3234 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 54, NO.

6, DECEMBER 2007

Fuzzy Identification Based on a Chaotic Particle

Swarm Optimization Approach Applied
to a Nonlinear Yo-yo Motion System
Leandro dos Santos Coelho and Bruno Meirelles Herrera

Abstract—The identification of uncertain and nonlinear systems The identification problem in TS modeling consists of two
is an important and challenging problem. Fuzzy models, partic- major parts: structure identification and parameter identifi-
ularly Takagi–Sugeno (TS), have received particular attention in cation. Furthermore, the TS model comprises premise-part
the area of nonlinear identification due to their potentialities to
approximate any nonlinear behavior. A method of nonlinear iden- identification and consequent-part identification. The iden-
tification based on the TS fuzzy model and optimization procedure tification of the premise part consists of determining the
is proposed in this paper. Chaotic particle swarm optimization premise-space partition and extracting the number of rules.
(CPSO) algorithms, based on chaotic Zaslavskii map sequences, The consequent-part identification consists of determining the
combined with efficient Gustafson–Kessel (GK) clustering algo- structure of the rules’ output parts. Lastly, the parameter learn-
rithm are proposed here for the design of the premise part of pro-
duction rules, while the least-mean-square technique is utilized for ing task consists of determining the system parameters, so
the subsequent part of the production rules of the TS fuzzy model. that a performance measurement based on the output errors is
An experimental case study using a nonlinear yo-yo motion control minimized.
system is analyzed by the proposed algorithms. The numerical TS fuzzy models can be designed by means of clustering,
results presented here indicate that the traditional particle swarm classical nonlinear optimization methods, evolutionary algo-
optimization algorithm and, particularly, the CPSO combined
with GK algorithms are effective in building a good TS fuzzy rithms, and others. An alternative is the investigation of particle
model for nonlinear identification. swarm optimization (PSO) for the optimization of membership
functions parameters in the TS model design.
Index Terms—Chaotic map, clustering algorithm, fuzzy identifi-
cation, nonlinear systems, optimization, particle swarm optimiza- Social-insect societies are distributed systems which, de-
tion, system identification. spite the simplicity of their individuals, present a highly
structured social organization. As a result of this organiza-
tion, insect societies can accomplish complex tasks that, in
I. I NTRODUCTION
some cases, far exceed the individual capabilities of a sin-

K NOWLEDGE of the behavior of real systems based on

dynamic models is important in many fields of science
and engineering. System identification can be described as the
gle ant [11]. The field of swarm intelligence is an emerging
research area that presents features of self-organization and
cooperation principles among group members bioinspired by
science of building mathematical models of dynamic systems social-insect societies [12]. Swarm intelligence is inspired by
based on observed input and/or output signals. The practical nature, based on the fact that the live animals of a group
importance of process-model identification has been recognized contribute with their individual experiences to the group, ren-
for many years [1]. dering it stronger to face other groups. The most familiar
Recently, nonlinear techniques using fuzzy identification and representatives of swarm intelligence in optimization prob-
control have received a great deal of attention [2]–[6]. A central lems are the food-searching behavior of ant colonies [11],
feature of fuzzy systems is that they are based on the concept PSO [13], artificial immune systems [14], and bacterial for-
of fuzzy coding of information and operating with fuzzy sets aging [15].
instead of numbers [7]. Takagi–Sugeno’s (TS) fuzzy model [8], Most heuristic optimization methods combined with fuzzy
[9] exhibits both high nonlinearity and simple structure. As systems for nonlinear identification applications are used for
reported in literature, it is capable of approximating a complex rule extraction, tuning of membership functions, and training
system using fewer fuzzy rules than conventional Mamdani- of fuzzy models using genetic algorithms. In this context,
type fuzzy models [10]. several hybrid methodologies of fuzzy models and genetic
algorithms have been intensively developed, with a variety of
works in nonlinear identification and control published in the
Manuscript received August 3, 2005; revised July 25, 2007. This work was literature in recent years [16]–[21]. The use of metaheuristics
supported by the National Council for Scientific and Technological Develop-
ment of Brazil—CNPq—under Grant 309646/2006-5/PQ. of PSO is an emerging approach in the TS fuzzy-model design
The authors are with the Industrial and Systems Engineering Graduate [22]–[26].
Program, Pontifical Catholic University of Paraná, PPGEPS/PUCPR, 80215- In this paper, the structural identifications of the premise and
901 Curitiba-PR, Brazil (e-mail: leandro.coelho@pucpr.br; bherrera@terra.
com.br). consequent parts of production rules of the TS fuzzy model are
Digital Object Identifier 10.1109/TIE.2007.896500 performed separately. This paper presents the design of the TS

0278-0046/$25.00 © 2007 IEEE

DOS SANTOS COELHO AND HERRERA: FUZZY IDENTIFICATION BASED ON A CHAOTIC PSO APPROACH 3235

fuzzy model based on the chaotic particle swarm optimization the consequent input term uji , and wj = [w0j , w1j . . . , wqj
j T
] are
(CPSO) approaches combined with the Gustafson–Kessel (GK) the polynomial coefficients that form the consequent parameter
clustering algorithm [27] for the premise-part design, and it also set. Each linguistic label Aji is associated with a membership
shows the least mean squares for calculation of the consequent function, µAj (zi ), which is described by
i
part of production rules of the TS fuzzy model for nonlinear


identification. An experimental case study using a yo-yo motion 1 (zi − mij )2
system is analyzed by the proposed design approach for nonlin- µAj (zi ) = exp − 2 (2)
i 2 σij
ear identification. The numerical results presented here indicate
that the PSO and, particularly, the CPSO combined with the GK
algorithm are effective in building a good TS fuzzy model for where mij and σij are the center and the spread of the
nonlinear identification. Gaussian-type membership function, respectively. The union of
The remaining sections of this paper are organized as fol- all these parameters formulates the set of premise parameters.
lows. Section II describes the fundamentals of TS fuzzy models. The firing strength of rule R(j) represents its excitation level
Section III then describes the optimization procedure based on and is given by
the concept of PSO and CPSO approaches. Section IV discusses
µj (z) = µAj (z1 ) · µAj (z2 ), . . . , µAjm (zm ). (3)
the yo-yo motion system prototype, while Section V ana- 1 2

lyzes the nonlinear identification results applied to the TS fuzzy

model. Lastly, Section VI presents our conclusion. The fuzzy sets pertaining to a rule form a fuzzy region
(cluster) within the premise space, Aj1 xAj2 x, . . . , xAjm , with
a membership distribution described by (3). Given the input
II. TS F UZZY M ODEL vectors z and uj , j = 1, . . . , M , the final output of the fuzzy
model is inferred by taking the weighted average of the local
Developing mathematical models of real systems is a central outputs gj (uj ), which is given by
topic in many fields of engineering and science. Models can
be used for computational simulations, analysis of complex 
M
systems, design of new industrial processes, and control of y= νj (z) · gj (uj ) (4)
systems [28]. j=1
For nonlinear dynamic systems, traditional techniques of
modeling and identification are difficult to implement and where M denotes the number of rules, and νj (z) is the normal-
sometimes impracticable. However, other techniques based ized firing strength of R(j) , which is defined as
on fuzzy systems, such as the TS fuzzy model, are increas-
ingly used for the identification of this kind of process. The µj (z)
νj (z) = M . (5)
TS fuzzy model is based on rules in which the consequent is j=1 µj (z)
not a linguistic variable, as in the Mamdani-type fuzzy model,
but a function of the input variables. A relevant aspect of the The TS model’s structure is identified based on the PSO or
TS fuzzy model is its power of representation, particularly the CPSO combined with a GK algorithm for the premise-part
in describing complex processes. This fuzzy system allows optimization, while the consequent part is optimized by the
complex systems to be decomposed into simple subsystems. least-mean-square method [1]. The PSO and the new CPSO
The identification of the TS fuzzy model involves two pri- approaches for fuzzy-model optimization are presented in the
mary tasks: parameter tuning and structure optimization. The section below.
parameter tuning procedure deals with the estimation of a
feasible set of parameters for a given structure. The structure
optimization procedure aims to find the optimal structure of the III. PSO AND CPSO A PPROACHES FOR O PTIMIZATION
local models, the relevant premise variables, and the suitable OF THE TS F UZZY M ODEL
partition of the premise space. The TS models consist of lin- A. Fundamentals of the Classical PSO Approach
guistic IF-THEN rules that can be represented by the following
general form: The proposal of PSO algorithm was put forward by sev-
eral scientists who developed computational simulations of the
R(j) : IF z1 IS Aj1 AND, . . . , AND zm IS Ajm movement of organisms such as flocks of birds and schools
of fish. Such simulations were heavily based on manipulating
THEN gj = w0j + w1j uj1 + · · · + wqj
j j
uqj . (1) the distances between individuals, i.e., the synchrony of the
behavior of the swarm was seen as an effort to keep an optimal
The IF preconditioned statements define the premise part, distance between them [29].
while the THEN rule functions constitute the consequent part of In theory, at least, individuals of a swarm may benefit from
the fuzzy model; z = [z1 , . . . , zm ]T , i = 1, . . . , m, is the input the prior discoveries and experiences of all the members of a
vector of the premise p, and Aji are the labels of fuzzy sets. The swarm when foraging. The fundamental point of developing
parameter u = [uj1 , . . . , ujqj ]T represents the input vector to the PSO is a hypothesis in which the exchange of information
consequent part of R(j) that comprises qj terms; gj = gj (uj ) among creatures of the same species offers some sort of evo-
denotes the jth rule output, which is a linear polynomial of lutionary advantage.
3236 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 54, NO. 6, DECEMBER 2007

The PSO originally developed by Kennedy and Eberhart in based on uniform probability distribution functions in the
1995 [30], [31] is a population-based swarm algorithm. Sim- range [0,1] of the jth design variable of ith particle.
ilarly to a genetic algorithm [32], the PSO is an optimization 6) Loop to step 2) until a stopping criterion is met, usually
tool based on a population, where each member is seen as a a sufficiently good fitness or a maximum number of
particle, and each particle is a potential solution to the problem iterations (generations), kmax .
under analysis. Each particle in the PSO has a randomized
The use of variable w, called the inertia weight, was proposed
velocity associated with it, which moves through the space of
by Shi and Eberhart [38]. This parameter is responsible for
the problem. However, unlike genetic algorithms, the PSO does
dynamically adjusting the velocity of the particles; therefore,
not have operators, such as crossover and mutation. The PSO
it is responsible for balancing between local and global search,
does not implement the survival of the fittest individuals; rather,
hence requiring fewer iterations for the algorithm to converge.
it implements the simulation of social behavior.
A low value of inertia weight implies a local search, while a
Each particle in the PSO keeps track of its coordinates in
high value leads to a global search.
the problem space, which are associated with the best solution
Applying a high inertia weight at the start of the algorithm
(fitness) it has achieved so far. This value is called personal
and making it decay to a low value through the PSO execution
best (pbest). Another “best” value that is tracked by the global
make the algorithm search globally, at the beginning of the
version of the particle-swarm optimizer is the overall best value
search, and search locally at the end of the execution. The
and its location obtained so far by any particle in the population.
following weighting function w is used in (6):
This location is called global best (gbest).
The PSO concept consists of, in each time step, changing
wmax − wmin
(accelerating) the velocity of each particle flying toward its w = wmax − k. (8)
pbest and gbest locations (global version of the PSO). Accel- kmax
eration is weighted by random terms, with separate random
Equation (8) shows how the inertia weight is updated, con-
numbers being generated for acceleration toward pbest and
sidering that wmax and wmin are the initial and final weights,
gbest locations, respectively. The procedure for implementing
respectively [39].
the global version of PSO is given by the following steps
Positive constants c1 and c2 are called cognitive and social
[33]–[35].
components, respectively. These are the acceleration constants
1) Initialize a population (array) of particles with random responsible for varying the particle velocity toward the pbest
positions and velocities in the n-dimensional problem and the gbest. Particle velocities in each dimension are clamped
space using a uniform probability distribution function. to a maximum velocity Vmax . If the velocity in that dimension
2) Evaluate the fitness value of each particle. exceeds Vmax , which is a parameter specified by the user, then
3) Compare each particle’s fitness with the particle’s pbest. the velocity in that dimension is limited to Vmax .
If the current value is better than the pbest, then set Vmax is a parameter serving to determine the resolution with
the pbest value equal to the current value and the pbest which the regions around the current solutions are searched.
location equal to the current location in n-dimensional If Vmax is too high, the PSO facilitates a global search, and
space. particles might fly past good solutions. Conversely, if Vmax is
4) Compare the fitness with the population’s overall previ- too small, the PSO facilitates a local search, and particles may
ous best. If the current value is better than gbest, then not explore sufficiently beyond locally good regions. Previous
reset gbest to the current particle’s array index and value. experience with PSO (trial and error, mostly) led us to set the
5) Change the velocity and position of the particle according Vmax to 20% of the dynamic range of the particle in each
to (6) and (7), respectively [36], [37] dimension.

νi (k + 1) = w · νi (k) + c1 · udi,j (k) · [pi (k) − xi (k)]

B. Chaotic PSO Approach
+ c2 · U di,j (k) · [pg (k) − xi (k)] (6)
Chaos theory is recognized as very useful in many engi-
xi (k + 1) = xi (k) + ∆t · νi (k + 1). (7) neering applications. An essential feature of chaotic systems is
that small changes in the parameters or the starting values for
Equation (7) represents the position updating, accord- the data lead to different future behaviors, such as stable fixed
ing to its previous position and its velocity [see (6)], points, periodic oscillations, bifurcations, and ergodicity. These
considering that ∆t = 1, where k is the current iter- behaviors can be analyzed based on the meaning of Lyapunov
ation number, xi = [xi1 , xi2 , . . . , xin ]T stands for the exponents and the attractor theory [40], [41].
position of the ith particle and nth decision variable, Based on these features, much of the chaos as a science
νi = [νi1 , νi2 , . . . , νin ]T stands for the velocity of the is connected with the notion of “sensitive dependence on
ith particle, and pi = [pi1 , pi2 , . . . , pin ]T represents the initial conditions” [60]. An apparent paradox is that chaos is
best previous position (the position giving the best fitness deterministic, which is generated by fixed rules that do not
value) of the ith particle. Index g represents the index themselves involve any elements of change. In principle, the
of the best particle among all the particles in the group. future is completely determined by the past, but, in practice,
Variables udi,j (t) and U di,j (t) are two random functions small uncertainties are unpredictable in the long term [42], [43].
DOS SANTOS COELHO AND HERRERA: FUZZY IDENTIFICATION BASED ON A CHAOTIC PSO APPROACH 3237

ods. New PSO approaches are proposed here based on the

Zaslavskii map.
The convergence properties of the PSO methods are strongly
connected to the random sequence applied on variation op-
erators of (6) during a run. In particular, it can be shown
that when different random sequences are used during the
evolution, the final results may effectively be very close but
not equal [48]. However, there are no analytical results that
guarantee an improvement of the performance indices of the
PSO methods depending on the choice of particular generator
in (6) instead of classical technique based on variables udi,j (t)
and U di,j (t). Recently, chaotic sequences have been adopted
instead of random ones, and interesting results have been shown
in optimization applications [49]–[55].
The performance of the simple PSO greatly depends on its
design parameters, and it often suffers the problem of being
trapped in local optima so as to be premature convergence.
Fig. 1. Example of evolution of Zaslavskii map for ν = 400, r = 3, and The choice of the chaotic Zaslavskii map in the PSO design is
a = 12.6695. justified theoretically by its unpredictability, i.e., by its spread-
spectrum characteristic and large Lyapunov exponent (a quan-
Optimization algorithms based on the chaos theory are titative measure of chaos) [56], [57]. Details of the Lyapunov
stochastic search methodologies that differ from any of the exponent are presented in [40]–[42], and [46]. In this context,
existing evolutionary algorithms. Due to the nonrepetition of due to ergodic and dynamic properties of the Zaslavskii-map
chaos, it can carry out overall searches at higher velocities than variables, the PSO using chaotic search based on the chaotic
stochastic ergodic searches that depend on probabilities. Zaslavskii map is more capable of escaping from local optima
In the PSO design, the main advantage of chaotic optimiza- than random search.
tion approaches is the maintenance of population diversity in These new PSO approaches combined with chaotic se-
the problem of interest. Based on Clerc and Kennedy [44] and quences (CPSO) based on the Zaslavskii map are described as
Trelea’s [45] studies, the parameters w, c1 , and c2 are generally follows:
the key factors that affect the classical PSO convergence. In Approach 1— CPSO1: Parameter c1 of (6) is modified by
fact, however, parameters c1 and c2 cannot entirely ensure the the following equation:
ergodicity of the optimization in phase search because they are
constant factors in the classical PSO. The variable w is also νi (k + 1) = w · νi (k) + z1 · udi,j (k) · [pi (k) − xi (k)]
limited for procedures of linear or exponential reduction with
the evolution of generations. +c2 · U di,j (k) · [pg (k) − xi (k)] (11)
Therefore, this paper provides approaches introducing
chaotic mapping with ergodicity, irregularity, and stochastic where z1 is a function based on the results
property in the PSO to improve the global convergence. The of y2 (t) of the Zaslavskii map with scaled
use of chaotic sequences in the PSO can be helpful to escape values between 0.5 and 2.5.
more easily from local minima than can be done through the Approach 2— CPSO2: Parameter c2 of (6) is modified by
traditional PSO. the following equation:
The literature is rich in chaotic time-series sequences, such as
logistic map, tent map, Hénon map, Ikeda map, Chua’s system, νi (k + 1) = w · νi (k) + c1 · udi,j (k) · [pi (k) − xi (k)]
Lorenz systems, Lozi map, and others [40], [41], [46]. The +z2 · U di,j (k) · [pg (k) − xi (k)] (12)
interesting dynamic system evidencing chaotic behavior is the
Zaslavskii map [47], whose equation is given by where z2 is a function of the results of y2 (t)
of the Zaslavskii map with scaled values
y1 (t) = mod [y1 (t − 1) + ν + ay2 (t), 1] (9) between 0.5 and 2.5.
y2 (t) = cos (2πy1 (t − 1)) + e−r y2 (t − 1) (10) Approach 3— CPSO3: Parameters c1 and c2 of (6) are
modified by the following equation:
where mod is the modulus (signed remainder after division).
νi (k + 1) = w · νi (k) + z1 · udi,j (k) · [pi (k) − xi (k)]
The Zaslavskii map shows a strange attractor with large
Lyapunov exponent for ν = 400, r = 3, and a = 12.6695, +z2 · U di,j (k) · [pg (k) − xi (k)] (13)
as shown in Fig. 1. In this case, the values of y2 (t)[−1.0512,
1.0512]. where z1 and z2 are functions of the results
The design of methods to improve the convergence of PSO of y2 (t) of the Zaslavskii map with scaled
is a challenging issue in the design of optimization meth- values between 0.5 and 2.5.
3238 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 54, NO. 6, DECEMBER 2007

Approach 4— CPSO4: Parameter w of (6) is modified by and Kessel [27] extended the standard fuzzy c-means algorithm
the following equation: by employing an adaptive distance norm in order to detect
clusters of different geometrical shapes in one data set. Details
νi (k + 1) = wz · νi (k) + c1 · udi,j (k) · [pi (k) − xi (k)] of GK clustering algorithm are presented in [2], [27], and [63].
+c2 · U di,j (k) · [pg (k) − xi (k)] (14) Clustering techniques, including the GK algorithm, belong
to the classes of unsupervised learning methods since they
where wz is a function based on the do not use prior class identifiers. The GK algorithm is based
Zaslavskii map with scaled values between on the minimization of a cost function (objective function)
0.4 and 0.9. regarding the degree to which the data belong to the clusters
Approach 5— CPSO5: Parameters w and c1 of (6) are mod- and the degree of dissimilarity between them. The combination
ified by the following equation: of the GK algorithm and the PSO or CPSO is very useful. In
this case, the GK algorithm optimizes the centers of Gaussian
νi (k + 1) = wz · νi (k) + z1 · udi,j (k) · [pi (k) − xi (k)] functions of membership functions (premise part of production
+c2 · U di,j (k) · [pg (k) − xi (k)] (15) rules) of the TS fuzzy model. The PSO or CPSO then employs
the solution of the GK algorithm in the initial population of
where wz is a function based on the results particles and optimizes the centers and spreads of Gaussian
of y2 (t) of the Zaslavskii map with scaled functions. In each evaluation of a solution, the classical method
values between 0.4 and 0.9, and z1 is a func- of least-mean-squares calculates the consequent part of produc-
tion based on the Zaslavskii map with values tion rules of the TS fuzzy model. Fig. 2 shows the diagram of
between 0.5 and 2.5. the proposed hybrid method.
Approach 6— CPSO6: Parameters w and c2 of (6) are mod-
ified by the following equation:
IV. D ESCRIPTION OF THE Y O -Y O M OTION S YSTEM
νi (k + 1) = wz · νi (k) + c1 · udi,j (k) · [pi (k) − xi (k)]
Yo-yo playing is considered a representative example of
+z2 · U di,j (k) · [pg (k) − xi (k)] (16) open-loop unstable control problems that involve intermittent
dynamic environments. Stable control of yo-yo playing relies
where wz is a function based on the results on a proper phase relationship between the controller’s action
of y2 (t) of the Zaslavskii map with scaled and the motion of the yo-yo [67].
values between 0.4 and 0.9, and z2 is a func-
Control of a yo-yo requires an asymmetric nontrivial con-
tion based on the Zaslavskii map with values troller with nonlinearity due to the unique features of the
between 0.5 and 2.5.
yo-yo system [68]. Due to its asymmetric nonlinearity, it seems
Approach 7— CPSO7: Parameters w, c1 , and c2 of (6) are
difficult to control a yo-yo by a linear controller. In this context,
modified by the following equation: the development of automatic control systems that efficiently
νi (k + 1) = wz · νi (k) + z1 · udi,j (k) · [pi (k) − xi (k)] control a yo-yo represents a challenge for the development of
electromechanical designs [68]–[71]. One of the main difficul-
+z2 · U di,j (k) · [pg (k) − xi (k)] (17) ties is the lack of sensors to obtain the motion measure of the
yo-yo position. Another difficulty is the lack of mathematical
where wz is a function based on the results models of this measurement device type, which justifies the use
of y2 (t) of the Zaslavskii map with scaled of the TS fuzzy model to identify the dynamic behavior of a
values between 0.4 and 0.9, and z2 is a func-
yo-yo motion in a real system.
tion based on the Zaslavskii map with values The control system prototype uses a yo-yo, and a dc motor
between 0.5 and 2.5. for its motion presents nonlinearity and complex behavior. The
reason for using this process is the possibility of proving the
C. PSO or CPSO Combined With GK Clustering Algorithm efficiency and flexibility of model-based control systems. A
block diagram of the described system and a photograph of
Clustering based on fuzzy approaches is a relevant applica-
the system are shown in Figs. 3 and 4, respectively [72], [73].
tion of fuzzy set theory and concepts of membership functions. The components of this prototype are divided into software and
The membership function of an object describes to what degree
hardware modules, which are described as follows.
that object is a member of a given set [2].
Cluster analysis is a technique that is used to seek out 1) Control module (software): consists of the implemen-
data, dividing all objects (samples) into smaller subgroups and tation of control techniques, such as proportional–
classifying them according to the similarities among them. A integral–derivative and fuzzy logic controllers integrated
fuzzy cluster is a fuzzy subset of the set of objects, with the into a computer with communication with the yo-yo
membership function of each object representing the degree to system using an input/output interface.
which it belongs to that cluster [58]. 2) Sensor module (hardware/firmware): The sensors em-
There are many conceptions of clustering method in the ployed include digital electronic circuits (power ampli-
literature [28], [59]–[66]. One of the most efficient of these fication), A/D and D/A converters, and microcontroller
clustering methods is the GK clustering algorithm. Gustafson running firmware.
DOS SANTOS COELHO AND HERRERA: FUZZY IDENTIFICATION BASED ON A CHAOTIC PSO APPROACH 3239

Fig. 2. Diagram of TS fuzzy system design using PSO or CPSO, GK clustering algorithm, and least-mean-square method.

Fig. 3. Block diagram of yo-yo motion system.

3) Actuator module (hardware/firmware): consists of dc mo-

tors integrated to the sensor module, electronic circuits,
and microcontroller running firmware.
4) Sensor submodule: made up of 16 infrared light-emitting
diodes (LEDs) that are able to inform the position of the
yo-yo.
The prototype modules are composed of hardware and
firmware and are connected to the same printed circuit board
(control board). The control board contains two hardware Fig. 4. Photograph of prototype of yo-yo motion system.
3240 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 54, NO. 6, DECEMBER 2007

Fig. 5. Procedure for identification of yo-yo motion system.

Fig. 7. Yo-yo motion system prototype’s input and output data.

A computer with a data-acquisition board for generating the

control signal (identification in closed loop using a proportional
controller) and position value of the yo-yo was used to obtain
system measurements. In the identification procedure based on
the TS fuzzy model, 290 samples of input (tension applied to
the dc motor in volts) and output (position of yo-yo in cm)
were collected with a time sampling of 40 ms (see Fig. 7). The
tension value corresponds to the maximum-value configuration
of the driver in PWM control of a dc motor.
Experiments for the estimation phase of the mathematical
model of the yo-yo motion system are carried out using samples
1 to 150. For the validation phase, the fuzzy model uses the in-
Fig. 6. Structure of NARX model for prediction of TS fuzzy system. put and output signals of samples 151 to 290. The system iden-
tification by the TS fuzzy model is appropriate if a performance
modules and communicates with a personal computer (main index is in permissible values for the user’s needs. In this case,
module) through the RS-232 I/O interface. All the components the fitness function for maximization proposes using the PSO
used for the yo-yo system are off-the-shelf items to keep the and CPSO approaches and is given by the harmonic mean of
cost minimal. multiple correlation indices of estimation and validation phases.
The fitness is calculated using the following expressions:

V. A NALYSIS R ESULTS OF N ONLINEAR I DENTIFICATION Rh2 = Rest

2 2
+ Rval (18)
The goal of the system identification is to allow the adjust- 150 2
t=1 [y(t) − ŷ(t)]
ment of a mathematical model to a dynamic system structure, Rest = 1 − 
2
150 2 (19)
based on measurements collected by the adjustment of parame- t=1 [y(t) − ȳ]
ters and/or of the model, until the system output is as close as 290 2
t=151 [y(t) − ŷ(t)]
possible to the samples of the measured outputs. The procedure 2
Rval =1 −  290 2 (20)
for experimental identification of a process comprises basic t=151 [y(t) − ȳ]
stages, as shown in Fig. 5.
In practice, system identification is an iterative procedure. where Rh2 is the harmonic mean of the multiple correlation
2
The lack of a priori information regarding the process model index (fitness function to be optimized), Rest is the multiple
2
will require that each step be initially examined in a superficial correlation index of the estimation phase, Rval is the multiple
manner. The mathematical model, which is employed in this correlation index of the validation phase, y(t) is the output of
paper, to represent the yo-yo motion system is a Nonlinear the real system, ŷ(t) is the output estimated by the TS fuzzy
AutoRegressive with eXogenous inputs (NARX), as shown model, and ȳ is the mean value of the system’s output.
in Fig. 6. In this case, the NARX model with series–parallel When R2 = 1.0, it indicates an overfitting phenomenon,
conception is used for one-step-ahead prediction of the TS i.e., a model error exists. An R2 value between 0.9 and 1.0
fuzzy model. In this case, the resulting fuzzy inference system is considered sufficient for applications in identification and
is a first-order TS model [28] [see (1)]. model-based control [74].
DOS SANTOS COELHO AND HERRERA: FUZZY IDENTIFICATION BASED ON A CHAOTIC PSO APPROACH 3241

TABLE I
SUMMARY OF THE RESULTS (BEST OF 30 INDEPENDENT RUNS WITH 50 GENERATIONS) FOR DIFFERENT
PSO AND CPSO STRATEGIES OF TS FUZZY SYSTEM DESIGN (NARX WITH TWO RULES) WITH
THREE INPUTS [u(t), y(t − 1), AND y(t − 2)] AND AN OUTPUT [ŷ(t)]

TABLE II
SUMMARY OF THE RESULTS (BEST OF 30 INDEPENDENT RUNS WITH 50 GENERATIONS) FOR DIFFERENT
PSO AND CPSO STRATEGIES OF TS FUZZY SYSTEM DESIGN (NARX WITH THREE RULES) WITH
THREE INPUTS [u(t), y(t − 1), AND y(t − 2)] AND AN OUTPUT [ŷ(t)]

Based on previous experience with the PSO and CPSO the CPSO approaches. The CPSO approaches present better
approaches (trial and error, mostly), we set the population size results in relation to the mean fitness than does the classical
N equal to 20, and kmax was set to 50 generations (stopping PSO1 algorithm. The CPSO5 shows the best performance with
criterion) for the 30 runs for the TS fuzzy-model optimization. fitness Rh2 = 0.9331. However, the CPSO2 and CPSO3 provide
The values of c1 = c2 = 0.7 were adopted for the classical results very close to those of the CPSO5. The CPSO3 presents
PSO1 with (6)–(8) and the CPSO1 to CPSO7 approaches. the best mean, and the CPSO6 has significant values with
The three chosen vectors of input for the TS fuzzy model a small standard deviation in convergence and also the best
were the following: [u(t), y(t − 1), and y(t − 2)]. The space minimum fitness of the approaches tested here.
searches for centers and spreads of Gaussian membership func- In general, for the case study of optimization of the TS
tions of TS fuzzy models by the CPSO optimization approaches fuzzy model with three rules, the PSO1 and CPSO approaches
are [−5; 5] and [0.001; 5], respectively. presented similar results. However, the CPSO2 is the best with
Tables I and II present the simulation results (best of Rh2 = 0.9357. This result is slightly better than the case of the
30 experiments with 50 generations for each run) for the TS fuzzy model with two rules. The CPSO5 is superior in terms
different PSO and CPSO strategies for optimization of the of mean, minimum, and standard deviations to the other tested
TS fuzzy model. In Table I, the PSO and CPSO approaches PSO1 and CPSO approaches in this particular case.
optimize 12 parameters (decision variables), e.g., antecedent of Tables I and II indicate that, for two and three rules, all
two rules (six centers and six spreads of Gaussian membership the results of the TS fuzzy model showed multiple correlation
functions), and in Table II, this optimization method tunes indexes of best results higher than 0.9230, which are nec-
18 parameters (9 centers and 9 spreads). essary and therefore appropriate for nonlinear identification
As indicated in Tables I and II, the results of the optimized and controller-design applications. The best results shown in
TS fuzzy model present precision and provide an appropriate Figs. 8 and 9 represent the TS design with two and three rules,
experimental mathematical model for the yo-yo motion system. respectively. Fig. 10 shows the membership functions of the
The TS fuzzy model presents a black-box model of a nonlinear best result of two cases using CPSO2 with three rules.
yo-yo system, with adequate treatment of the nonlinearities of
the dynamic system due to the inherent features of the TS fuzzy
VI. C ONCLUSION
model, which deal with complex processes.
For the case study of the TS fuzzy-model optimization with In multimodal problems such as the TS fuzzy design, the
two rules, there is a consistent performance pattern across all classical PSO tends to suffer from premature convergence. This
3242 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 54, NO. 6, DECEMBER 2007

Fig. 8. Best result of Rh 2 for TS fuzzy design with two rules in yo-yo motion
Fig. 9. Best result of Rh2 for TS fuzzy design with three rules in yo-yo motion
system identification using CPSO5 (see Table I). (a) Outputs. (b) Error signal.
system identification using CPSO2 (see Table II). (a) Outputs. (b) Error signal.

is due to a decrease of diversity in the search space, which leads All CPSO approaches improve the performance and con-
to a fitness stagnation in the optimization process. vergence of classical PSO for the optimization of the premise
In this paper, alternative optimization methods, called the part of the TS fuzzy model. The proposed CPSO approaches
CPSO approaches combined with the GK algorithm, for the perform better in terms of mean fitness than the classical PSO1.
premise part of the design of the TS fuzzy models are analyzed For the case study of the TS fuzzy-model optimization with
and compared. These approaches are the PSO algorithms based two rules, it is interesting to note that the CPSO5 presents the
on Zaslavskii chaotic sequences in the design of w, c1 , and c2 best result of fitness but worse in terms of mean convergence
parameters. than the CPSO2 and CPSO3. For the case study with three
The most striking feature of chaos is the unpredictability of rules, CPSO5 outperforms the other tested approaches in terms
the future despite a deterministic time evolution, a feature that of mean-fitness results.
is useful in the design of stochastic optimization methods. The However, the precision, computational complexity, and or-
proposed CPSO approaches deal with the maintenance of the ders of the input vectors of the TS fuzzy model must be
diversity of particle populations of classical PSO for preventing analyzed in detail in future works.
premature convergence. As prospects for future works are linked to the yo-yo motion
The experimental results showed that the TS fuzzy model system design, there is a possibility of assessing a comparison
with the CPSO approaches and the GK algorithm presented of the CPSO approaches with other chaotic sequences in the
successful results due to precision in the case study of the PSO design reported in recent literature [75]–[77] in multivari-
prediction of the nonlinear dynamics of a yo-yo motion system. able nonlinear identification problems.
DOS SANTOS COELHO AND HERRERA: FUZZY IDENTIFICATION BASED ON A CHAOTIC PSO APPROACH 3243

referees for their valuable comments which improved the initial

manuscript.

R EFERENCES
[1] L. Ljung, System Identification: Theory for the User. New York:
Prentice-Hall, 1987.
[2] A. L. Elshafei and F. Karray, “Variable-structure-based fuzzy-logic iden-
tification of a class of nonlinear systems,” IEEE Trans. Control Syst.
Technol., vol. 13, no. 4, pp. 646–653, Jul. 2005.
[3] F. Betin, A. Sivert, A. Yazidi, and G. Capolino, “Determination of scaling
factors for fuzzy logic control using the sliding-mode approach: Applica-
tion to control of a dc machine drive,” IEEE Trans. Ind. Electron., vol. 54,
no. 1, pp. 296–309, Feb. 2007.
[4] P. Baranyi, A. R. Várkonyi-Kóczy, Y. Yam, and R. J. Patton, “Adap-
tation of TS fuzzy models without complexity expansion: HOSVD-
based approach,” IEEE Trans. Instrum. Meas., vol. 54, no. 1, pp. 52–60,
Feb. 2005.
[5] M. Cheng, Q. Sun, and E. Zhou, “New self-tuning fuzzy PI control of a
novel doubly salient permanent-magnet motor drive,” IEEE Trans. Ind.
Electron., vol. 53, no. 3, pp. 814–821, Jun. 2006.
[6] C.-F. Juang and J.-S. Chen, “Water bath temperature control by a recur-
rent fuzzy controller and its FPGA implementation,” IEEE Trans. Ind.
Electron., vol. 53, no. 3, pp. 941–949, Jun. 2006.
[7] B. Wu and X. Yu, “Fuzzy modelling and identification with genetic al-
gorithm based learning,” Fuzzy Sets Syst., vol. 113, no. 3, pp. 351–365,
Aug. 2000.
[8] T. Takagi and M. Sugeno, “Fuzzy identification of systems and its ap-
plications to modeling and control,” IEEE Trans. Syst., Man Cybern.,
vol. SMC-15, no. 1, pp. 116–132, 1985.
[9] M. Sugeno and G. T. Kang, “Structure identification of fuzzy model,”
Fuzzy Sets Syst., vol. 28, no. 1, pp. 15–33, Oct. 1988.
[10] E. H. Mamdani and S. Assilian, “An experiment in linguistic synthesis
with a fuzzy logic controller,” Int. J. Man-Mach. Studies, vol. 7, no. 1,
pp. 1–13, 1975.
[11] M. Dorigo and T. Stützle, Ant Colony Optimization, A Bradford Book.
Cambridge, MA: MIT, 2004.
[12] C. Sierakowski and L. S. Coelho, “Study of two swarm intelligence
techniques for path planning of mobile robots,” in Proc. 16th IFAC Word
Congr., Prague, Czech Republic, 2005.
[13] J. F. Kennedy, R. C. Eberhart, and R. C. Shi, Swarm Intelligence. San
Francisco, CA: Morgan Kaufmann, 2001.
[14] L. N. Castro and J. I. Timmis, Artificial Immune Systems: A New Compu-
tational Intelligence Approach. London, U.K.: Springer-Verlag, 2002.
[15] K. M. Passino, “Biomimicry of bacterial foraging for distributed opti-
mization and control,” IEEE Control Syst., vol. 22, no. 3, pp. 52–67,
Jun. 2002.
[16] K. M. Chow and A. B. Rad, “On-line fuzzy identification using genetic
algorithms,” Fuzzy Sets Syst., vol. 132, no. 2, pp. 147–171, Dec. 2002.
[17] S.-K. Oh, W. Pedrycz, and H.-S. Park, “A new approach to the de-
velopment of genetically optimized multilayer fuzzy polynomial neural
networks,” IEEE Trans. Ind. Electron., vol. 53, no. 4, pp. 1309–1321,
Jun. 2006.
[18] B.-J. Park, W. Pedrycz, and S.-K. Oh, “Identification of fuzzy models with
the aid of evolutionary data granulation,” Proc. Inst. Electr. Eng.—Control
Theory Applications, vol. 148, no. 5, pp. 406–418, Sep. 2001.
[19] M. Iwasaki, M. Miwa, and N. Matsui, “GA-based evolutionary identifica-
tion algorithm for unknown structured mechatronic systems,” IEEE Trans.
Ind. Electron., vol. 52, no. 1, pp. 300–305, Feb. 2005.
[20] M. F. Azeem, M. Hanmandlu, and M. H. N. Ahmad, “Evolutive learning
algorithms for fuzzy modeling,” Int. J. Smart Eng. Syst. Design, vol. 5,
no. 4, pp. 205–224, Oct.–Dec. 2003.
[21] P. P. Angelov and R. A. Buswell, “Automatic generation of fuzzy rule-
based models from data by genetic algorithms,” Inf. Sci., vol. 150, no. 1/2,
pp. 17–31, Mar. 2003.
[22] A. P. Ghoshal, “Optimizations of PID gains by particle swarm optimiza-
tions in fuzzy-based automatic generation control,” Elect. Power Syst.
Res., vol. 72, no. 3, pp. 203–212, 2004.
Fig. 10. Designed membership functions for TS fuzzy model using CPSO2. [23] A. Chatterjee, K. Pulasinghe, K. Watanabe, and K. Izumi, “A particle-
swarm-optimized fuzzy-neural network for voice-controlled robot sys-
tems,” IEEE Trans. Ind. Electron., vol. 52, no. 6, pp. 1478–1489,
Dec. 2005.
ACKNOWLEDGMENT [24] H. Zhenya, W. Chengjian, Y. Luxi, G. Xigi, Y. Susu, R. C. Eberhart, and
Y. Shi, “Extracting rules from fuzzy neural network by particle swarm op-
The authors would like to thank Prof. M. Kazmierkowski for timisation,” in Proc. IEEE World Congr. Comput. Intell., Evol. Comput.,
the review process. They are deeply grateful to the anonymous Anchorage, AL, 1998, pp. 74–77.
3244 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 54, NO. 6, DECEMBER 2007

[25] A. A. A. Esmin, A. R. Aoki, and G. Lambert-Torres, “Particle swarm [53] L. Li, Y. Yang, H. Peng, and X. Wang, “Parameters identification of
optimization for fuzzy membership functions optimization,” in Proc. chaotic systems via chaotic ant swarm,” Chaos Solitons Fractals, vol. 28,
IEEE Int. Conf. Syst., Man, Cybern., Hammamet, Tunisia, 2002, vol. 3, no. 5, pp. 1204–1211, Jun. 2006.
pp. 108–113. [54] X. Q. Zuo and Y. S. Fan, “A chaos search immune algorithm with its
[26] K. E. Parsopoulos, E. I. Papageorgiou, P. P. Groumpos, and application to neuro-fuzzy controller design,” Chaos Solitons Fractals,
M. N. Vrahatis, “A first study of fuzzy cognitive maps learning using par- vol. 30, no. 1, pp. 94–109, Oct. 2006.
ticle swarm optimization,” in Proc. IEEE Congr. Evol. Comput., Canberra, [55] D. Yang, G. Li, and G. Cheng, “On the efficiency of chaos optimization
Australia, 2003, vol. 2, pp. 1440–1447. algorithms for global optimization,” Chaos Solitons Fractals, vol. 34,
[27] D. E. Gustafson and W. C. Kessel, “Fuzzy clustering with a fuzzy covari- no. 4, pp. 1366–1375, Nov. 2007.
ance matrix,” in Proc. IEEE Control and Decision Conf., San Diego, CA, [56] P. Grossberger and I. Procaccia, “Measuring the strangeness of strange
1979, pp. 761–766. attractors,” Physica D: Nonlinear Phenomena, vol. 9, no. 1/2, pp. 189–
[28] R. Babus̆ka, Fuzzy Modeling for Control, International Series in Intelli- 208, Oct. 1983.
gent Technologies. London, U.K.: Kluwer, 1998. [57] D. A. Russell, J. D. Hanson, and E. Ott, “Dimension of strange attractors,”
[29] B. Brandstätter and U. Baumgartner, “Particle swarm optimization— Phys. Rev. Lett., vol. 45, no. 14, pp. 1175–1180, Oct. 1980.
Mass-spring systems analogon,” IEEE Trans. Magn., vol. 38, no. 2, [58] J. C. Bezdek, Pattern Recognition With Fuzzy Objective Function Algo-
pp. 997–1000, Mar. 2002. rithms. New York: Plenum, 1981.
[30] R. C. Eberhart and J. Kennedy, “A new optimizer using particle swarm [59] I. Gath and A. B. Geva, “Unsupervised optimal fuzzy clustering,”
theory,” in Proc. 6th Int. Symp. Micro Mach. and Human Sci., Nagoya, IEEE Trans. Pattern Anal. Mach. Intell., vol. 11, no. 7, pp. 773–781,
Japan, 1995, pp. 39–43. Jul. 1989.
[31] J. Kennedy and R. C. Eberhart, “Particle swarm optimization,” in Proc. [60] J. Yen and R. Langari, Fuzzy Logic: Intelligence, Control, and Informa-
IEEE Int. Conf. Neural Netw. IV, Perth, Australia, 1995, pp. 1942–1948. tion. Upper Saddle River, NJ: Prentice-Hall, 1999.
[32] D. E. Goldberg, Genetic Algorithms in Search, Optimization, and Machine [61] L. Ren, G. W. Irwin, and D. Flynn, “Nonlinear identification of turbo-
Learning. Reading, MA: Addison-Wesley, 1989. generator AVR loop dynamics using fuzzy clustering,” in Proc. Int. Conf.
[33] R. A. Krohling and L. S. Coelho, “Coevolutionary particle swarm opti- Power Syst. Technol., Kumming, China, 2002, pp. 1503–1508.
mization using Gaussian distribution for solving constrained optimization [62] X. Jiang and H. Adeli, “Fuzzy clustering approach for accurate embedding
problems,” IEEE Trans. Syst., Man Cybern. B, Cybern., vol. 36, no. 6, dimension identification in chaotic time series,” Integr. Comput.-Aided
pp. 1407–1416, Dec. 2006. Eng., vol. 10, no. 3, pp. 287–302, Aug. 2003.
[34] R. A. Krohling, F. Hoffmann, and L. S. Coelho, “Co-evolutionary particle [63] J. Abonyi, Fuzzy Model Identification for Control. Boston, MA:
swarm optimization for min–max problems using Gaussian distribution,” Birkhäuser, 2003.
in Proc. Congr. Evol. Comput., Portland, OR, 2004, pp. 959–964. [64] J. D. Holliday, S. L. Rodgers, P. Willett, M.-Y. Chen, M. Mahfouf,
[35] L. S. Coelho and R. A. Krohling, “Predictive controller tuning using K. Lawson, and G. Mullier, “Clustering files of chemical structures using
modified particle swarm optimization based on Cauchy and Gaussian dis- the fuzzy k-means clustering method,” J. Chem. Inf. Comput. Sci., vol. 44,
tribution,” in Soft Computing: Methodologies and Applications, ser. Ad- no. 3, pp. 894–902, 2004.
vances in Soft Computing, F. Hoffmann, M. Köppen, F. Klawonn, and [65] U. Kaymak and M. Setnes, “Fuzzy clustering with volume prototypes
R. Roy, Eds. London, U.K.: Springer-Verlag, 2005, pp. 287–298. and adaptive cluster merging,” IEEE Trans. Fuzzy Syst., vol. 10, no. 6,
[36] Y. Shi and R. C. Eberhart, “A modified particle swarm optimiser,” in Proc. pp. 705–712, Dec. 2002.
IEEE World Congr. Comput. Intel., Evol. Comput., Anchorage, AL, 1998, [66] F. Höppner and F. Klawonn, “A contribution to convergence theory of
pp. 69–73. fuzzy c-means and derivatives,” IEEE Trans. Fuzzy Syst., vol. 11, no. 5,
[37] Y. Shi and R. C. Eberhart, “Empirical study of particle swarm pp. 682–694, Oct. 2003.
optimization,” in Proc. Congr. Evol. Comput., Washington, DC, 1999, [67] H.-L. Jin and M. Zacksenhouse, “Oscillatory neural networks for robotic
pp. 1945–1950. yo-yo control,” IEEE Trans. Neural Netw., vol. 14, no. 2, pp. 317–325,
[38] Y. Shi and R. C. Eberhart, “Parameter selection in particle swarm op- Mar. 2003.
timizer,” in Proc. 7th Annu. Conf. Evol. Programming, V. W. Porto, [68] Z. Bein, Y.-J. Lee, S.-H. Lee, K.-H. Shim, and S.-W. Bang, “A new bench-
N. Saravan, and D. Waagen, A. E. Eiben, Eds., 1998, pp. 591–601. mark system for evaluation of intelligent controllers: An yo-yo system,”
[39] Y. Shi and R. C. Eberhart, “Fuzzy adaptive particle swarm optimization,” in Proc. Int. Joint Conf. 4th IEEE Int. Conf. Fuzzy Syst. and the 2nd Int.
in Proc. Congr. Evol. Comput., Honolulu, HI, 2002, vol. 1, pp. 101–106. Fuzzy Eng. Symp., Yokohama, Japan, 1995, vol. 3, pp. 1361–1366.
[40] S. H. Strogatz, Nonlinear Dynamics and Chaos with Applications to [69] L. Zlajpah and B. Nemec, “Control strategy for robotic yo-yo,” in Proc.
Physics, Biology, Chemistry, and Engineering. Cambridge, MA: Perseus IEEE/RSJ Int. Conf. Intell. Robots and Syst., Las Vegas, NV, 2003,
Publishing, 2000. pp. 767–772.
[41] H. Kantz and T. Schreiber, Nonlinear Time Series Analysis. Cambridge, [70] K. Hashimoto and T. Noritsugu, “Modeling and control of robotic
U.K.: Cambridge Univ. Press, 2000. yoyo with visual feedback,” in Proc. IEEE Int. Conf. Robot. Autom.,
[42] H.-O. Peitgen, H. Jürgens, and D. Saupe, Chaos and Fractals: New Fron- Minneapolis, MN, 1996, vol. 3, pp. 2650–2655.
tiers of Science, 2nd ed. New York: Springer-Verlag, 2004. [71] S.-H. Lee, K.-H. Shim, and Z. Bien, “A neuro-fuzzy controller for yo-yo
[43] B. Li and W. Jiang, “Optimizing complex functions by chaos search,” system,” in Proc. Biennial Conf. the North Amer. Fuzzy Inf. Process. Soc.,
Cybern. Syst., vol. 29, no. 4, pp. 409–419, Jun. 1998. 1996, pp. 215–218.
[44] M. Clerc and J. Kennedy, “The particle swarm: Explosion stability and [72] B. M. Herrera and L. V. Ribas, “Artificial intelligence approaches applied
convergence in a multi-dimensional complex space,” IEEE Trans. Evol. to the control of yo-yo motion system,” in Project of Undergraduate
Comput., vol. 6, no. 1, pp. 58–73, Feb. 2002. Course in Computer Engineering. Curitiba, Brazil: Pontifical Catholic
[45] I. C. Trelea, “The particle swarm optimization algorithm,” Inf. Process. Univ. Paraná, PUCPR, 2003. (in Portuguese).
Lett., vol. 85, no. 6, pp. 317–325, Mar. 2003. [73] B. M. Herrera, L. V. Ribas, and L. S. Coelho, “Nonlinear identification
[46] T. S. Parker and L. O. Chua, Practical Numerical Algorithms for Chaotic method of a yo-yo system using fuzzy model and fast particle swarm
System. Berlin, Germany: Springer-Verlag, 1989. optimization,” in Applied Soft Computing Technologies: The Challenge of
[47] G. M. Zaslavskii, “The simplest case of a strange attractor,” Phys. Lett. A, Complexity, A. Abraham, B. De Baets, M. Köppen, and B. Nickolay, Eds.
vol. 69, no. 3, pp. 145–147, Dec. 1978. London, U. K.: Springer-Verlag, 2006, pp. 302–316.
[48] R. Caponetto, L. Fortuna, S. Fazzino, and M. G. Xibilia, “Chaotic se- [74] B. Schaible, H. Xie, and Y. C. Lee, “Fuzzy logic models for ranking
quences to improve the performance of evolutionary algorithms,” IEEE process effects,” IEEE Trans. Fuzzy Syst., vol. 5, no. 4, pp. 545–556,
Trans. Evol. Comput., vol. 7, no. 3, pp. 289–304, Jun. 2003. Nov. 1997.
[49] G. Zilong, W. Sun’an, and Z. Jian, “A novel immune evolutionary algo- [75] J. Chuanwen and E. Bompard, “A self-adaptive chaotic particle swarm
rithm incorporating chaos optimization,” Pattern Recognit. Lett., vol. 27, algorithm for short term hydroelectric system scheduling in deregulated
no. 1, pp. 2–8, Jan. 2006. environment,” Energy Convers. Manage., vol. 46, no. 17, pp. 2689–2696,
[50] L. Wang and K. Smith, “On chaotic simulated annealing,” IEEE Trans. Oct. 2005.
Neural Netw., vol. 9, no. 4, pp. 716–718, Jul. 1998. [76] B. Liu, L. Wang, Y.-H. Jin, F. Tang, and D.-X. Huang, “Improved parti-
[51] M. S. Tavazoei and M. Haeri, “Comparison of different one-dimensional cle swarm optimization combined with chaos,” Chaos Solitons Fractals,
maps as chaotic search pattern in chaos optimization,” Appl. Math. Com- vol. 25, no. 5, pp. 1261–1271, Sep. 2005.
put., vol. 187, no. 2, pp. 1076–1085, Apr. 2007. [77] J. Chuanwen and E. Bompard, “A hybrid method of chaotic particle
[52] J. Mingjun and T. Huanwen, “Application of chaos in simulated anneal- swarm optimization and linear interior for reactive power optimization,”
ing,” Chaos Solitons Fractals, vol. 21, no. 4, pp. 933–941, Aug. 2004. Math. Comput. Simul., vol. 68, no. 1, pp. 57–65, Feb. 2005.
DOS SANTOS COELHO AND HERRERA: FUZZY IDENTIFICATION BASED ON A CHAOTIC PSO APPROACH 3245

Leandro dos Santos Coelho received the B.S. de- Bruno Meirelles Herrera received the B.S. in com-
gree in computer science and the B.S. degree in puter engineering from the Pontifical Catholic Uni-
electrical engineering from the Federal University of versity of Paraná, Curitiba-PR, Brazil, in 2004. He
Santa Maria, Santa Maria, Brazil, in 1994 and 2000, is currently working toward the M.S. degree in in-
respectively. He earned the M.S. and Ph.D. degrees dustrial and systems engineering at the Pontifical
in computer science and electrical engineering from Catholic University of Paraná, Curitiba-PR, Brazil.
the Federal University of Santa Catarina, Florianópo- His research interests include hardware design,
lis, Brazil, in 1997 and 2000, respectively. optimization, and quantum computation.
He is currently an Associate Professor of industrial
and systems engineering with the Pontifical Catholic
University of Paraná, Curitiba-PR, Brazil. He has
published several papers on computational intelligence applied to control
systems and engineering and is a Coauthor with M. Jamshidi, R. A. Krohling,
and P. Fleming of the book Robust Control Systems with Genetic Algorithms
(CRC Press, 2002). His research interests are electrical power systems, compu-
tational intelligence, nonlinear identification, optimization methods, quantum
computation, chaos theory, and advanced control systems.
Dr. dos Santos Coelho serves as a Reviewer for the IEEE TRANSACTIONS
ON I NDUSTRIAL E LECTRONICS ; IEEE T RANSACTIONS ON S YSTEMS ,
MAN, AND CYBERNETICS (PARTS B AND C); IEEE TRANSACTIONS ON
EVOLUTIONARY COMPUTATION; Applied Soft Computing; International Jour-
nal of Electrical Power and Energy Systems; IEE Proceedings—Generation,
Transmission, and Distribution; Engineering Applications of Artificial Intelli-
gence; Instrumentation, Systems, and Automation Transactions; and Interna-
tional Journal of Production Research.