Information Sciences 161 (2004) 21–35

www.elsevier.com/locate/ins

A fuzzy controller with evolving structure

Plamen Angelov *

Loughborough University, Loughborough, Leicestershire, LE11 3TU, UK

Received 10 March 2002; accepted 18 March 2003

Abstract
An approach to on-line design of fuzzy controllers of Takagi–Sugeno type with
gradually evolving structure is treated in the paper. Fuzzy rules, representing the
structure of the controller are generated based on data collected during the process of
control using newly introduced technique for on-line identification of Takagi–Sugeno
systems. The output of the plant under control (including its dynamic) and the
respective control signal has been memorised and stored in on-line mode. These data has
been used to train in a non-iterative, recursive way the fuzzy controller. The indirect
adaptive control approach has been used in combination with the novel on-line iden-
tification technique. This approach exploits the quasi-linear nature of Takagi-Sugeno
models and builds-up the control rule-base structure and adapts it in on-line mode with
recursive, non-iterative learning. The method is illustrated with an example from air-
conditioning systems, though it has wider potential applications.

2003 Elsevier Inc. All rights reserved.

Keywords: Evolving fuzzy rule-based systems; On-line identification; On-line clustering;

Recursive least squares estimation; Indirect adaptive control

1. Introduction

Fuzzy logic controllers of the so-called Takagi–Sugeno (TS) type are gaining
impetus recently in applications to complex systems, because of their compu-
tational efficiency, transparency and flexibility [13,14]. The main reason for this

*
Address: Department of Communication Systems, Lancaster University, Bailrigg, Lancaster,
A1 4YR, UK. Tel.: +44-1524-594188; fax: +44-1524-592713.
E-mail addresses: p.p.angelov@lboro.ac.uk, p.angelov@lancaster.ac.uk (P. Angelov).

0020-0255/$ - see front matter
2003 Elsevier Inc. All rights reserved.
doi:10.1016/j.ins.2003.03.006
22 P. Angelov / Information Sciences 161 (2004) 21–35

is their dual, quasi-linear nature. Generally, they are non-linear and therefore
suitable for complex systems, but they could also be treated as linear in respect
to the consequent parameters, which makes possible the application of efficient
recursive techniques. TS models serve as a bridge between the classical control
theory [8] and the soft computing [12] generalizing the well-established con-
cepts of gain scheduling and local linear modeling and control [4].
The problem of fuzzy rule-based systems design is already treated in the
literature for the off-line case when a batch set of data is used [7,9–13,17,18]. In
reality, however, even if a validation test has been made beforehand, there is no
guarantee that a controller designed in this way will perform satisfactory if the
object of control or its environment change [2]. The reasons could be aging,
wearing, change in the mode of operation or development of a fault, seasonal
changes in the environment, etc. [3]. An effective mechanism for tackling such
problems known from the classical control theory is the adaptation [8]. It is
well developed for the linear models and controllers, but not for a general
(normally non-linear, complex and uncertain) case [2]. It usually supposes a
fixed model structure and applies to parameters only [8].
Exploiting the specific dual, quasi-linear, semi-fuzzy, semi-crisp nature of the
TS rule-based fuzzy systems we have developed and tested a new approach for
on-line adaptation of their structure and parameters [2–6]. It is based on a
combination of unsupervised on-line clustering and a version of recursive
weighted least squared estimation of the parameters of the consequent part [4,5].
In the present paper, this approach is applied to on-line design of fuzzy
controllers combined with the indirect learning technique, proposed in [15] and
developed in [1]. Instead of a neural network trained off-line as in [1,15] we use
evolving rule-based models [6] to design the controller structure on-line by
adding more informative rules than those, that previously formed the model. In
a driven evolution of the controller structure the existing rules are replaced
with new ones based on the informative potential of the data samples, mea-
sured by their spatial proximity to the potential rules centers. In this way, when
new informative data become available, in addition to the parameters update
the rule-base also gradually evolves inheriting and updating the existing one.
The evolving nature of these controllers in combination with the high degree
of their transparency, compact form, and computational efficiency makes them
a promising candidate for on-line control. An illustrative example of indoor
temperature control is presented.

2. Fuzzy controller with evolving structure

The proposed control scheme combines two basic paradigms:

• indirect learning (IL) mechanism introduced initially in [15];

• evolving fuzzy rules [2–6].
 P. Angelov / Information Sciences 161 (2004) 21–35 23

2.1. Indirect learning control

The first one is based on the approximation of the inverse dynamics of

the plant. The IL-based control scheme supposes model-free concept and
instead of feeding back the error between the plant output and the reference
signal it feeds back the integrated (or delayed one-step back) output signal
(Fig. 1).
Fig. 1 represents the basic IL-based control scheme. The switch has two
positions: ÔC’ for control and ÔL’ for learning. During the learning phase, the
actual output signal ðykþ1 Þ at the time-step (k þ 1) is fed back and the switch is
in position L. The other signal supplied to the input of the controller comes
from the Delay block, which performs an integration of the output signal. It
can be realised in discrete form as memorisation of the past output signals ðyk Þ.
Additionally, we consider a measured disturbance signal ðFk Þ. The output of
the controller is the control signal ðuk Þ. During the control phase (the switch is
at the position C) the reference signal ðrÞ is supplied as an input instead of the
predicted next step output ðykþ1 Þ. In this way, the controller already trained in
the previous learning phases produces such a control signal ðuk Þ, which brings
the output of the plant at the next time-step ðykþ1 Þ close to the reference signal
ðrÞ.
In the original works on IL-based control [1,15] the controller has been
realised as a neural network trained off-line using the sets of data (yk , ykþ1 ,
uk ; k ¼ 1; 2; . . . ; TD; where TD is the number of training data samples). As it
is known, however, the learning techniques for NN are iterative and,
therefore training of the neuro-controller in [1] and [15] has been done off-
line. Additional important disadvantage of neural networks in comparison to
the fuzzy rule-based models is their lack of interpretability and transparency
[11,12].

Fk

C yk+1
uk
r Controller Plant

yk
Delay

yk+1

Fig. 1. IL-based control scheme.

24 P. Angelov / Information Sciences 161 (2004) 21–35

2.2. Evolving fuzzy rules

Owing to the facts that IL control scheme is inherently convenient for

recursive, on-line implementation, and the fuzzy rule-based models are proven
universal approximators [16] we apply evolving TS models [4–6] to represent
the controller. Additionally, we also consider a measured disturbance signal,
which for the case of indoor temperature control is representing the ambient
temperature. At each time instance, it performs consecutive time-steps of
learning and control. In this way, the controller learns and evolves its structure
(fuzzy rules and fuzzy sets) and its parameters (focal points of the rules and
local model parameters) on-line in the process of control itself. The rules are
supposed of TS form with three inputs and a linear consequence:

Ri : IFðyk
1 is Yi Þ AND ðr is Ri Þ AND ðFk
1 is Ui Þ

ð1Þ
THEN ðuki ¼ ai0 þ ai1 yk
1 þ ai2 r þ ai3 Fk
1 Þ

where Ri denotes the ith fuzzy rule; i 2 f1; Rg; R is the number of fuzzy rules;
Yi ; Ri , and Ui denote fuzzy sets of the antecedent part of the controller; uki
denotes output (control signal) of the ith local controller at time instance
k ¼ 2; 3; . . .; ail are its parameters, l 2 f0; 3g.
For Gaussian-like antecedent fuzzy sets the level of contribution of the
corresponding linear model to the overall output is determined as:

lij ¼ e
akxij
xi k ; i ¼ 1; 2; . . . ; R; j ¼ f1; 2; 3g ð2Þ

where a is a positive constant, which defines the zone of influence of the ith
T
model; x ¼ ½yk ; r; Fk  is the inputs vector; xi is the focal point of the ith rule
antecedent.
The firing level of the rules are defined as Cartesian product or conjunction
of respective fuzzy sets for this rule:

\
3
si ¼ li1 ðyk Þ  li2 ðrÞ  li3 ðFk Þ ¼ lij ðxj Þ i ¼ 1; 2; . . . ; R ð3Þ
j¼1

The controller output is calculated by weighted averaging of individual

rules’ contributions:

X
R X
R
si
uk ¼ ki uki ¼ ki xTe pi ; ki ¼ PR ð4Þ
i¼1 i¼1 j¼1 sj

where ki is the normalized firing level of the ith rule; uki represents the output of
the ith linear model; pi ¼ ½ai0 ai1 ai2 ai3 T ; i ¼ 1; 2; . . . ; R, is the vector of para-
meters of the ith linear model; xe ¼ ½1xT T is the expanded data vector.
 P. Angelov / Information Sciences 161 (2004) 21–35 25

It should be noted that because of the Ôblending’ of the partial outputs,

which are locally linear (consequences in (1)) the general output determined by
the TS system (4) is non-linear.

3. Fuzzy controller on-line training

The data are collected continuously, in on-line mode during the process of
control. Some of the new data reinforce and confirm the information contained
in the previous data. Other data, however, bring new information, which could
indicate a change in operating conditions, development of a fault or simply a
more significant change in the dynamic of the process [2–6]. They may posses
enough new information to form a new rule or to modify an existing one. The
value of the information they bring is closely related to the information the
data collected so far already possesses. The judgement of the informative po-
tential and importance of the data is made based on their spatial proximity,
which corresponds to operating conditions, possibly seasonal variations or
different faults.
On-line training of the controller includes on-line clustering under assumption
of a gradual change of the rule-base and modified recursive least squares [4–6].

3.1. On-line clustering

The on-line clustering procedure starts with the first data point established as
the focal point of the first cluster. Its co-ordinates are used to form the ante-
cedent part of the fuzzy rule (1) using for example Gaussian membership
functions (2). Any other type of membership functions could also be used in-
stead. Its potential is assumed equal to 1.
Starting from the next data point onwards the potential of the new data
points is calculated recursively. As a measure of the potential, we use a Cauchy
type function of first order [4–6]:
1
Pk ðzk Þ ¼ 1
Pk
1 P4 j 2
; k ¼ 2; 3; . . . ð5Þ
1þ ðk
1Þ i¼1 j¼1 ðdlk Þ

where Pk ðzk Þ denotes the potential of the data point zk ¼ ½xk ; uk T ¼

½yk ; r; Fk ; uk T ; dlkj ¼ zjl
zjk , denotes projection of the distance between two data
points on the axis y; r; F or u, respectively.
This function is monotonic and inversely proportional to the distance and
enables recursive calculation [5]
k
1
Pk ðzk Þ ¼ ; k ¼ 2; 3; . . . ð6Þ
ðk
1Þð#k þ 1Þ þ rk
2mk
P P P4 P4 j j j Pk
1 j
where #k ¼ 4j¼1 ðzjk Þ2 ; rk ¼ k
1
l¼1
j 2
j¼1 ðzl Þ ; mk ¼ j¼1 zk bk ; bk ¼ l¼1 zl .
26 P. Angelov / Information Sciences 161 (2004) 21–35

Parameters #k and mk in are calculated from the current data point zk , while
bjk and rk are recursively updated as follows:
X
4
rk ¼ rk
1 þ ðzjk
1 Þ2 ; bjk ¼ bjk
1 þ zjk
1 :
j¼1

The recursive formula (6) is especially important for on-line implementation

of the training algorithm. After the new data are available in on-line mode, they
influence the potentials of the centres of the clusters ðzl ; l ¼ ½1; RÞ, which are
respective to the focal points of the existing rules ðxl ; l ¼ ½1; RÞ. The reason is
that by definition the potential depends on the distance to all data points,
including the new ones (the sum in the denominator by l in (5) has an
increasing number of components). The recursive formula for update of the
potentials of the existing clusters’ focal points can easily be derived from (5):
ðk
1ÞPk
1 ðzl Þ
Pk ðzl Þ ¼ P ; k ¼ 2; 3; . . . ð7Þ
k
2 þ Pk
1 ðzl Þ þ Pk
1 ðzl Þ 4j¼1 ðdkðk
1Þ
j
Þ2

where Pk ðzl Þ is the potential of the data point zl at time k which is a prototype
of the lth rule.
Potentials of the new data points are compared to the updated potential of
the centres of the existing clusters.
If the potential of the new data point is higher than the potential of the
existing centres then the new data point it is accepted as a new centre and a new
rule is formed with a focal point based on the projection of this centre on the
axis x ðR :¼ R þ 1; xR ¼ xk Þ. The rationale is that in this case the new data point
is more descriptive, has more summarisation power than all the other data
points. It should be noted that the condition to have higher potential is a very
strong one. The reason is that with the growing number of data, their con-
centration is usually decreasing except in the cases some new important region
of data space reflecting a new operating regime or new condition appears. In
such cases a new rule is formed, while outlying data are automatically rejected
because their potential is significantly lower due to their distance from the
other data.
If in addition to the previous condition (the potential of the new data point is
higher that the potential of all the previously existing centres) the new data
point is close to an old centre:

arg minRi¼1 kzk
zl k Pk ðzk Þ

þ <1 ð8Þ
r maxRl¼1 Pk ðzi Þ

then the new data point ðzk Þ replaces this centre ðzj :¼ zk Þ. This mechanism for
rule-base adaptation called modification ensures a replacement of a rule with
another one built around the projection of the new data point on the axis x.
 P. Angelov / Information Sciences 161 (2004) 21–35 27

This on-line clustering approach [4,5] ensures an evolving fuzzy rule-based

system by dynamically upgrading and modifying it while inheriting the bulk of
the rules (R
1 of the rules are preserved even when a modification or an
upgrade take place).

3.2. On-line recursive estimation of consequence parameters

TS models (1) underlying the structure of the controller are quasi-linear

(having linear consequent part), which tempts to apply well-known recursive
least squares estimation (Kalman filter) [8]. Because of the gradual evolution of
the controller structure, however, the normalized firing strengths of the rules, ki
(4) will change. Since this effects all the data (including the data collected before
time of the change) the straightforward application of the Kalman filter is not
correct. A proper resetting of the co-variance matrices and parameters ini-
tialisation is needed at each time a rule is added to and/or removed from the
rule base [5]. In [5] we estimate the co-variance matrices and parameters of the
new ðR þ 1Þth rule as a weighted average of the respective co-variance and
parameters of the remaining R rules. This is possible, since the approach of
rule-base innovation, we consider concerns one rule only the other R rules
remain unchanged.
The estimation of parameters of local models is based on the wRLS [5]:

p ^ik
1 þ cik xek
1 ki ðxk
1 Þðyk
xTek
1 p

^ik ¼ p ^ik
1 Þ; k ¼ 2; 3; . . . ð9Þ

ki ðxk
1 Þcik
1 xek
1 xTek
1 cik
1

cik ¼ cik
1
; i ¼ ½1; R ð10Þ
1 þ ki ðxk
1 ÞxTek
1 cik
1 xek
1

^1 ¼ 0 and ci1 ¼ XI.

with initial conditions p
In this case, the co-variance matrices are separate for each rule and have
smaller dimensions ðcik 2 Rðnþ1Þðnþ1Þ ; i ¼ 1½1; RÞ. Parameters of the newly
added rule are determined as weighted average of the parameters of the rest R
rules. Parameters of the other R rules are inherited ðpik :¼ piðk
1Þ ; i ¼ ½1; RÞ
from the previous time-step.
When a rule is replaced by another rule, which have close antecedent
parameter (centre) then parameters of all rules are inherited ðpik :¼ piðk
1Þ ;
i ¼ ½1; RÞ.
The co-variance matrix of the newly added rule is initialised by

cRþ1k ¼ XI ð11Þ

The co-variance matrices of the rest R rules are then inherited ðc:ik ¼ ciðk
1Þ ;
i ¼ ½1; RÞ.
28 P. Angelov / Information Sciences 161 (2004) 21–35

3.3. Procedure for fuzzy controller structure evolution

The recursive procedure for on-line training the fuzzy controller includes the
following stages:

Stage 1. Initialisation of the controller structure (antecedent part of the fuzzy

rules).
Stage 2. At the next time-step reading the next data sample zk .
Stage 3. Recursive calculation of the potential of each new data sample to
influence the structure of the controller.
Stage 4. Recursive up-date of the potentials of old centres taking into account
the influence of the new data sample.
Stage 5. Possible modification or up-grade of the controller structure based on
the potential of the new data sample in comparison to the potential of
the existing rules’ centres.
Stage 6. Recursive calculation of the consequent parameters.
Stage 7. Generation of the control signal for the next time step ðukþ1 Þ.

The execution of the algorithm continues for the next time-step from stage 2.
Stage 1. The controller rule-base could contain one single rule only, based,
for example, on the first data sample. Then
k :¼ 1; R :¼ 1; x1 :¼ xk ; P1 ðz1 Þ :¼ 1; h1 ¼ p1 ¼ 0; C1 ¼ XI
ð12Þ
where z1 is the first cluster centre; x1 par is focal point of the first rule being a
projection of z1 on the axis x.
In principle, the rule-base could be initialised by existing expert knowledge
or off-line trained controller. In this case:

R :¼ Rini ; P1 ðzi Þ :¼ 1; i ¼ ½1; Rini  ð13Þ

where Rini denotes the number of rules defined initially off-line.

Stages 2–7 are performed on-line. They form the distinctive characteristics of
the proposed approach.
Stage 2. At the next time-step ðk :¼ k þ 1Þ the new data sample ðzk Þ is col-
lected.
At stage 3 the potential of each new data sample is recursively calculated by
(6). The use of already calculated values #k and bjk leads to significant time and
calculation savings the number of training data in on-line mode is continuously
growing. At the same time, they have accumulated information regarding the
spatial proximity of all previous data.
At stage 4 the potentials of the focal points (centers) of the existing clusters/
rules are recursively updated by (7).
 P. Angelov / Information Sciences 161 (2004) 21–35 29

At stage 5 the potential of the new data sample is compared to the updated
potential of existing centres and a decision whether to modify or up-grade the
controller rule-base is taken.

(a) If (the potential of the new data point is higher than the potential of the
existing centres: Pk ðzk Þ > Pk ðzi Þ; i ¼ ½1; RÞ
AND (the new data point is close to an old centre (8))
THEN the new data point ðzk Þ replaces it.

In this case, the new data point is used as a prototype of a focal point. Let us
suppose that it has index j
R
zj ¼ arg minkzk
zi k : xj :¼ xk ; Pk ðzj Þ :¼ Pk ðzk Þ ð14Þ
i¼1

Consequence parameters and co-variance matrices are inherited from the

rule to be replaced:

p ^j
^k :¼ p k ; Ck :¼ Ckj ð15Þ

It should be noted that when a rule is replaced by another rule the weights
ðkÞ are changing according to (4) and the summation in the denominator in (4)
should change. R
1 addends in this summation are the same and only one
change. Moreover, since the new centre is close to the replaced one by defini-
tion (8), this change is marginal. The disturbance caused by this change could
be ignored, because the Kalman filter is able to cope with using the existing
estimations of the parameters and co-variance matrices. This is illustrated by
the experimental results (next section).

(b) ELSE IF (the potential of the new data point is higher than the potential of
the existing centres: Pk ðzk Þ > Pk ðzi Þ; i ¼ ½1; RÞ
THEN it is added to the rule-base as a new rule’s centre.

In this case, the new data point becomes a prototype of a focal point of a
new rule
R :¼ R þ 1; xR ¼ xk ; Pk ðzR Þ ¼ Pk ðzk Þ ð16Þ

Consequence parameters and co-variance matrices are reset by (11).

END IF
At stage 6 parameters of the consequence are recursively updated by the
Kalman filter.
At stage 7 the control signal ðukþ1 Þ for the next time-step is generated.
The algorithm continues from stage 2 by reading the next data sample at the
next time-step. All steps are non-iterative.
30 P. Angelov / Information Sciences 161 (2004) 21–35

Using this approach, a transparent, compact, and efficient controller

structure can be found by on-line rule base evolution with the simultaneous
recursive estimation of the fuzzy set parameters. The proposed controller uses
non-iterative training, which is very appropriate for real-time applications
(calculations for each time-step take fraction of a second on a 550 MHz PC in
Matlab). Most of the reported approaches for rule-base evolution by genetic
algorithms [9–11] are applicable off-line only. The others require computa-
tionally expensive iterative procedures sing gradient-based back-propagation
or genetic algorithms [13,14]. Other approaches, like the one considered in
[17,18], concern Mamdani type fuzzy systems (with no functional consequence)
and perform on-line tuning of a defuzzification parameter and correction of the
rules’ output. The rule-structure itself, however, is not changed, but a variable
part of it (FSAR: fuzzy set of active rules) is considered. In our approach the
whole structure of the fuzzy controller has evolved and been upgraded.
Additionally, it is of TS type, which has been proven to be significantly more
effective [12].
It is interesting to note that the rate of upgrade with new rules does not lead
to an excessively large rule base in comparison to the other methods [9–
11,13,14,17]. The reason for this is that the condition for the new data point to
have higher potential than the focal points of rules of all existing rules is a hard
requirement. Additionally, the possible proximity of a candidate centre to the
already existing focal points leads to just a replacement of the existing focal
point, i.e. modification of its co-ordinates without enlarging the rule-base size.

4. Indoor temperature controller with evolving structure

A subsystem of an indoor climate control system serving a real building,

located in Iowa, USA is considered (Fig. 2). Data generated from the ASH-
RAE funded research project RP1020 (courtesy of ASHRAE) has been used.
The ambient temperature on 22 August 1998 (dotted line in Fig. 3) has been
considered as a measured disturbance. The cooling coil (Fig. 2) cools the warm
air that flows on to the coil. The cool air is used to maintain comfortable
conditions in an occupied space. One of the principle loads on the coil is
generated due to the supply of ambient air required to maintain a minimum
standard of indoor air quality.
The temperature of the air which is supplied to the occupied zone and is
outlet from the cooling coil ðT out Þ is the output variable of the controlled plant
(y). The reference temperature ðT ref Þ determines the reference signal (r). The
control signal (u) commands an electrically driven actuator, which operates the
control valve diverting the water flow from one port to the other.
The measurements of air inlet (in fact, the ambient temperature, T amb Þ and
outlet temperatures ðT out Þ are used as inputs to the controller. Volumetric flow
 P. Angelov / Information Sciences 161 (2004) 21–35 31

Cooling coil Axial Fan

Air Flow
•T out supplied
Air Flow •T amb to the
Direction occupied
Zone

T ref
Cooling
water
intake

Actuator
Controller
Valve

Fig. 2. Supply air temperature control sub-system.

Fig. 3. Control of the indoor temperature by eTS controller.

rate of air ðma ¼ 1:0 kg/s), moisture content ðg ¼ 0:008 kg/kg), and the tem-
perature of the water ðTw ¼ 10, C) are supposed to be constant. The sample
32 P. Angelov / Information Sciences 161 (2004) 21–35

interval is 1 min. Results of simulation with the evolving fuzzy controller are
depicted in Fig. 3.
In this simulation study we consider the reference signal ðT ref Þ to be fixed at
19 C for the first 100 min, at 17 C for the next 100 min, and at 20 C for the
rest 100 min (dashed line in Fig. 3). A set of data (100 data points) has been
used to generate off-line the initial structure of the controller.
It consists of three fuzzy rules:
out
R1 : IFðTk
1 is HÞ AND ðT ref is HÞ AND ðTk
1
amb
is LÞ
THEN ðuk1 ¼ a10 þ a11 yk
1 þ a12 r þ a13 Fk
1 Þ
out
R2 : IF ðTk
1 is MÞ AND ðT ref is MÞ AND ðTk
1
amb
is HÞ
ð17Þ
THEN ðuk2 ¼ a20 þ a21 yk
1 þ a22 r þ a23 Fk
1 Þ
out
R3 : IF ðTk
1 is HÞ AND ðT ref is HÞ AND ðTk
1 amb
is VLÞ
THEN ðuk3 ¼ a30 þ a31 yk
1 þ a32 r þ a33 Fk
1 Þ
where the linguistic labels (H for High, L for Low, VL for Very Low, M for
Medium) are assigned to membership functions representing the respective
fuzzy rule.
In the on-line phase, the incremental learning and evolving control are
sequentially performed at each time instant (in respect to each new data point)
as described in the previous section. The new triplet of data ðTkout ; Tkamb ; uk Þ has
been recorded in real time and the value of Tkout has been memorised. In this
out
way, at the next time instance we have Datak ¼ ðTk
1 ; Tkout ; Tk
1
amb
; uk
1 Þ. This
quadruple forms a data point for training the eTS controller. Then a recursive
procedure for up-date of the controller structure and parameters [4,5] takes
place. Generally, based on the informative potential (spatial proximity measure
to the focal point of a rule) of a new data point Datak a decision to upgrade or
modify the controller rule-base is taken [4–6]. This process is repeated for every
time instant (from 1 to 300 in this example) and calculations take several
seconds making possible real-time application. For most of the new data
points, no changes are required because the informative potential of the new
data sets is not high enough to replace, modify or add a new rule and to correct
the parameters. Two of the new data points, however, have high enough po-
tential and they have triggered the modification and upgrade of the rule-base.
At time instant k ¼ 3 (3 min after the on-line control starts) the data point is
informative enough to generate a new fuzzy rule, which has been added to the
controller rule-base:
out
R4 : IF ðTk
1 is LÞ AND ðT ref is LÞ AND ðTk
1
amb
is MÞ
ð18Þ
THEN ðuk4 ¼ a40 þ a41 yk
1 þ a42 r þ a43 Fk
1 Þ
There was also one data point, which has informative potential high enough
to modify the rule-base though not high enough to be added as a new rule.
 P. Angelov / Information Sciences 161 (2004) 21–35 33

Fig. 4. Membership functions and linguistic labels of the fuzzy variables T amb and T out .

Additionally, this data point was close to the centre of the rule R3 . Based on
this, the fuzzy rule R3 has been replaced by the more informative rule R5 :
out
R5 : IF ðTk
1 is MÞ AND ðT ref is RHÞ AND ðTk
1 amb
is MÞ
ð19Þ
THEN ðuk5 ¼ a50 þ a51 yk
1 þ a52 r þ a53 Fk
1 Þ
Fuzzy sets with respective linguistic labels are depicted in the Fig. 4 for the
variables T amb and T out .

5. Conclusions

A new approach for on-line design of fuzzy controllers of TS type with

gradually evolving structure is treated in the paper. Fuzzy rules, representing
the structure of the controller are generated based on data collected during the
process of control using newly introduced technique for on-line identification of
TS type of fuzzy models. The output of the plant under control (including its
dynamic) and the respective control signal applied has been memorised and
stored in on-line mode. These data has been used to train in a non-iterative,
recursive way the fuzzy controller. The indirect adaptive control approach has
been used in combination with the novel on-line identification technique. This
approach exploits the quasi-linear nature of TS models and builds-up the
controller rule-base structure and adapts it in on-line mode with recursive, non-
iterative learning. The method is illustrated with an example from air-condi-
tioning systems, though it has wider potential applications. The problems of
stability of the controller require further investigation. They will be addressed
in a future work using the latest results on stability of TS models with a fixed
structure.
34 P. Angelov / Information Sciences 161 (2004) 21–35

Acknowledgements

The author acknowledges use of the data courtesy of ASHRAE; generated

from the ASHRAE funded research project RP1020.

References

[1] H.C. Andersen, F.C. Teng, A.C. Tsoi, Single net indirect learning architecture, IEEE
Transactions on Neural Networks 5 (1994) 1003–1005.
[2] P.P. Angelov, Evolving Rule-based Models: A Tool for Design of Flexible Adaptive Systems,
Springer, Physica-Verlag, Heidelberg, Germany, 2002.
[3] P.P. Angelov, R.A. Buswell, Identification of evolving fuzzy rule-based models, IEEE
Transactions on Fuzzy Systems 10 (2002) 667–677.
[4] P.P. Angelov, D.P. Filev, An approach to on-line identification of Takagi–Sugeno fuzzy
models, IEEE Transactions on System, Man and Cybernetics, Part B, vol. 33, December, 2003,
to appear.
[5] P.P. Angelov, D.P. Filev, Flexible models with evolving structure, in: Proceedings of the IEEE
Symposium on Intelligent Systems IS’02, Varna Bulgaria, 10–12 September, vol. 2, 2002,
pp. 28–33.
[6] P.P. Angelov, R.A. Buswell, Evolving rule-based models––a tool for intelligent adaptation, in:
Proceedings of the Joint 9th IFSA World Congress and 20th NAFIPS Annual Conference,
Vancouver, BC, Canada, 25–28 July, 2001, pp. 1062–1067.
[7] P.P. Angelov, V.I. Hanby, R.A. Buswell, J.A. Wright, Automatic generation of fuzzy rule-
based models from data by genetic algorithms, in: R. John, R. Birkenhead (Eds.),
Developments in Soft Computing, Physica-Verlag, Heidelberg, 2000, pp. 31–40.
[8] K.J. Astroem, B. Wittenmark, Computer Controlled Systems: Theory and Design, Prentice-
Hall, Englewood Cliffs, NJ, USA, 1984.
[9] B. Carse, T.C. Fogarty, A. Munro, Evolving fuzzy rule-based controllers using GA, Fuzzy Sets
and Systems 80 (1996) 273–294.
[10] C.K. Chiang, H.-Y. Chung, J.J. Lin, A self-learning fuzzy logic controller using genetic
algorithms with reinforcements, IEEE Transactions on Fuzzy Systems 5 (1996) 460–
467.
[11] M.G. Cooper, J.J. Vidal, Genetic design of fuzzy controllers, in: S.K. Pal, P.P. Wang (Eds.),
Genetic Algorithms and Pattern Recognition, CRC Press, USA, 1996, pp. 283–298 (Chapter
13).
[12] D. Driankov, H. Hellendoorn, M. Reinfrank, An Introduction to Fuzzy Control, Springer-
Verlag, Berlin, Germany, 1993.
[13] W.S. Lin, C.-H. Tsai, Self-organizing Fuzzy control of multi-variable systems using learning
vector quantization network, Fuzzy Sets and Systems 124 (2001) 197–212.
[14] F.-J. Lin, C.-H. Lin, P.-H. Shen, Self-constructing fuzzy neural network speed controller for
permanent-magnet synchronous motor drive, IEEE Transactions on Fuzzy Systems 9 (2001)
751–759.
[15] D. Psaltis, A. Sideris, A.A. Yamamura, A multilayered neural network controller, IEEE
Transactions on Control Systems Management 8 (1988) 17–21.
[16] L.-X. Wang, Fuzzy Systems are Universal Approximators, in: Proceedings of the IEEE
Conference on Fuzzy Systems FUZZ_IEEE, San Diego, CA, USA, 1992, pp. 1163–
1170.
 P. Angelov / Information Sciences 161 (2004) 21–35 35

[17] H. Pomares, I. Rojas, J. Gonzalez, M. Damas, A. Prieto, An Enhanced Self-organizing

Controller for Real-time Process-control Applications, in: Proceedings of the 10th IEEE
Conference on Fuzzy Systems FUZZ-IEEE’01, Melbourne, Australia, December 2001.
[18] I. Rojas, H. Pomares, A. Prieto, On-line adaptive fuzzy controller: application to helicopter
stabilization, in: Proceedings of the EUNITE-2002, on CD-ROM, Albufeira, Portugal, 19–21
September 2002, pp.621–626.