Hu, R. Shonkwiler, and M. Spruill. Random restarts in global optimization.

Technical Report Tech Report 110592-015, School of Mathematics, Georgia Institute
of Technology, January 1994.
[226] N. Intrator and S. Edelman. Competitive learning in biological and artificial neural
computation. Trends in Cognitive Sciences, 1(7):268–272–, October 1997.
[227] Y. Ishida. Fully distributed diagnosis by pdp learning algorithm: towards immune
network pdp model. In IJCNN International Joint Conference on Neural Networks,
volume 1, pages 777–782–. IEEE Computer Society, USA, 1990.
[228] Y. Ishida. Distributed and autonomous sensing based on the immune network.
Artificial Life and Robotics: Computer Science and Engineering, 2(1):1–7–, March
1998.
363
[229] Y. Ishida. Immunity-Based Systems - A Design Perspective. Springer-Verlag, Berlin,
Germany, 2004.
[230] A. Ishiguro, T. Kondo, Y. Watanabe, and Y. Uchikawa. Dynamic behavior arbitra-
tion of autonomous mobile robots using immune networks. In IEEE International
Conference on Evolutionary Computation, volume 2, pages 722–727–. IEEE Press,
USA, 1995.
[231] A. Ishiguro, Y. Watanabe, T. Kondo, Y. Shirai, and Y. Uchikawa. A robot with
a decentralized consensus-making mechanism based onthe immune system. In Pro-
ceedings. ISADS 97., Third International Symposium on Autonomous Decentralized
Systems, pages 231–237–. IEEE Press, USA, 1997.
[232] N. K. Jerne. The natural-selection theory of antibody formation. Proceedings of
the National Academy of Sciences of the United States of America, 41(11):849–857–,
1955.
[233] N. K. Jerne. Clonal selection in a lymphocyte network. In Cellular Selection and
Regulation in the Immune Response, volume 29 of Society of General Physiologists
Series, pages 39–48. Raven Press, 1974.
[234] N. K. Jerne. Towards a network theory of the immune system. Annales
d’immunologie (Annals of Immunology), Institut Pasteur (Paris, France), Societe
Francaise d’Immunologie, 125(C):373–389–, January 1974.
[235] N. K. Jerne. Idiotypic networks and other preconceived ideas. Immunological Re-
views, 79:5–24–, 1984.
[236] D. S Johnson. A theoreticians guide for experimental analysis of algorithms. In
D. S. Johnson and C. C. McGeoch, editors, Proceedings of the 5th and 6th DIMACS
Implementation Challenges, pages 215–250–. American Mathematical Society, 2002.
[237] J. Jost. External and internal complexity of complex adaptive systems. Theory in
biosciences, 123(1):69–88–, 2004.
[238] J. H. Jun, D. W. Lee, and K. B. Sim. Realization of cooperative strategies and swarm
behavior in distributed autonomous robotic systems using artificial immune system.
In Systems, Man, and Cybernetics, 1999. IEEE SMC ’99 Conference Proceedings.
1999 IEEE International Conference on, volume 6, pages 614–619vol.6, 12-15 Oct.
1999.
[239] H. Kargupta and P. Chan. Advances in Distributed and Parallel Knowledge Discov-
ery. AAAI Press/MIT Press, 2000.
[240] S. Kauffman. The Origins of Order: Self-Organization and Selection in Evolution.
Oxford University Press, New York, USA, 1993.
[241] S. Kauffman and S. Levin. Towards a general theory of adaptive walks on rugged
landscapes. Journal of Theoretical Biology, 128(1):11–45–, 1987.
[242] S. A. Kauffman and E. D. Weinberger. The nk model of rugged fitness landscapes
and its application to maturation of the immune response. Journal of theoretical
biology, 141(2):211–245–, 1989.
[243] S. A. Kauffman, E. D. Weinberger, and A. S. Perelson. Maturation of the immune
response via adaptive walks on affinity landscapes. In Theoretical Immunology I:
Santa Fe Institute Studies in the Sciences of Complexity, pages –. Addison Wesley,
USA, 1988.
364
[244] M. Kearns. Thoughts on hypothesis boosting. Unpublished manuscript, 1988.
[245] M. A. Keller and E. R. Stiehm. Passive immunity in prevention and treatment of

infectious diseases. Clinical Microbiology Reviews, 13(4):602–614–, 2000.
[246] J. Kelsey and J. Timmis. Immune inspired somatic contiguous hypermutation for
function optimisation. In Lecture Notes in Computer Science, pages 207–218–.
Springer, Berlin / Heidelberg, 2003.
[247] J. Kelsey, J. Timmis, and A. Hone. Chasing chaos. In The 2003 Congress on
Evolutionary Computation, (CEC ’03), volume 1, pages 413–419–. IEEE Computer
Society, USA, 2003.
[248] J. O. Kephart. A biologically inspired immune system for computers. In Pattie Maes,
editor, Artificial Life IV, pages 130–139–. MIT Press, Cambridge, Massachusetts,
USA, 1994.
[249] J. O. Kephart, G. B. Sorkin, W. C. Arnold, D. M. Chess, G. J. Tesauro, and S. R.

White. Biologically inspired defences against computer viruses. In Proceedings of
the 14th International Joint Conference on Artificial Intelligence (IJCA ’95), pages
985–996–. Morgan Kaufmann Publishers, Inc., 1995.
[250] J. O. Kephart, G. B. Sorkin, and M. Swimmer. An immune system for cyberspace. In

IEEE International Conference on Systems, Man, and Cybernetics, ’Computational
Cybernetics and Simulation’, volume 1, pages 879–884–. IEEE Press, USA, 1997.
[251] J. Kim and P. Bentley. Immune memory and gene library evolution in the dynamic
clonal selection algorithm. Genetic Programming and Evolvable Machines, 5(4):361–
391–, December 2004.
[252] J. Kim and P. J. Bentley. Negative selection and niching by an artificial immune
system for network intrusion detection. In A late-breaking paper, Genetic and Evo-
lutionary Computation Conference (GECCO ’99), pages –, 1999.
[253] J. Kim and P. J. Bentley. Towards an artificial immune system for network intrusion
detection: an investigation of clonal selection with a negative selection operator. In
Proceedings of the 2001 Congress on Evolutionary Computation, volume 2, pages
1244–1252–, USA, 2001. IEEE Computer Society.
[254] J. Kim and P. J. Bentley. Towards an artificial immune system for network intrusion
detection: An investigation of dynamic clonal selection. In Congress on Evolutionary
Computation (CEC-2002), pages 1015–1020–, 2002.
[255] J. W. Kim. Integrating Artificial Immune Algorithms for Intrusion Detection. PhD
thesis, Department of Computer Science, University College London, UK, 2002.
[256] J. Klein and N. Nikolaidis. The descent of the antibody-based immune system by
gradual evolution. Proceedings of the National Academy of Sciences of the United
States of America, 102(1):169–174–, January 2005.
[257] V. B. Klimovich. Actual problems of evolutionary immunology. Journal of Evolu-

tionary Biochemistry and Physiology, 38(5):562–574–, September 2002.
[258] T. Knight and J. Timmis. Aine: An immunological approach to data mining. In

T. Lin and Xindon Wu, editors, First IEEE International Conference on Data Min-
ing (ICDM’01), pages 297–304–. IEEE Computer Society, USA, 2001.
365
[259] T. P. Knight. MARIA: A Multilayered Unsupervised Machine Learning Algorithm
Based on the Vertebrate Immune System. PhD thesis, The University of Kent at
Canterbury, UK, 2005.
[260] G. Kohler and C. Milstein. Continuous cultures of fused cells secreting antibody of
predefined specificity. Nature, 256:495–497–, 1975.
[261] T. Kohonen. Self-Organizing Maps, volume Third. Springer-Verlag, Berlin Heidel-
berg, 2001.
[262] M. Kudo and J. Sklansky. Comparison of algorithms that select features for pattern
classifiers. Pattern Recognition, 33:25–41, 2000.
[263] V. Kumar. Algorithms for constraint-satisfaction problems : a survey. The AI
magazine (AI mag.), 13, 1992.
[264] L. Kuncheva and J. Bezdek. An integrated framework for generalized nearest proto-
type classifier design. Int. Journal of Uncertaintly, Fuzziness and Knowledge-Based
Systems, 6:437–457, 1998.
[265] L.I. Kuncheva and J.C. Bezdek. Nearest prototype classification: clustering, genetic
algorithms, or random search? IEEE Transactions on Systems, Man and Cybernet-
ics, Part C, 28(1):160–164, 1998.
[266] H. Y. K. Lau, V. W. K. Wong, and I. S. K. Lee. Immunity-based autonomous guided
vehicles control. Applied Soft Computing, pages –, 2005.
[267] H. Y.K. Lau and V. W.K. Wong. A strategic behavioral-based intelligent transport
system with artificial immune system. IEEE International Conference on Systems,
Man and Cybernetics, 4:3909–3914–, 2004.
[268] H. Y.K. Lau and V. W.K. Wong. An immunity-based distributed multiagent-control
framework. IEEE Transactions on Systems, Man and Cybernetics, Part A, 36(1):91–
108–, January 2006.
[269] J. Lederberg. Ontogeny of the clonal selection theory of antibody formation. re-
flections on darwin and ehrlich. Annals of the New York Academy of Sciences,
546:175–182–, 1988.
[270] D. W. Lee, H. B. Jun, and K. B. Sim. Artificial immune system for realization of
cooperative strategies and group behavior in collective autonomous mobile robots.
In Proceedings of the Forth International Symposium on Artificial Life and Robotics,
AAAI, pages 232–235–, 1999.
[271] V.R. Lesser. An Overview of DAI: Viewing Distributed AI as Distributed Search.
Journal of Japanese Society for Artificial Intelligence-Special Issue on Distributed
Artificial Intelligence, 5(4):392–400, January 1990.
[272] R. M. Lewis, V. T., and M. W. Trosset. Direct search methods: then and now.
Journal of Computational and Applied Mathematics, 124:191–207, 2000.
[273] Z. Li, C. J. Woo, M. D. Iglesias-Usse, D. Ronai, and M. D. Scharff. The generation of
antibody diversity through somatic hypermutation and class switch recombination.
Genes and Development, 18:1–11–, 2004.
[274] P. Liu, P. Liu, and D. W. Wang. Reduction optimization in heterogeneous clus-
ter environments. In Da-Wei Wang, editor, Proc. 14th International Parallel and
Distributed Processing Symposium IPDPS 2000, pages 477–482, 2000.
366
[275] Y. J. Liu and C. Arpin. Germinal center development. Immunological Reviews,
156(1):111126–, 1997.
[276] G. F. Luger and W. A. Stubblefield. Artificial Intelligence: Structures and Strategies

for Complex Problem Solving. Addison Wesley Longman Inc, USA, 1997.
[277] I. C. M. MacLennan. Germinal centers. Annual review of immunology, 12:117–139–,

1994.
[278] M. Magdon-ismail and A. F. Atiya. The early restart algorithm. Neural Computa-
tion, 12:1303–1312, 2000.
[279] S. Mahfoud. Evolutionary Computation 2: Advanced Algorithms and Operations,

chapter Niching Methods, pages 87–92. IoP, 2000.
[280] E. N. Marieb. Essentials of human anatomy & physiology. Pearson/Benjamin Cum-

mings, San Francisco, CA, USA, 2006.
[281] P Marrow. Nature-inspired computing technology and applications. BT Technology

Journal, 18(4):13–23–, October 2000.
[282] T. Martinetz. Competitive hebbian learning rule forms perfectly topology preserving
maps. In Proceedings of the International Conference on Artificial Neural Networks
(ICANN93), pages 427–434–. Springer, Amsterdam, 1993.
[283] T. Martinetz and K. Schulten. A ”neural gas” network learns topologies. In Kai
Mkisara, Olli Simula, and Jari Kangas, editors, Artificial Neural Networks, pages
397–402–. Elsevier, North-Holland, Amsterdam, 1991.
[284] T. Martinetz and K. Schulten. Topology representing networks. Neural Networks,

7(3):507–522–, 1994.
[285] T. M. Martinetz, S. G. Berkovich, and K. J. Schulten. ’neural-gas’ network for

vector quantization and its application totime-series prediction. IEEE Transactions
on Neural Networks, 4(4):558–569–, July 1993.
[286] G. Marwah and L. Boggess. Artificial immune systems for classification: Some
issues. In First International Conference on Artificial Immune Systems, pages 149–
153–, 2002.
[287] P. Matzinger. Tolerance, danger, and the extended family. Annual Review of Im-
munology, 12:991–1045–, 1994.
[288] P. Matzinger. The danger model: A renewed sense of self. Science, 296(5566):301–
305–, 2002.
[289] T. W. McDade and C. M. Worthman. Evolutionary process and the ecology of

human immune function. American Journal of Human Biology, 11(6):705–717–,
1999.
[290] A. R. McLean, M. M. Rosado, F. Agenes, R. Vasconcellos, and A. A. Freitas. Re-

source competition as a mechanism for b cell homeostasis. Proc. Natl. Acad. Sci.
USA, Immunology, 94:5792–5797, May 1997.
[291] H. Meshref and H. VanLandingham. Artificial immune systems: application to

autonomous agents. In IEEE International Conference on Systems, Man, and Cy-
bernetics, volume 1, pages 61–66–. IEEE Press, USA, 2000.
367
[292] M. E. Meyer-Hermann and P. K. Maini. Cutting edge: Back to ”one-way” germinal
centers. The Journal of Immunology, 174:2489–2493–, 2005.
[293] R. S. Michalski, J. G. Carbonell, and T. M. Mitchell. Machine Learning: An Arti-

ficial Intelligence Approach. Morgan Kaufmann, 1983.
[294] R. Michelan and F. J. Von Zuben. Decentralized control system for autonomous
navigation based on an evolved artificial immune network. In Proceedings of the
2002 Congress on Evolutionary Computation, (CEC ’02), volume 2, pages 1021–
1026–. IEEE Press, USA, 2002.
[295] F. Mizessyn and Yoshiteru Ishida. Immune networks for cement plants. In Proceed-
ings International Symposium on Autonomous Decentralized Systems, ISADS 93,
pages 282–288–. IEEE Computer Society, USA, 1993.
[296] P. H. Mohr, N. Ryan, and J. Timmis. Exploiting immunological properties for ubiq-
uitous computing systems. In Vincenzo Cutello, Peter J. Bentley, and Jon Timmis,
editors, Lecture Notes in Computer Science, pages 277–289–. Springer, Berlin / Hei-
delberg, Germany, 2004.
[297] K. Mori, M. Tsukiyama, and T. Fukuda. Immune algorithm with searching diversity
and its application to resource allocation problem. The Transactions of the Institute
of Electrical Engineers of Japan, 113-C(10):872–878–, 1993.
[298] H. Mühlenbein. How genetic algorithms really work: I. mutation and hillclimbing.
In Parallel Problem Solving from Nature 2, PPSN-II, pages 15–26–. Elsevier, 1992.
[299] C. Murgia, J. K Pritchard, S. Y. Kim, A. Fassati, and R. A Weiss. Clonal origin

and evolution of a transmissible cancer. Cell, 126(3):477–487, Aug 2006.
[300] M. Muselli. A theoretical approach to restart in global optimization. Journal of

Global Optimization, 10:1–16, 1997.
[301] M. V. Nagendra Prasad, V. R. Lesser, and S. E. Lander. Retrieval and reasoning in

distributed case bases. Journal of Visual Communication and Image Representation,
Special Issue on Digital Libraries, 7(1):74–87, 1996.
[302] A. Nakada, A. Nakada, T. Shibata, M. Konda, T. Morimoto, and T. Ohmi. A

fully parallel vector-quantization processor for real-time motion-picture compression.
IEEE Journal of Solid-State Circuits, 34(6):822–830, 1999.
[303] R. M. Nesse. Maladaptation and natural selection. The Quarterly Review of Biology,
80:6270–, 2005.
[304] J. Newborough and S. Stepney. A generic framework for population-based al-

gorithms, implemented on multiple fpgas. In Marcin L. Pilat, Peter J. Bentley,
and Jonathan Timmis, editors, Lecture Notes in Computer Science, pages 43–55–.
Springer, Berlin, Heidelberg, 2005.
[305] G. J. V. Nossal. Negative selection of lymphocytes. Cell, 76(2):229–239–, 1994.
[306] G. J. V. Nossal and J. Lederberg. Antibody production by single cells. Nature,

181:1419–1420–, 1958.
[307] M. Nussenzweig. Immune receptor editing: revise and select. Cell, 95(7):875–878–,
1998.
368
[308] W. H. Oddy. Breastfeeding protects against illness and infection in infants and
children: a review of the evidence. Breastfeeding review : professional publication of
the Nursing Mothers’ Association of Australia, 9(2):11–18–, 2001.
[309] W. L. Olszewski. The lymphatic system in body homeostasis: Physiological condi-

tions. Lymphatic Research and Biology, 1(1):11–24–, January 2003.
[310] D. Opitz and R. Maclin. Popular ensemble methods: An empirical study. Journal
of Artificial Intelligence Research, 11:169–198, 1999.
[311] C. H. Papadimitriou and K. Steiglitz. Combinatorial Optimization: Algorithms and

Complexity. Courier Dover Publications, 1998.
[312] K.K. Parhi, K.K. Parhi, F.H. Wu, and K. Genesan. Sequential and parallel neural
network vector quantizers. IEEE Transactions on Computers, 43(1):104–109, 1994.
[313] R. Paton. Introduction to computing with biological metaphors. In Computing With

Biological Metaphors, pages 1–8–. Chapman & Hall, London, UK, 1994.
[314] William E. Paul. Immunology - Recognition and Response : Readings from Scientific
American magazine / edited by William E. Paul. Scientific American Inc., W. H.
Freeman and Company, USA, 1991.
[315] L. Pauling. A theory of the structure and process of formation of antibodies. Journal
of the American Chemical Society, 62:2643–2657–, 1940.
[316] G. Paun. Bio-inspired computing paradigms (natural computing). Unconventional

Programming Paradigms, 3566:155–160–, 2005.
[317] W. Pedrycz. Computational Intelligence: An Introduction. CRC Press, New York,

1997.
[318] E. S. Peer, A. P. Engelbrecht, and F. van den Bergh. Cirg@up optibench: a sta-
tistically sound framework for benchmarking optimisation algorithms. In The 2003
Congress on Evolutionary Computation, (CEC ’03), volume 4, pages 2386–2392–.
IEEE Press, USA, 2003.
[319] M. Pelikan and D. E. Goldberg. Hierarchical problem solving by the bayesian opti-
mization algorithm. IlliGAL Report No. 2000002, Illinois Genetic Algorithms Lab-
oratory, University of Illinois at Urbana-Champaign, Urbana, IL, 2000.
[320] A. S. Perelson. Immune network theory. Immunological Reviews, 110:5–36–, August

1989.
[321] A. S. Perelson and G. F. Ostera. Theoretical studies of clonal selection: Minimal

antibody repertoire size and reliability of self-non-self discrimination. Journal of
Theoretical Biology, 81(4):645–670–, December 1979.
[322] A. S. Perelson and G. Weisbuch. Immunology for physicists. Reviews of Modern

Physics, 69(4):12191268–, 1997.
[323] C. C. Pettey. Diffusion (cellular) models. In David B Fogel and Zbigniew Michalwicz,
editors, Evolutionary Computation 2 - Advanced Algorithms and Operations, pages –.
Institute of Physics (IOP) Publishing Ltd., UK, 2000.
[324] L. J. Picker and E. C. Butcher. Physiological and molecular mechanisms of lympho-

cyte homing. Annual Review of Immunology, 10:561–591–, 1992.
369
[325] D. M. Pierre, D. Goldman, Y. Bar-Yam, and A. S. Perelson. Somatic evolution in
the immune system: the need for germinal centers for efficient affinity maturation.
Journal of Theoretical Biology, 186(2):159–171–, 1997.
[326] E. Plaza and J. Llus A. F. Martn. Cooperative case-based reasoning. In Distributed

Artificial Intelligence Meets Machine Learning Learning in Multi-Agent Environ-
ments, 1997.
[327] R. Polikar. Ensemble based systems in decision making. IEEE Circuits and Systems
Magazine, 6(3):21–45, 2006.
[328] M. J. D. Powell. Radial basis functions for multivariable interpolation: a review.

In Algorithms for approximation, pages 143–167–. Clarendon Press, New York, NY,
USA, 1987.
[329] W. L. Price. A controlled random search procedure for global optimisation. The
Computer Journal, 20:367–370, 1977.
[330] Computational Science Education Project. Mathematical optimization. Online,

1995.
[331] R. C. Purshouse and P. J. Fleming. Conflict, harmony, and independence: Relation-

ships in evolutionary multi-criterion optimisation. In Proceedings of the Second In-
ternational Conference on Evolutionary Multi-Criterion Optimization (EMO), pages
16–30, 2003.
[332] K. Rajewsky. Immunology: Burnet’s unhappy hybrid. Nature, 394:624–625–, August

1998.
[333] I. Rechenberg. Evolutionsstrategie: Optimierung Technischer Systeme nach Prinzip-

ien der Biologischen Evolution. Frommann-Holzboog, 1973.
[334] R. D. Reed and R. J. Marks II. Neural Smithing - Supervised Learning in Feedforward
Artificial Neural Networks. MIT Press, USA, 1999.
[335] R. R. Regoes, M. A. Nowak, and S. Bonhoeffer. Evolution of virulence in a hetero-

geneous host population. Evolution Int J Org Evolution, 54(1):64–71, Feb 2000.
[336] L. Riechmann, M. Clark, H. Waldmann, and G. Winter. Reshaping human antibod-

ies for therapy. Nature, 332:323–327–, 1988.
[337] B. Rinkevich. Primitive immune systems: Are your ways my ways? Immunological
Reviews, 198(1):2535–, April 2004.
[338] H. Robbins. Some aspects of the sequential design of experiments. Bull. Amer.
Math. Soc., 58:527–535, 1952.
[339] H. Rosen, G. Sanna, and C. Alfonso. Egress: a receptor-regulated step in lymphocyte

trafficking. Immunological Reviews, 195(1):160177–, 2003.
[340] I. M Rouzine, J. Wakeley, and J. M Coffin. The solitary wave of asexual evolution.
Proc Natl Acad Sci U S A, 100(2):587–592, Jan 2003.
[341] D. E. Rumelhart and D. Zipser. Feature discovery by competitive learning. Cognitive

Science, 9(1):75–112–, 1985.
[342] S. J. Russell. and P Norvig. Artificial Intelligence: A Modern Approach. Prentice

Hall, 1995.
370
[343] M. Salmi and S. Jalkanen. Lymphocyte homing to the gut: attraction, adhesion,
and commitment. Immunological Reviews, 206(1):100–113–, 2005.
[344] D. Salmon, S. Teret, C. MacIntyre, D. Salisbury, M. Burgess, and N. Halsey. Com-

pulsory vaccination and conscientious or philosophical exemptions: past, present,
and future. The Lancet, 367(9508):436–442–, 2006.
[345] S. Sarafijanovic and J. Y. L Boudec. An artificial immune system for misbehavior

detection in mobile ad-hoc networks with virtual thymus, clustering, danger signal,
and memory detectors. In Vincenzo Cutello, Peter J. Bentley, and Jon Timmis,
editors, Lecture Notes in Computer Science, pages 342–356–. Springer, Berlin /
Heidelberg, Germany, 2004.
[346] S. Sarafijanovic and J. Y. L Boudec. An artificial immune system approach with

secondary response for misbehavior detection in mobile ad hoc networks. IEEE
Transactions on Neural Networks, 16(5):1076–1087–, September 2005.
[347] J. David Schaffer, Richard A. Caruana, Larry J. Eshelman, and Rajarshi Das. A
study of control parameters affecting online performance of genetic algorithms for
function optimization. In Proceedings of the third international conference on Ge-
netic algorithms, pages 51–60–. Morgan Kaufmann Publishers Inc., San Francisco,
CA, USA, 1989.
[348] J. David Schaffer and Larry J. Eshelman. On crossover as an evolutionarily viable

strategy. In Lashon B. Booker, editor, Proceedings of the 4th International Confer-
ence on Genetic Algorithms, pages 61–68–. Morgan Kaufmann, 1991.
[349] R. E. Schapire. The strength of weak learnability. Machine Learning, 5(2):197–227,

1992.
[350] R. E. Schapire. The boosting approach to machine learning: An overview. In

D. D. Denison, M. H. Hansen, C. Holmes, B. Mallick, and B. Yu, editors, Nonlinear
Estimation and Classification, 2003.
[351] T. Schnekenburger. Parallel randomized algorithms in heterogeneous environments.

In Int. Conference on Systems Engineering, 1993.
[352] H. P. Schwefel. Numerical optimization of computer models. John Wiley & Sons,
Inc., New York, NY, USA, 1981.
[353] A. Seeker and A. A. Freitas. Wairs: improving classification accuracy by weighting

attributes in the airs classifier. In Proc. IEEE Congress on Evolutionary Computa-
tion CEC 2007, pages 3759–3765, 25–28 Sept. 2007.
[354] L. A. Segel. Some spatio-temporal models in immunology. International Journal of

Bifurcation and Chaos, 12(11):2343–2347, 2002.
[355] L. A. Segel and A. S. Perelson. Computations in shape space: A new approach

to immune network theory. Theoretical Immunology, Part Two, SFI Studies in the
Sciences of Complexity, pages 377–401–, 1988.
[356] M. Shannon and R. Mehr. Reconciling repertoire shift with affinity maturation: The
role of deleterious mutations. The Journal of Immunology, 162:3950–3956–, 1999.
[357] A. J. C. Sharkey. Combining Artificial Neural Nets: Ensemble and Modular Multi-
Net Systems. Springer, 1999.
371
[358] A. M. Silverstein. The clonal selection theory: what it really is and why modern
challenges are misplaced. Nature Immunology, 3:793–796–, 2002.
[359] S. P. N. Singh and S. M. Thayer. Immunology directed methods for distributed
robotics: A novel, immunity-based architecture for robust control & coordination.
In Proceedings of SPIE: Mobile Robots XVI, pages –, 2001.
[360] S. P. N. Singh and S. M. Thayer. Kilorobot search and rescue using an immunolog-
ically inspired approach. Distributed Autonomous Robotic Systems, 5:–, June 2002.
[361] S. P. N. Singh, S. M. Thayer, and W. P. Thayer. A foundation for kilorobotic
exploration. In Proceedings of the 2002 Congress on Evolutionary Computation,
CEC ’02., volume 2, pages 1033–1038–. IEEE Press, USA, 2002.
[362] D. Skillicorn. Strategies for parallel data mining. IEEE Concurrency, 7(4):26–35,
1999.
[363] H. B. Slade and S. A. Schwartz. Mucosal immunity: The immunology of breast milk.
Journal of Allergy and Clinical Immunology, 80(3, Part 1):348–358–, September
1987.
[364] A. Sloman. Must intelligent systems be scruffy? In G T McKee and G C Dean,
editors, Evolving Knowledge in Natural Science and Artificial Intelligence, pages –.
Pitman, London, 1990.
[365] D. J Smith, S. Forrest, and A. S. Perelson. Immunological memory is associative.
In ICMAS ‘96: Multiagent systems workshop on immunity-based systems, pages 62–
70–, United States, 1996.
[366] J. E. Smith and T. C. Fogarty. Operator and parameter adaptation in genetic algo-
rithms. Soft Computing - A Fusion of Foundations, Methodologies and Applications,
1:81–87, 1997.
[367] W. M. Spears. Crossover or mutation? In Proceedings of Foundations of Genetic
Algorithms Workshop, volume 2, pages 221–237–. Morgan Kaufmann, San Mateo,
CA, USA, 1992.
[368] W. M. Spears. The Role of Mutation and Recombination in Evolutionary Algorithms.
PhD thesis, George Mason University, Fairfax, Virginia, USA, 1998.
[369] W. Srikusalanukul, F. De Bruyne, and P. McCullagh. Modelling of peripheral lym-
phocyte migration: System identification approach. Immunology and Cell Biology,
78(3):288–, June 2000.
[370] D. J. Stekel. The simulation of density-dependent effects in the recirculation of t
lymphocytes. Scandinavian journal of immunology, 47(5):426–430, 1998.
[371] D. J. Stekel, C. E. Parker, and M. A. Nowak. A model of lymphocyte recirculation.
Immunology Today, 18(5):216–221, May 1997.
[372] S. Stepney, R. E. Smith, J. Timmis, and A. M. Tyrrell. Towards a conceptual
framework for artificial immune systems. In Vincenzo Cutello, Peter J. Bentley, and
Jon Timmis, editors, Lecture Notes in Computer Science (LNCS), pages 53–64–.
Springer-Verlag, Germany, 2004.
[373] S. Stepney, R. E. Smith, J. Timmis, A. M. Tyrrell, M. J. Neal, and Andrew N. W.
Hone. Conceptual frameworks for artificial immune systems. International Journal
of Unconventional Computing, 1(3):315–338–, July 2005.
372
[374] T. Stibor. On the Appropriateness of Negative Selection for Anomaly Detection
and Network Intrusion Detection. PhD thesis, Darmstadt University of Technology,
Germany, 2006.
[375] T. Stibor, J. Timmis, and C. Eckert. On the appropriateness of negative selection

defined over hamming shape-space as a network intrusion detection system. In The
2005 IEEE Congress on Evolutionary Computation, (CEC 2005), volume 2, pages
995–1002–. IEEE Press, USA, 2005.
[376] R. S. Sutton and A. G. Barto. Reinforcement Learning: An Introduction. The MIT

Press, Cambridge, Massachusetts; London, England, 1998.
[377] M. A. Swartz. The physiology of the lymphatic system. Advanced drug delivery
reviews, 50(1-2):3–20–, 2001.
[378] H. Takagi. Interactive evolutionary computation: fusion of the capabilities of ec

optimization and human evaluation. Proceedings of the IEEE, 9(89):1275–1296,
September 2001.
[379] E. Talbi. A taxonomy of hybrid metaheuristics. Journal of Heuristics, 8:541–564,

2001.
[380] D. W. Talmage. Allergy and immunology. Annual Review of Medicine, 8:239–256–,

1957.
[381] D. W. Talmage. Immunological specificity: Unique combinations of selected natural

globulins provide an alternative to the classical concept. Science, 129:1643–1648–,
1959.
[382] R. Tanese. Distributed genetic algorithms. In Proceedings of the third international

conference on Genetic algorithms, 1989.
[383] J. C. Tay and A. Jhavar. Cafiss: a complex adaptive framework for immune system
simulation. In Proceedings of the 2005 ACM symposium on Applied computing, pages
158–164–. ACM Press, New York, NY, USA, 2005.
[384] S. M. Thayer and S. P. N. Singh. Development of an immunology-based multi-

robot coordination algorithm for exploration and mapping domains. In IEEE/RSJ
International Conference on Intelligent Robots and System, volume 3, pages 2735–
[385] P. Theodosopoulos and T. Theodosopoulos. Quantitative biology, abstract. Techni-

cal report, arXiv, arXiv.org, June 2004.
[386] P. Theodosopoulos and T. Theodosopoulos. Robustness and evolvability of the b

cell mutator mechanism. Technical report, arXiv, arXiv.org, June 2005.
[387] G. J. Thorbecke, A. R. Amin, and V. K. Tsiagbe. Biology of germinal centers in

lymphoid tissue. The The Federation of American Societies for Experimental Biology
Journal (FASEB), 8:832–840–, 1994.
[388] J. Timmis. Artificial immune systems - today and tomorrow. Natural Computing,
6(1):1–18–, March 2007.
[389] J. Timmis, M. Amos, W. Banzhaf, and A. Tyrrell. ”going back to our roots”: second
generation biocomputing. Technical Report arXiv:cs/0512071v1, arXiv, Dec 2005.
373
[390] J. Timmis, C. Edmonds, and J. Kelsey. Assessing the performance of two immune
inspired algorithms and a hybrid genetic algorithm for function optimisation. In
Proceedings of the Congress on Evolutionary Computation (CEC04), volume 1, pages
1044–1051–. IEEE Press, USA, 2004.
[391] J. Timmis and M. Neal. Investigating the evolution and stability of a resource limited
artificial immune system. In Workshop Program, pages 40–41–. AAAI Press, USA,
2000.
[392] J. Timmis, M. Neal, and J. Hunt. An artificial immune system for data analysis.
Biosystems, 55(1):143–150–, 2000.
[393] J. Timmis and M. J. Neal. A resource limited artificial immune system for data
analysis. Knowledge Based Systems Journal: Special Issue, 14(3-4):121–130–, 2001.
[394] K. M. Ting and I. H. Witten. Issues in stacked generalization. Journal of Artificial

Intelligence Research, 10:271–289, 1999.
[395] N. Toma, S. Endo, K Yamada, and H. Miyagi. The immune distributed competitive
problem solver with majorhistocompatibility complex and immune network. In IEEE
International Conference on Systems, Man, and Cybernetics, volume 3, pages 1865–
[396] S. Tonegawa. Somatic generation of antibody diversity. Nature, 302(5909):575–581–,

April 1983.
[397] S. Tonegawa. Somatic generation of immune diversity. Bioscience Reports, Biomed-

ical and Life Sciences, Nobel Lecture, 8(1):3–26–, February 1988.
[398] F. Trepel. Number and distribution of lymphocytes in man. a critical analysis.

Klinische Wochenschrift, 52(11):511–515–, 1974.
[399] E. Tsang. Foundations of Constraint Satisfaction. Academic Press, 1993.
[400] J. Twycross and U. Aickelin. Towards a conceptual framework for innate immunity.
In LNCS, pages 112–125–. Springer, Germany, 2005.
[401] J. Twycross and U. Aickelin. Biological inspiration for artificial immune systems.
In Proceedings of the 6th International Conference on Artificial Immune Systems
(ICARIS-2007), pages –, 2007.
[402] A. Trn, M.M. Ali, and S. Viitanen. Stochastic global optimization: Problem classes
and solution techniques. Journal of Global Optimization, 14:437–447, 1999.
[403] F. J. Varela. A cognitive view of the immune system. World Futures, 42(1-2):31–40–,
1994.
[404] P. A. Vargas, L. N. de Castro, R. Michelan, and F. J. Von Zuben. An immune

learning classifier network for automated navigation. In Peter J. Bentley and Emma
Hart, editors, Lecture Notes in Computer Science, pages 69–80–. Springer, Berlin
Heidelberg, 2003.
[405] P. A. Vargas, L. N. de Castro, and F. J. Von Zuben. Artificial immune systems

as complex adaptive systems. In Proc. of ICARIS (International Conference on
Artificial Immune Systems), pages 115–123, 2002.
374
[406] P. A. Vargas, L. N. de Castro, and F. J. Von Zuben. Mapping artificial immune
systems into learning classifier systems. In Lecture Notes in Computer Science, pages
163–186–. Springer, Berlin / Heidelberg, 2003.
[407] T. J. Vaughan, J. K. Osbourn, and P. R. Tempest. Human antibodies by design.

Nature Biotechnology, 16:535–539–, 1998.
[408] M. Villalobos-Arias, C. A. Coello Coello, and O. Hernndez-Lerma. Convergence

analysis of a multiobjective artificial immune system algorithm. In Vincenzo Cutello,
Peter J. Bentley, and Jon Timmis, editors, Lecture Notes in Computer Science, pages
[409] P. Vineis. Cancer as an evolutionary process at the cell level: an epidemiological

perspective. Carcinogenesis, 24(1):1–6, Jan 2003.
[410] P. Vineis and M. Berwick. The population dynamics of cancer: a darwinian per-
spective. Int J Epidemiol, 35(5):1151–1159, Oct 2006.
[411] U. H. von Andrian and T. R. Mempel. Homing and cellular traffic in lymph nodes.
Nature reviews. Immunology, 3(11):867–878–, 2003.
[412] C. von der Malsburg. Self-organization of orientation sensitive cells in the striate
cortex. Biological Cybernetics, 14(2):85–100–, December 1973.
[413] S. Voulgaris, M. Jelasity, and M. van Steen. A robust and scalable peer-to-peer
gossiping protocol. In Agents and Peer-to-Peer Computing, 2005.
[414] T. A. Waldmann. Immunotherapy: past, present and future. Nature Medicine,

9(3):269–277–, 2003.
[415] J. H. Walker and S. M. Garrett. Dynamic function optimisation: Comparing the

performance of clonal selection and evolutionary strategies. In Lecture Notes in
Computer Science, pages 273–284–. Springer-Verlag, Germany, 2003.
[416] R. Aaron Warnock, Sanaz Askari, Eugene C. Butcher, and Ulrich H. von Andrian.
Molecular mechanisms of lymphocyte homing to peripheral lymph nodes. Journal
of Experimental Medicine (JEM), 182(2):205–216–, 1998.
[417] A. Watkins, X. Bi, and A. Phadke. Parallelizing an immune-inspired algorithm for

efficient pattern recognition. In A. Buczak, J. Ghosh, M. Embrechts, and O. Ersoy,
editors, Fuzzy Logic, Evolutionary Programming, Complex Systems and Artificial
Life, volume 13, pages 225–230–. ASME Press, New York, USA, 2003.
[418] A. Watkins and L. Boggess. A new classifier based on resource limited artificial
immune systems. In Part of the 2002 IEEE World Congress on Computational
Intelligence, pages 1546–1551–. IEEE, May 2002.
[419] A. Watkins and J. Timmis. Artificial immune recognition system (airs): Revisions
and refinements. In P.J. Bentley, editor, 1st International Conference on Artificial
Immune Systems (ICARIS2002), pages 173–181–. University of Kent at Canterbury
Printing Unit, UK, 2002.
[420] A. Watkins and J. Timmis. Exploiting parallelism inherent in airs, an artificial

immune classifier. In V. Cutello, P. Bentley, and J. Timmis, editors, Lecture Notes
in Computer Science (LNCS), volume 3239, pages 427–438–. Springer-Verlag GmbH,
2004.
375
[421] A. Watkins, J. Timmis, and L. Boggess. Artificial immune recognition system
(airs): An immune-inspired supervised learning algorithm. Genetic Programming
and Evolvable Machines, 5(3):291–317–, September 2004.
[422] A. B. Watkins. Airs: A resource limited artificial immune classifier. Master’s thesis,
Mississippi State University, USA, 2001.
[423] A. B. Watkins and L. C. Boggess. A resource limited artificial immune classifier.

In Part of the 2002 IEEE World Congress on Computational Intelligence held in
Honolulu, pages 926–931–, USA, May 2002. IEEE Computer Society.
[424] R. G. Weinand. Somatic mutation, affinity maturation and the antibody repertoire:
a computer model. Journal of Theoretical Biology, 143(3):343–382–, 1990.
[425] G. Weiss. Multiagent Systems: A Modern Approach to Distributed Artificial Intelli-

gence. MIT Press, 1999.
[426] D. Wettschereck, D. W. Aha, and T. Mohri. A review and empirical evaluation of

feature weighting methods for a class of lazy learning algorithms. Artificial Intelli-
gence Review, 11(1-5):273–314–, February 1997.
[427] J. White and S. M. Garrett. Improved pattern recognition with artificial clonal
selection? In Lecture Notes in Computer Science, pages 181–193–. Springer-Verlag,
Berlin / Heidelberg, Germany, 2003.
[428] R. H. White. Competitive hebbian learning. In Proceedings of the International

Joint Conference on Neural Networks (IJCNN-91), volume 2, pages 949–. IEEE
Press, USA, 1991.
[429] R. H. White. Competitive hebbian learning: algorithm and demonstrations. Neural

Networks, 5(2):261–275–, 1992.
[430] R. H. White. Recurrent competitive hebbian learning. In Proceedings of the Inter-

national Joint Conference on Neural Networks (IJCNN), volume 4, pages –. IEEE
Press, USA, 1992.
[431] D. Whitley, S. Rana, J. Dzubera, and K E. Mathias. Evaluating evolutionary algo-

rithms. Artificial Intelligence - Special volume on empirical methods, 85(1-2):245–
276–, 1996.
[432] P. D. Williams, K. P. Anchor, J. L. Bebo, G. H. Gunsch, and G. D. Lamont. Cdis:

Towards a computer immune system for detecting network intrusions. In L. M and
A. Wespi, editors, Lecture Notes in Computer Science, pages 117–133–. Springer,
Berlin / Heidelberg, 2001.
[433] I. H. Witten and E. Frank. Data Mining: Practical machine learning tools with Java
implementations. Morgan Kaufmann, San Francisco, 2000.
[434] D. H. Wolpert. Stacked generalization. Neural Networks, 5(LA-UR-90-3460):241–

259, 1992.
[435] D. H. Wolpert and W. G. Macready. No free lunch theorems for search. Technical
report, Santa Fe Institute, Sante Fe, NM, USA, 1995.
[436] D. H. Wolpert and W. G. Macready. No free lunch theorems for optimization. IEEE
Transactions on Evolutionary Computation, 1(67):67–82–, 1997.
376
[437] S. Wright. The roles of mutation, inbreeding, crossbreeding, and selection in evolu-
tion. Proceedings of the Sixth International Congress on Genetics, pages 355–366–,
1932.
[438] L J Wysocki and M L Gefter. Gene conversion and the generation of antibody
diversity. Annual Review of Biochemistry, 58:July 1989–, July 1989.
[439] L. Xu, A. Krzyzak, and E. Oja. Rival penalized competitive learning for cluster-
ing analysis, rbfnet, and curve detection. IEEE Transactions on Neural Networks,
4(4):636–649–, July 1993.
[440] X. Yao. A review of evolutionary artificial neural networks. International Journal

of Intelligent Systems, 8(4):539–567, 1993.
[441] X. Yao, Y. Liu, and G. Lin. Evolutionary programming made faster. IEEE Trans-
actions on Evolutionary Computation, 3(2):82–102, July 1999.
[442] A. J. Young and J. Hay. Lymphocyte migration in the development and disease.
Seminars in Immunology, 11(2):71–, 1999.
[443] Y. Yu and C. Z. Hou. A clonal selection algorithm by using learning operator. In

Proceedings of 2004 International Conference on Machine Learning and Cybernetics,
volume 5, pages 2924–2929–. IEEE Press, USA, 2004.
[444] M. J. Zaki. Parallel and distributed data mining: An introduction. In Revised

Papers from Large-Scale Parallel Data Mining, Workshop on Large-Scale Parallel
KDD Systems, SIGKDD, pages 1–23, 1999.
[445] Y. Zhang and C. Hou. A clone selection algorithm with niching strategy inspiring by
biological immune principles for change detection. In IEEE International Symposium
on Intelligent Control, pages 1000–1005–. IEEE Press, USA, 2003.
[446] R. M. Zinkernagel, S. Ehl, P. Aichele, S. Oehen, T. Kundig, and H. Hengartner. Anti-

gen localisation regulates immune responses in a dose- and time-dependent fashion:
a geographical view of immune reactivity. Immunological Reviews, 156:199–209–,
1997.
377

Jason Brownlee Thesis PDF

Uploaded by

Copyright:

Available Formats

Jason Brownlee Thesis PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Jason Brownlee Thesis PDF

Uploaded by

Copyright:

Available Formats

Clonal Selection as an Inspiration for

Adaptive and Distributed Information

A dissertation presented for fulfillment of the requirements for the award of

The development of adaptive and intelligent computational methods is an important fron-

Professionally, I would like to thank:

Personally, I would like to thank:

3 Clonal Selection Paradigm 35

4 Cellular Clonal Selection 76

6 Host Clonal Selection 209

7 Hierarchical Frameworks 261

8 Suitability and Application 290

A Tables of Empirical Results 332

B Works Arising from the Study 342

4.1 Depiction of external stimulation resulting in internal activation. . . . . . . 80

8.1 Depiction that the suitability of application of a strategy to a problem is

3.1 Clonal selection algorithm taxonomy overview. . . . . . . . . . . . . . . . . 46

4.1 List of goals for a colour-space domain. . . . . . . . . . . . . . . . . . . . . 87

5.1 Summary of information processing needs, and proposed algorithms that

6.1 Summary of the diﬀerent types of immunity. . . . . . . . . . . . . . . . . . . 210

An in depth study of immunological structure and function, in particular the

1. Identify a systematic methodology for realising a novel biological inspired

3. Identify immunological structures and/or functions which clearly exhibit

4. Study the behaviours, capabilities, and limitations of devised computational

5. Propose an integrated clonal selection framework that unites the base

5. The identification and abstraction of host immunisation and evolution as a frame-

7. The integration of the three perspectives of the computational properties of clonal

1.5 Thesis Organisation

Chapter 2 provides background documentation on the general fields of study in which

Chapter 6 considers the clonal selection paradigm in the context of a population of

Chapter 7 provides a systematic integration firstly of the computational clonal selec-

Chapter 8 considers the applicability of the developed clonal selection algorithms. A

Chapter 9 provides a detailed assessment of the research hypothesis by highlighting the

2.1 Chapter Overview

2.2 Artificial and Computational Intelligence

2.2.2 Computational Intelligence

Swarm Intelligence A paradigm that considers collective intelligence emerges through

Connectionist Intelligence An approach that is concerned with the investigation of

2.2.3 Natural Computation

Biologically Inspired Computation (Computation inspired by biological metaphor )

Computationally Motivated Biology (Biology with digital computers) The intent of

2.3 Natural and Artificial Immune Systems

2.3.1 Overview of the Immune System

2.3.2 Artificial Immune Systems

Artificial Immune Systems (AIS) is a sub-field Computational Intelligence motivated by

• Disposability: There is no single cell or molecule that is essential to the functioning

• Autonomy: There is no central or coordinating organ controlling the immune system

• Fault Tolerance: The complementary roles performed by several immune components

• Self-Organisation: There is no specific information as to how to respond to a given

2.4 Three Schools of AIS

2.4.1 Clonal Selection Paradigm

2.5 Distributed Artificial Immune Systems

2.5.1 Promise of Distributed Information Processing

2.5.2 Algorithms, Systems, and Applications

Distributed Detectors for Computer Security

Control, Robots, and Autonomous Navigation

2.6 Open Problems in AIS

2.6.1 Open Problems

2.7.1 Conceptual Framework

Although a progression from inspiring biological system to inspired computation system