To appear in the Proceedings of EuroGP 2005

Extending Particle Swarm Optimisation via

Genetic Programming

Riccardo Poli, William B. Langdon, and Owen Holland

Department of Computer Science, University of Essex, UK

Abstract. Particle Swarm Optimisers (PSOs) search using a set of in-

teracting particles flying over the fitness landscape. These are typically
controlled by forces that encourage each particle to fly back both to-
wards the best point sampled by it and towards the swarm’s best. Here
we explore the possibility of evolving optimal force generating equations
to control the particles in a PSO using genetic programming.

1 Introduction
The class of complex systems sometimes referred to as swarm systems is a rich
source of novel computational methods that can solve difficult problems effi-
ciently and reliably. When swarms solve problems in nature, their abilities are
usually attributed to swarm intelligence; perhaps the best-known examples are
colonies of social insects such as termites, bees and ants. In recent years it has
proved possible to identify, abstract and exploit the computational principles
underlying some forms of swarm intelligence, and to deploy them for scientific
and industrial purposes. One of the best-developed techniques of this type is
Particle Swarm Optimisation (PSO) [9].
In PSOs, which are inspired by flocks of birds and shoals of fish, a number
of simple entities — the particles — are placed in the parameter space of some
problem or function, and each evaluates the fitness at its current location. Each
particle then determines its movement through the parameter space by combin-
ing some aspect of the history of its own fitness values with those of one or more
members of the swarm, and then moving through the parameter space with a
velocity determined by the locations and processed fitness values of those other
members, along with some random perturbations. The next iteration takes place
after all particles have been moved. Eventually the swarm as a whole, like a flock
of birds collectively foraging for food, is likely to move close to the best location.
This simple model can deal with difficult problems efficiently. Naturally, dif-
ferent variations of the basic recipe have been tried and compared to existing
techniques and different application areas have been investigated. However, the
situation overall is still as it was in 2001 when Kennedy and Eberhart wrote:
“...we are looking at a paradigm in its youth, full of potential and fertile with
new ideas and new perspectives...Researchers in many countries are experiment-
ing with particle swarms...Many of the questions that have been asked have not
yet been satisfactorily answered.”[9]
2 R. Poli, W.B. Langdon and O. Holland

This research is part of a project that aims to systematically explore the ex-
tension of particle swarms by including strategies from biology, by extending the
physics of the particles, and by providing a solid theoretical and mathematical
basis for the understanding and problem-specific design of new particle swarm
algorithms. In this paper, in an effort to extend PSO models beyond real biol-
ogy and physics and push the limits of swarm intelligence into the exploration
of swarms as they could be, we study the possibility of evolving, through the use
of genetic programming (GP) [10, 12], the force generating equations to control
the particles in a PSO. This gives us a methodology for routinely “inventing”
specialised PSOs which are near-optimum for specific domains. Our aim is to
verify the feasibility of this approach and to understand, through the analysis of
the evolved components, what types of PSOs are best for different landscapes.
Section 2 provides a review of the work to date on particle swarms, while
Section 3 describes how to use GP to automatically generate PSO tailored to
particular tasks. The results section (4) is followed, in Section 5, by a brief
restatement of our findings and future direction.

2 Particle Swarm Optimisation

The initial ideas of James Kennedy (a social psychologist) and Russ Eberhart
(an engineer and computer scientist) were essentially aimed at producing compu-
tational intelligence by exploiting simple analogues of social interaction, rather
than purely individual cognitive abilities. Their 1995 simulations, influenced by
Heppner’s work [7], involved analogues of bird flocks searching for corn. Their
continuing emphasis on the social nature of intelligence, even in humans, can be
seen in their book, Swarm Intelligence [9].
In the simplest (and original) version of PSO, each particle is moved by two
elastic forces, one attracting it with random magnitude to the fittest location so
far encountered by the particle, and one attracting it with random magnitude
to the best location encountered by any member of the swarm.1 Suppose we are
dealing with an N dimensional problem. Each particle’s position, velocity and
acceleration, can each be represented as a vector with N components (one for
each dimension). Starting with the acceleration vector, a = (a1 , · · · , aN ), each
component, ai , is given by

ai = ψ1 R1 (xsi − xi ) + ψ2 R2 (xpi − xi )

where xsi is the ith component of the best point visited by the swarm, xi is
the ith component of the particle’s current location, xpi is the ith component of
its personal best, R1 and R2 are two independent random variables uniformly
distributed in [0,1] and ψ1 and ψ2 are two learning rates. ψ1 and ψ2 control the
relative proportion of cognition and social interaction in the swarm. The same
formula is used independently for each dimension of the problem.
1
Here we limit ourselves to the case where particles can socially interact freely. More
general models constrain inter-particle interactions via a neighbourhood structure.
 Extending Particle Swarm Optimisation via GP 3

The velocity of a particle v = (v1 , · · · , vn ) and its position are updated every
time step using the equations:

vi (t) = vi (t − 1) + ai xi (t) = xi (t − 1) + vi (t)

This system can lead the particles to become unstable, with their speed increas-
ing without control. This is harmful to the search and need to be controlled. The
standard technique is to bound velocities so that vi ∈ [−Vmax , +Vmax ].
Early variations in PSO techniques involved the addition of analogues of
physical characteristics to the members of the swarm, such as the “inertia weight”
ω [17], where the velocity update equation is modified as follows:

vi (t) = ωvi (t − 1) + ai

In vector notation, the velocity change can be written as ∆v = a − (1 − ω)v.

That is, the constant 1 − ω acts effectively a friction coefficient.
Following Kennedy’s graphical examinations of the trajectories of individual
particles and their responses to variations in key parameters [8] the first real
attempt at providing a theoretical understanding of PSO was the “surfing the
waves” model presented by Ozcan [14]. Shortly afterwards, Clerc developed a
comprehensive 5-dimensional mathematical analysis of the basic PSO system [5].
A particularly important contribution of that work was the use and analysis of
a modified update rule:

vi (t) = κ(vi (t − 1) + ai )

where the constant κ is called a constriction coefficient. If κ is correctly chosen,

it guarantees the stability of the PSO without the need to bound velocities.
More recently, there have been further explorations of physics-based effects
in the swarm. For example, Blackwell has investigated quantum swarms [3] and
charged particles [2], and Poli and Stephens have proposed a scheme in which
the particles do not “fly above” the fitness landscape, but actually slide over it
[15]. Krink and collaborators have looked at a range of modifications of PSO,
including ideas from physics (spatially extended particles [11], self-organised
criticality [13]) and biology (e.g. division of labour [16]). Finally, there has also
been cross-fertilisation between PSOs and evolutionary computing. For example,
Angeline [1] introduced selection and Brits et al. [4] have explored niching.

3 Evolution of PSOs via Genetic Programming

In the original PSO and well as in most improvements proposed in the literature
(see previous section), the equation controlling the particles is of the form:

ai = F (xi , xsi , xpi , vi )

for some function F . Our approach to explore the space of possible PSOs is to
use GP to evolve the function F so as to maximise some performance measure.
4 R. Poli, W.B. Langdon and O. Holland

Clearly, we cannot expect to be able to evolve a PSO that can beat all other
PSOs on all possible problems [18]. We should, however, be able to evolve PSOs
that can outperform known PSOs on specific classes of problems of interest.
Evolving specialised search algorithms is typically a heavy computational
task. So, a very efficient implementation of both PSO and GP were required for
this work. Since nothing particularly fancy is required of the GP environment
(except efficiency), we used a small and highly efficient C implementation of GP.
For efficiency, we implemented our own minimalist PSO engine in C. This has
to be compact and efficient in that it is invoked multiple times and for many
time steps during the fitness evaluation of each GP program.
The function set for GP included the functions +, −, × and the protected
division DIV. I.e. if |y| <= 0.001 DIV(x, y) = x else DIV(x, y) = x/y. The
terminal set included the coordinate of a particle xi , the corresponding compo-
nent of the velocity vi , the coordinate of the best point visited by the particle
xpi , the coordinate of the best point visited by the swarm xsi , the numerical
constants 1.0, -1.0, 0.5, -0.5, and finally a zero-arity function R which returns
random numbers uniformly distributed within the range [−1, 1].
In order to evolve PSOs that are able to solve a class of problems as opposed
to just one problem, we need to build a fitness function which uses the program
being evaluated as the F function in a PSO, and evaluates the performance of
the resulting PSO on a training set of problems taken from the given class.
In our study we considered two classes of benchmark problems, the city-
block sphere problem class and the Rastrigin’s problem class, of two different
dimensions, N = 2 and N = 10. Problem instances from the city-block sphere
class have the following form:
N
X
f (x) = |xi − ci |.
i=1

Every city-block problem has a single global optimum (at x = (c1 , . . . , cN ),

where f (x) = 0) and no local optima. Problem instances from the Rastrigin’s
class have the following form:
N
X
(xi − ci )2 − 10 cos(2π(xi − ci ))


f (x) = 10N +
i=1

This has many local optima and one global optimum x = (c1 , . . . , cN ) with
f (x) = 0.
At the beginning of each GP run, 10 random problems were generated from
the chosen class of functions (either the city-block sphere or the Rastrigin func-
tion class) by choosing the values ci uniformly at random from the range [−1, 1].
To limit the computational load of the simulations, during fitness evaluation we
used PSOs with 10 particles and run them for only 30 iterations on each of the
10 problems. The initial coordinates for the particles were drawn uniformly at
random from the interval [−5, 5]. Since performance can vary substantially with
the initial random position of the particles, for each problem the PSO was run 5
 Extending Particle Swarm Optimisation via GP 5

times with different initial random positions for the 10 particles. Initial particle
velocity is 0. To ensure stability we updated velocities using Clerc’s update rule
(see previous section), using a constriction factor κ = 0.7. We also clipped the
components vi of particle velocities within the range [−2.0, +2.0].
In each of the 10 × 5 = 50 training cases, the performance of the PSO
needs to be evaluated. More than one performance measure could be used. For
example, if one wanted to encourage the convergence of the swarm at the global
optimum, performance could be evaluated as the sum of the distances between
the global optimum and each particle at the end of the 30 PSO iterations. If one
is only interested in the ability of the PSO to find the global optimum, then the
distance between the swarm best and the global optimum at the end of the PSO
iterations should be used as a performance measure. If one does not care about
global optima, but is only interested in achieving good values of the objective
function, then the difference between the best objective function value observed
in a PSO run and the objective function value at the global optimum could be
used as a performance measure. Fitness functions that encourage the swarm not
to collapse onto the swarm best could be defined for dynamic problems which
require the ability to track moving optima. And so on.
We experimented with two fitness functions: a) we measured and accumulated
(over 50 fitness cases) the (city-block) distance between the swarm best and the
global optimum at the end of each PSO run
X
|xsi − ci |
i

and b) we measured and accumulated (over 50 fitness cases and 10 particles) the
(city-block) distance between each particle’s position and the global optimum at
the end of each PSO run
XX
|xi − ci |
x i

The negation of either (a) or (b) minus a parsimony pressure term (see below)
was returned as the fitness of the GP program controlling the PSO.
In the GP system we used steady state binary tournaments for parent selec-
tion and binary negative tournaments to decide who to remove from the popu-
lation. The initial population was created using the “grow” method with max
depth of 6 levels (the root node being at level 0). We used population sizes of
1000 and 5000 individuals. We used 90% standard sub-tree crossover (with uni-
form random selection of crossover points) and 10% point mutation with a 2%
chance of mutation per tree node. Runs were terminated either manually when
fitness appeared to be sufficiently good or automatically at generation 100. To
favour readability and understandability of the evolved solutions, a mild parsi-
mony pressure (parsimony coefficient=0.01) was applied to the fitness function
to encourage the evolution of a simpler F .
6 R. Poli, W.B. Langdon and O. Holland

4 Results

In order to be able to evaluate the PSOs produced by GP, we compared them

with a number of human-designed update rules, most of which have previously
appeared in the literature. The update rules included:

PSO a version of the standard PSO where ψ1 = ψ2 = 1.0, that is

˙ p − xi ) + R2 (x
F = R1 (x ˙ s − xi )
i i

PSOD1 a deterministic (no random coefficients) 100% social version (ψ1 = 0,

ψ2 = 1) of the standard PSO:

F = (xsi − xi )

PSOR0 a PSO controlled by random forces

F = 2.0Ṙ − 1.0

PSOR1 a 100% social (ψ1 = 0, ψ2 = 1) version of the standard PSO

˙ s − xi )
F = R(x i

In our GP runs we evolved several high performance PSOs. Three of the most
interesting ones are:

PSOG1 was evolved when the training set was the shifted city-block sphere
functions of dimension N = 2, and so it was expected to perform well on uni-
modal objective functions:

F = (xsi − xi ) − (vi Ṙ)

This is particularly interesting since it includes a deterministic, 100% social

component as well as a random friction component.

PSOG2 was also evolved when the training set included shifted city-block
sphere functions of dimension N = 2. Its equation is equivalent to

F = 0.5 ((xsi − xi ) + (xpi − xi ) − vi )

This is interesting because it is completely deterministic (particles are attracted

towards the middle between swarm best and particle best) and because it in-
cludes standard friction.
 Extending Particle Swarm Optimisation via GP 7

PSOG3 was evolved when the training set included shifted Rastrigin functions
of dimension N = 2, and so it was expected to perform well on highly multimodal
objective functions. Its equation is equivalent to:

F = R1 (xsi − xi ) − 0.75R2 R1 xi x2si − 0.25R3 R2 R1 xi xsi

This is interesting for a number of reasons. Firstly, it does not use information
about each particle’s best. Probably this is because, in a highly multimodal
landscape, particles should not trust their own observations too much. Their
personal best is likely to belong to the basin of attraction of a deceptive local
optimum. The swarm best, instead, has a higher chance of being in the basin of
attraction of the global optimum, and so particles should aim at exploring its
surroundings. PSOG3’s second term is also interesting. Unless the swarm best
is near the origin, this component will tend to push the particles towards the
origin. The push is very mild if the swarm best is not too far from the origin,
but it becomes quite strong otherwise. Clearly, the reason why this component
is useful is that GP has found a regularity in the training set: global optima tend
to be near the origin, and so that is an area of the search space that should be
explored preferentially.
In order to compare the behaviour of the hand-designed and evolved PSOs,
we tested them on 30 random problems taken from the city-block sphere and
Rastrigin function classes for two and ten dimensions. The problems were gener-
ated by selecting the components of the global optimum ci uniformly at random
from the interval [−C, +C], with C = 1.0 and C = 2.0. Note that PSOG1 and
PSOG2 were evolved using the city-block sphere problem class, N = 2 and
C = 1.0, and that PSOG3 was evolved using the Rastrigin function problem
class, N = 2 and C = 1.0. Thus, all other conditions represent off-sample test
problems and are useful to assess the generalisation capabilities of these PSOs.
For each problem instance we performed 30 independent runs of each PSO.
The results are reported in Tables 1 and 2. The tables show the average over
the 30 problems and the standard deviation (in brackets) of the mean (over
30 independent runs) absolute error between the coordinates
P of the swarm best
and the coordinates of the global optimum (i.e., i |xs − ci |/N ) at the end of
30 PSO iterations. This gives an idea of how far the swarm best was from the
global optimum in each dimension. Data in boldface represent the PSO with the
best average performance in each condition.
On the two-dimensional city block sphere problem class, all PSOs do quite
well, with the exception of PSOR0 which cannot really be expected to do very
well on unimodal functions. PSOG1 is better than the standard PSO. However,
to our surprise also PSOG3 (which wasn’t evolved on sphere functions) does
better than the standard PSO (and in fact is even better than PSOG1 for ten
dimensions), suggesting PSOG3 may be a good all-rounder. On the Rastrigin
function problem class, PSOG3 outperforms all other PSOs by a considerable
margin. In ten dimensions PSOG2 is second best, while in two dimensions
PSOR1 is second best. In other tests (not reported) where each PSO was run
for 300 iterations instead of 30, we obtained essentially the same results.
8 R. Poli, W.B. Langdon and O. Holland

Table 1. City Block Sphere. Normalised mean (standard deviation) of the distance
between best location found by each PSO and global optima. Best results in bold.

N C PSO PSOD1 PSOR0 PSOR1 PSOG1 PSOG2 PSOG3

2 1 .046 (.089) .002 (.0002) .26 (.029) .003 (.0003) .001 (.0016) .048 (.014) .01 (.007)
2 2 .054 (.093) .002 (.0002) .27 (.036) .003 (.0004) .001 (.0019) .066 (.029) .04 (.028)
10 1 .52 (.47) .31 (.025) 1.6 (.021) .27 (.022) .45 (.025) .65 (.023) .17 (.037)
10 2 .62 (.45) .38 (.028) 1.6 (.036) .31 (.027) .55 (.045) .8 (.056) .45 (.096)

Table 2. Rastrigin. Normalised mean (standard deviation) of the distance between

best location by each PSO and global optima. Best results in bold.

N C PSO PSOD1 PSOR0 PSOR1 PSOG1 PSOG2 PSOG3

2 1 .66 (.22) .81 (.081) .77 (.072) .64 (.072) .94 (.084) .72 (.07) .28 (.066)
2 2 .71 (.2) .85 (.098) .79 (.066) .66 (.091) .94 (.12) .75 (.12) .47 (.16)
10 1 1.2 (.35) 1.3 (.05) 1.9 (.042) 1.3 (.07) 1.3 (.055) 1.1 (.054) .59 (.057)
10 2 1.4 (.28) 1.4 (.073) 1.9 (.049) 1.4 (.064) 1.4 (.079) 1.3 (.063) .94 (.11)

It is interesting to compare the behaviour of PSOG3 and PSO. Figure 1

illustrates the behaviour of PSO on a city block sphere function with N = 2. The
particles tend to rapidly focus towards the global optimum at the origin. Figure 2
shows the behaviour of PSOG3 in exactly the same conditions (including same
starting positions for the particles). Here the particles tend to focus less rapidly.
Figures 3 and 4 illustrate the behaviour of PSO on a Rastrigin function
with N = 2. Here the swarm controlled by PSO is quickly attracted towards a
deceptive local optimum, while the particles in PSOG3 perform bigger orbits
and eventually discover and start converging towards the global optimum at the
origin.

5 Conclusions

GP has been able to evolve a variety of particle swarm optimisers that work
as well or considerably better than standard human-designed PSOs. Analysis of
the evolved programs has led to new insights in the design of PSOs tailored for
specific classes of landscapes.
To evolve our PSOs we have used the state of the art in GP, but we have
not proposed a great deal in terms of new GP technology. However, in the more
general context of machine intelligence, this work represents an important step
within a new research trend: using search algorithms to discover new search al-
gorithms. This approach has become possible thanks to the growth in computing
power. We can already foresee that the results of this may be spectacular (see,
for example, the case of Fukunaga’s award winning work on evolving human-
competitive SAT problem solvers [6]). The main contribution of this paper is to
show that genetic programming can evolve better than human-designed PSOs
in a few hours on a standard PC. A second contribution is to give us new ideas
 Extending Particle Swarm Optimisation via GP 9

All PSO particles All GP particles

4 4

2 2

0 0

-2 -2

-4 -4

-4 -2 0 2 4 -4 -2 0 2 4

Fig. 1. Trajectories of the particles in one Fig. 2. A run of PSOG3 on the 2–D city-
prototypical run of PSO on the 2–D city- block sphere problem (same initial condi-
block sphere problem. tions as in Fig.1).

All PSO particles All GP particles

4 4

2 2

0 0

-2 -2

-4 -4

-4 -2 0 2 4 -4 -2 0 2 4

Fig. 3. A run of PSO on the 2–D Rastrigin Fig. 4. A run of PSOG3 on the 2–D Ras-
function problem (global optimum at the trigin function problem.
intersection of cross hairs).

on what types of particle behaviours are most appropriate for which type of
landscape.
In future research we intend to apply the approach to a variety of problem
domains (including real-world problems) and to extend it by allowing GP to use
more information on the past history of the swarm to control the particles and
by allowing the evolution of coupled force-generating equations. We also intend
to explore the effects and benefits of using different performance measures for
PSO evolution.

Acknowledgments

The authors would like to thank EPSRC (grant GR/T11234/01) for financial
support.
10 R. Poli, W.B. Langdon and O. Holland

References
1. P. J. Angeline. Using selection to improve particle swarm optimization. In IEEE
World Congress on computational intelligence, ICEC-98, pages 84–89, Anchorange,
Alaska, 1998.
2. T. M. Blackwell and P. J. Bentley. Dynamic search with charged swarms. In
GECCO 2002: Proceedings of the Genetic and Evolutionary Computation Confer-
ence, pages 19–26, New York, 9-13 July 2002. Morgan Kaufmann Publishers.
3. T. M. Blackwell and J. Branke. Multi-swarm optimization in dynamic environ-
ments. In Applications of Evolutionary Computing. Springer, 2004.
4. R. Brits, A. P. Engelbrecht, and B. Bergh. A Niching Particle Swarm Optimizer. In
Proceedings of the 4th Asia-Pacific Conference on Simulated Evolution and Learn-
ing (SEAL’02), volume 2, pages 692–696, Orchid Country Club, Singapore, Nov.
2002. Nanyang Technical University.
5. M. Clerc and J. Kennedy. The particle swarm-explosion, stability, and conver-
gence in a multidimensional complex space. IEEE Transactions on Evolutionary
Computation, 6(1):58–73, 2002.
6. A. S. Fukunaga. Evolving local search heuristics for SAT using genetic program-
ming. In Genetic and Evolutionary Computation – GECCO-2004, Part II, volume
3103 of Lecture Notes in Computer Science, pages 483–494, Seattle, WA, USA,
26-30 June 2004. Springer-Verlag.
7. F. Heppner and U. Grenander. A stochastic nonlinear model for coordinated bird
flocks. In The ubiquity of Chaos. AAAS publications, Washington DC, 1990.
8. J. Kennedy. The behavior of particles. In Evolutionary Programming VII: Pro-
ceedings of the Seventh Annual Conference on evolutionary programming, pages
581–589, San Diego, CA, 1998.
9. J. Kennedy and R. C. Eberhart. Swarm Intelligence. Morgan Kaufmann Publish-
ers, San Francisco, California, 2001.
10. J. R. Koza. Genetic Programming: On the Programming of Computers by Means
of Natural Selection. MIT Press, Cambridge, MA, USA, 1992.
11. T. Krink, J. S. Vesterstrøm, and R. Riget. Particle swarm optimisation with
spatial particle extension. In Proceedings of the 2002 Congress on Evolutionary
Computation CEC2002, pages 1474–1479. IEEE Press, 2002.
12. W. B. Langdon and R. Poli. Foundations of Genetic Programming. Springer-
Verlag, 2002.
13. M. Lovbjerg and T. Krink. Extending particle swarm opimisers with self-organized
criticality, July 11 2002.
14. E. Ozcan and C. K. Mohan. Particle swarm optimization: surfing the waves. In
Proceedings of the IEEE Congress on evolutionary computation (CEC 1999), Wash-
ington DC, 1999.
15. R. Poli and C. R. Stephens. Constrained molecular dynamics as a search and
optimization tool. In Genetic Programming 7th European Conference, EuroGP
2004, Proceedings, volume 3003 of LNCS, pages 150–161, Coimbra, Portugal, 5-7
Apr. 2004. Springer-Verlag.
16. J. Riget, J. S. Vesterstrm, and K. Krink. Division of labor in particle swarm
opimisation, July 11 2002.
17. Y. Shi and R. C. Eberhart. A modified particle swarm optimizer. In Proceedings
of the IEEE Congress on Evolutionary Computation (CEC 1999), pages 69–73,
Piscataway NJ, 1999.
18. D. H. Wolpert and W. G. Macready. No free lunch theorems for optimization.
IEEE Transactions on Evolutionary Computation, 1(1):67–82, Apr. 1997.