ISA Transactions 47 (2008) 32–44

www.elsevier.com/locate/isatrans

Fuzzy rule base design using tabu search algorithm for nonlinear
system modeling
Aytekin Bagis ∗
Department of Electrical Electronic Engineering, Erciyes University, 38039 Kayseri, Turkey

Received 15 February 2007; received in revised form 29 August 2007; accepted 4 September 2007
Available online 22 October 2007

Abstract

This paper presents an approach to fuzzy rule base design using tabu search algorithm (TSA) for nonlinear system modeling. TSA is used to
evolve the structure and the parameter of fuzzy rule base. The use of the TSA, in conjunction with a systematic neighbourhood structure for the
determination of fuzzy rule base parameters, leads to a significant improvement in the performance of the model. To demonstrate the effectiveness
of the presented method, several numerical examples given in the literature are examined. The results obtained by means of the identified fuzzy
rule bases are compared with those belonging to other modeling approaches in the literature. The simulation results indicate that the method based
on the use of a TSA performs an important and very effective modeling procedure in fuzzy rule base design in the modeling of the nonlinear or
complex systems.
c 2007, ISA. Published by Elsevier Ltd. All rights reserved.

Keywords: Fuzzy systems; Fuzzy modeling; Fuzzy rule base design; Nonlinear system modeling; Tabu search

1. Introduction a fast and efficient solution to complex or nonlinear system

Modeling is an essential part of control design as it provides
a way of formally proving properties of the closed loop system.
There are many reasons why a model of a process may be
required. It could, for example, be used to simulate the real
process or it could be used to design a controller. If a good
model of the system can be obtained by using knowledge and
observations, accurate and efficient solutions to engineering
problems such as system design and simulation, process
analysis, reliability of the measured data, optimization of the
systems can be provided. Conventional system theory relies
on crisp mathematical models of systems, such as algebraic
and differential equations. However, many real world systems
are inherently nonlinear and cannot be represented by linear
models used in conventional system identification. For a large
number of practical problems, the gathering of an acceptable
degree of knowledge needed for physical modeling is a difficult,
time consuming and expensive or even impossible task. Expert
system techniques based on artificial intelligence can provide
∗ Tel.: +90 (0.352) 437 49 37x32255; fax: +90 (0.352) 437 57 84.
E-mail address: bagis@erciyes.edu.tr.
modeling problems. One of the most successful expert system
methods is fuzzy logic based modeling approach [1].
Rule based fuzzy modeling is a powerful technique for
the modeling of partly known nonlinear systems. In the
modeling of complicated and/or ill-defined processes, precise
mathematical models may fail to give satisfactory results. In
such cases, fuzzy models can effectively integrate information
from different sources (such as physical laws, empirical models,
and experimental measurements) and they may be used to
properly depict the system uncertainty. If fuzzy rule based
models can be identified, fuzzy controllers can be obtained from
the fuzzy models and genuine control engineering issues can
be addressed for fuzzy controllers, namely stability, robustness
etc. [2].
In designing fuzzy models, a major difficulty in the
identification of an optimized fuzzy rule base is encountered.
Traditional fuzzy rule design is based upon a human operator’s
experienced knowledge. Not only is this design method time
consuming since it uses trial and error to find good fuzzy
rules and membership functions, it is also not guaranteed
to find optimal or near optimal fuzzy rules and membership
functions [3,4].

c 2007, ISA. Published by Elsevier Ltd. All rights reserved.

0019-0578/$ - see front matter

doi:10.1016/j.isatra.2007.09.001
 A. Bagis / ISA Transactions 47 (2008) 32–44
There are many studies on fuzzy modeling in the literature.
Takagi and Sugeno proposed a new type of fuzzy model
called “Takagi–Sugeno fuzzy model (T–S fuzzy model)” [5].
Furthermore, this study proposes a procedure to identify the
T–S fuzzy model from input–output data of systems. Sugeno
and Yasukawa discussed a general approach to qualitative
modeling based on fuzzy logic [6]. It proposes the use of a
fuzzy clustering method to identify the structure of a fuzzy
model. Wang and Lee proposed a method based on the
back-propagation algorithm to determine significant variables
without the need of human inspection [7]. A novel fuzzy
modeling algorithm was presented by Kim et al. [8]. In the
study, fuzzy C regression model clustering algorithm and a
gradient descent algorithm are used. Dynamic versions of
fuzzy logic systems and their nonsingleton generalizations
were investigated by Mouzouris and Mendel [9]. They used
a dynamic learning algorithm to train the system parameters.
Wang and Langari presented a fuzzy logic approach to
complex system modeling that is based on fuzzy discretization
technique [10]. A great advantage of the presented approach
was that it could determine the premise structure of a model
off-line. In [11], a method of Takagi–Sugeno–Kang fuzzy
model generation using numerical data as a starting point was
described by Kukolj.
In addition, many other methods such as neural networks
based methods were suggested and were used to build fuzzy
models [12–15]. Several models for the identification and
control of nonlinear dynamic systems were suggested by
Narendra and Parthasarathy [13]. These models, which include
multilayer neural networks as well as linear dynamics, can be
viewed as generalized neural networks. For the adjustment of
parameters, static and dynamic back-propagation methods are
used. Jang described an adaptive network based fuzzy inference
system [12]. The presented method employed a hybrid learning
procedure to construct an input–output mapping based on
both human knowledge and stipulated input–output data
pairs. By using the architecture of the fuzzy-neural network,
simple and effective fuzzy rule based models of complex
systems from input–output data were developed by Lin and
Cunningham [15]. Babuska and Verbruggen presented an
overview of neuro-fuzzy modeling methods for nonlinear
system identification [14].
Different optimization algorithms were also used for
fuzzy system modeling [16–23]. Hanss presented a special
fuzzy modeling method for developing multivariable fuzzy
models [20]. This fuzzy model identification procedure was
carried out by applying a special clustering method. Kilic
et al. proposed a fuzzy system modeling approach based on
a probabilistic hill-climbing algorithm as a data analysis and
approximate reasoning tool [19]. Evsukoff et al. used a least
squares error minimization method for nonlinear fuzzy model
identification [16]. A genetic based neuro-fuzzy approach for
modeling and control of dynamical systems was presented by
Farag et al. [23]. The proposed approach combined the merits of
the fuzzy logic theory, neural networks, and genetic algorithms.
Wu and Yu used a genetic algorithm and gradient descent based
approach to optimize the structure of a Takagi–Sugeno–Kang
33
model [17]. An evolutionary design method of fuzzy rule base
for nonlinear system modeling and control was presented by
Kang et al. [22]. Cordon et al. presented an overview of the
representative methods in genetic fuzzy systems proposed in
the literature [21].
This paper presents an approach to fuzzy rule base design
using a tabu search algorithm (TSA) for nonlinear system
modeling. TSA is used to evolve the structure and the
parameters of fuzzy rule base. To check the effectiveness of the
presented approach, several numerical examples given in the
literature are examined. The results obtained by means of the
identified fuzzy rule bases are compared with those belonging
to other modeling approaches in the literature. The remainder
of the paper is organized as follows. In Section 2, a brief review
of the TSA is introduced. Structure of the identified fuzzy rule
base and design method is developed in Section 3. Simulation
results and some conclusions are given in Sections 4 and 5,
respectively.
2. Tabu search algorithm (TSA)
TSA was proposed by Glover as an intelligent optimization
technique to overcome local optimality in solution processes for
hard combinatorial optimization problems [24]. This algorithm
consists of the systematic prohibition of some solutions to
prevent cycling and to avoid the risk of trapping at a local
optimum. New solutions are searched in the neighbourhood of
the current one. In TSA, there are several important concepts:
Neighbourhood: To improve the quality of the solutions visited,
the search moves from one solution to another using a
neighbourhood structure. The neighbourhood of a solution is
the set of all formations that can be arrived at by a move.
The move is a process that transforms the current solution to
its neighbouring solution. If, after making a move, a solution
is found which is better than all solutions found in prior
moves than the new solution is saved as the new best solution.
This neighbourhood procedure is applied to the only one best
solution of each generation. Considering a chromosome with n
bits, the number of the neighbour solutions is n, as depicted in
Fig. 1. To obtain the neighbour solutions for the best solution,
encoding and decoding operations are applied. In this work,
using different bit sizes for each of the parameters we have tried
to achieve a good balance between the algorithm performance
and the computational cost, and we finally adopt a bit size
of 8. Therefore, total number of bits required for defining a
fuzzy rule base parameter set is [rule number × parameter
number × bit number]. According to simulation results, this bit
size is suitable for the main process in the algorithm described
in this paper.
Tabu list: One of the main ideas of TSA is the use of a
flexible memory (tabu list). The objective of the tabu list is to
exclude moves which would bring the algorithm back to where
it was at some previous iteration and keep it trapped in a local
minimum. To avoid performing a move returning to a recently
visited region, the reverses of last moves are forbidden. In other
words, these moves are considered as “tabu” and are recorded
in the tabu list. The tabu list stores all tabu moves that are
34 A. Bagis / ISA Transactions 47 (2008) 32–44
Fig. 1. Neighbourhood structure for a solution.
Table 1
The values of the control parameters for TSA
Tabu search algorithm
Recency factor (r ) 0.2
Frequency factor ( f ) 2.0
Generation number 2000
Tabu conditions (1) Recency (k) > (r × M),
(2) Frequency (k) < ( f × avgfreq)
Number of variables for a rule 9, for system with 2 inputs and 1 output
13, for system with 3 inputs and 1 output
Number of neighbourhood per 8 frequently among them but it is still classified as tabu. In this
variable
Tabu list size Rule number × number of the variables
not permitted to be applied to the present solution. This list is
initialized empty, constructed in consecutive iterations of the
search and updated circularly in later iterations.
Tabu conditions: The number of the possible moves is
determined by the tabu list. If a solution produced by a move
that is not on the tabu list is better than all solutions found in
prior iterations, then this solution is saved as the best solution.
At each of the iterations in the optimization process, candidate
solutions are checked with respect to “tabu conditions”, and
hence, the next solution is determined depending on evaluation
values and tabu conditions. The re-generation of a solution
previously obtained is avoided by using the tabu conditions
(or restrictions) on the possible moves. The tabu conditions
are usually based on two important factors: frequency memory
and recency memory. Frequency memory keeps the knowledge
of how often the same solutions have been made in the past.
Recency memory prevents cycles of length less than or equal to
a predetermined number of iterations. The tabu conditions used
in this work are as the following: If the element k of a solution
vector does not satisfy the conditions (i) recency (k) > r.M,
(ii) frequency (k) < f.avgfreq, then it is accepted as tabu and
not used to create a neighbour. Here, r and f are recency and
frequency factors, M is the number of elements in the solution
vector, and avgfreq is the average frequency value. In this work,
after some experimentation, the algorithm parameters r and f
have been set to the values of 0.2 and 2.0, respectively. The
control parameter values for the TS algorithm are given in
Table 1. For all of the simulations, the iteration or generation
number is taken as 2000. The number of bits per variable is 8.
Thus, if a rule has the n parameters, the total number of bits in
the string describing each member of the solutions is (n × 8).
Fig. 2. System modeling scheme.
Aspiration criterion: TS has another important element called
aspiration mechanism. If a move on the tabu list leads to a
solution with a value strictly better than the best obtained so
far, it is possible to allow this move to get out of the tabu list.
This property is used to avoid removing very good moves from
consideration and plays an essential role in the search process.
In this work, if all available moves are classified as tabu, then a
least tabu move is selected for a new solution. The least tabu
move means that this solution is changed less recently and
work, maximum number of iterations reached is selected as the
stopping criterion of the search process.
The basic procedure of TSA to be utilized in this paper is
described as follows:
Initial solution;
While predetermined stopping criteria not satisfied;
{
Create a set of neighbour solutions;
Evaluate the neighbour solutions;
Choose the best admissible solution;
Test the tabu conditions;
Perform the aspiration test;
Update tabu list;
}
This procedure is repeated until a satisfactory solution is
obtained or predetermined number of generations is elapsed.
For all of the modeling examples, the number of the generation
was selected as 2000.
3. Structure of the identified fuzzy rule base
Fig. 2 shows the system modeling scheme. In this figure,
y p (t + 1) is the output of the system to be modeled, and
ym (t + 1) is the output of the fuzzy model optimized by the
TSA. Input of the TSA is the error value, and it is calculated
by {error = y p (t + 1) − ym (t + 1)}. The major objective of
the algorithm is to determine the most appropriate fuzzy model
parameters and to reduce the error factor.
In fuzzy models, structure selection involves the following
choices:
(a) input and output variables
(b) structure of the rules
 A. Bagis / ISA Transactions 47 (2008) 32–44
Fig. 3. (a) Parameter matrix representing the membership functions, (b) Fuzzy rule base represented by parameter matrices, (c) The rule base coding.
(c) number and type of membership functions for each variable
(d) type of the inference mechanism, connective operators,
defuzzification method.
After the structure is fixed, the performance of a fuzzy
model can be fine-tuned by adjusting its parameters. Tunable
parameters of linguistic models are the antecedent parameters
and consequent membership functions and the rules. Rules
in the fuzzy rule base are described by the membership
functions. Fuzzy membership functions can have different
shapes depending on the designer’s preference or experience.
In general, fuzzy engineers have found that triangular and
trapezoidal shapes help capture the modeller’s sense of
fuzzy numbers and simplify computation. In this study, input
variables are characterized by triangular membership functions,
and the output variables are characterized by fuzzy singletons.
A rule in the rule base has a general form expressed by
Ri : If x1 is Ai1 (x1 ) and x2 is Ai2 (x2 ) then y is Bi
where Ri is the ith rule, xj is the input variable, and y is the
output variable. Aij (xj ) is the fuzzy linguistic value defined as
in Eq. (2) and Bi takes the role of fuzzy singleton, i.e., real
number.
Aij = max(min[(xj − a)/(b − a), (c − xj )/(c − b)], 0).
In Eq. (2); a, b, and c are the main parameters of a triangular
membership function. These parameters represent the locations
of starting point, peak point, and the ending point for a triangle
shaped membership function, respectively.
Deciding the number of fuzzy rules is a very critical
issue since it plays key roles in the fuzzy modeling of the
systems. Except for some propositions, unfortunately, there is
no systematic and efficient procedure for choosing the most
35
appropriate rule number. A reasonable number of fuzzy rules
without losing too much information about the system to
be modeled must be carefully obtained. In this work, using
different rule sizes we have tried to achieve a good balance
between the model performance and rule complexity.
Another important element in combining a TSA and a
fuzzy model is the encoding method used to represent a
fuzzy rule base as a probable solution element on which the
TSA operates. Selecting an appropriate encoding procedure
is a critical factor that affects the efficiency of the TSA and
the final solution. The membership functions of the input
variables are characterized by three numerical values. On
the other hand, the output variables are defined by singleton
numerical values. These numerical values represent the position
of the membership function on the universe of discourse. The
parameter matrix defining the membership functions and fuzzy
rule base represented by the parameter matrices is formulated
(1) in Fig. 3. The parameter matrix that contains the parameters
for defining the membership functions within the fuzzy rule
base will be of one-dimensional matrix form. For each rule
in the fuzzy rule base, the adjustment parameters used in the
reasoning procedure are also involved in this parameter matrix.
In Fig. 3, these parameters are stated by di for each input
variable. Thus, the size of matrix is determined by {the number
(2)
of rules × [(the number of input variables × 3) + (the number
of output variables × 1) + (the number of input variables)]}.
This value of the matrix is also equal to the size of the tabu
list. For a fuzzy model with two inputs and one output, each
fuzzy rule is represented by nine parameters. The rule structure
coding is organized as shown in Fig. 3(c). For example, if the
modeling system has two inputs and one output, 144 bits are
used to represent a probable solution with two rules optimized
by the algorithm (Fig. 3(c)). Initializing the solution is based
36 A. Bagis / ISA Transactions 47 (2008) 32–44

Table 2
Optimized parameter matrix for fuzzy rule base

Rules Input 1, u(t − 4) Input 2, y(t − 1) Output, y(t) Adjustment parameters

ai1 ai2 ai3 bi1 bi2 bi3 ci di1 di2
Normalized parameters
(1) 0.1579 0.1720 0.4845 0.1485 0.3595 1.0196 0.7658 0.4863 0.2813
(2) 0.0552 0.3294 2.5020 0.0000 0.2667 0.8046 0.0156 0.1489 0.5020
(3) 0.0000 0.1569 0.8908 0.3476 1.8823 2.7922 0.9412 0.4706 1.1214
(4) 0.5640 0.7529 1.2549 0.2813 0.4072 0.8314 0.2500 0.2510 0.0784
Denormalized parameters
(1) −2.0526 −1.9680 −0.0930 46.6730 50.4710 62.3528 57.7844 0.4863 0.2813
(2) −2.6688 −1.0236 12.0120 44.0000 48.8006 58.4828 44.2808 0.1489 0.5020
(3) −3.0000 −2.0586 2.3448 50.2568 77.8814 94.2596 60.9416 0.4706 1.1214
(4) 0.3840 1.5174 4.5294 49.0634 51.3296 58.9652 48.5000 0.2510 0.0784

on the randomly generated values from the given searching 4. Simulation results
intervals.
In this study, the following reasoning procedures are Five numerical examples are provided in this section to
considered: illustrate the performance of the presented approach. All these
(a) Given input data x1 and x2 , calculate the degree of the examples have their origin in the existing literature. The
fulfillment ωi in the premise for the ith rule as in datasets used in the first and third examples are derived from
practical processes; while the dataset used in the second and
ωi = Ai1 (x1k ).di1 + Ai2 (x2k ).di2 , fourth examples is generated from the simulation of analytical
k = input data number. (3) models. In the first example, the fuzzy modeling of a dynamical
(b) Calculate the inferred value yi by taking the weighted process using a famous example of the system identification
average of Bi with respect to ωi as in given by Box and Jenkins [25] is discussed. The third example
is used to show how to build a model of a human operator’s
r
P
ωi .Bi control action in a chemical plant. The structure of the human
i=1 operation is shown in Sugeno and Yasukawa [6]. Furthermore,
yi = r
(4) input–output data for human operation at the chemical process
ωi
P
are also listed in [17]. In the fifth example, modeling of a real
i=1
reservoir system is considered [3,4]. And, the datasets used in
where r is the number of fuzzy rules. this example are real flood hydrograph data of the dynamical
To evaluate the performance of a fuzzy model, there are two system.
important performance indices in the literature. These are the (a) Box and Jenkins’ gas furnace: The fuzzy model of a
mean squared error (MSE) and a performance parameter stated dynamic process using a well-known example of the system
as PI. In the MSE, the error is the difference between the actual identification given by Box and Jenkins is considered [6,22,25].
output and the estimated output by the fuzzy model. The MSE The process is a gas furnace with a single input u(t) and a single
and PI are calculated from N data point as output y(t): gas flow rate and CO2 concentration, respectively.
N The dataset consists of N = 296 pairs of input–output
1 X measurements. The inputs to the fuzzy model were selected as
MSE = (O d − Ok )2 (5a)
N k=1 k u(t −4), y(t −1), respectively. The model output is y(t). For the
s variables of u(t − 4), y(t − 1), and y(t), normalization intervals
N
P
(Okd − Ok )2 were selected as [−3, 3], [44, 62], and [44, 62], respectively.
k=1 Parameters of the fuzzy model were searched into the interval
PI = (5b) of [0, 4]. During the evolution, the modeling error of the best
N  
Od 
P
k individual at each generation was calculated by Eq. (5a). The
k=1
where Okd , k = 1, . . . , N are the actual or desired output values
and Ok , k = 1, . . . , N are the outputs from the model.
These performance indices provide a means for evaluating
the performance of a fuzzy model with the selected fuzzy
rule base in the process of evolution, so that an optimized
model would be developed by the best individual. During the
evolution, the quality of the best solution at each generation is
calculated by Eq. (5a) or (5b).
optimized fuzzy model for a gas furnace and the comparison of
model outputs and the original output data are shown in Figs. 4
and 5, and the modeling error is 0.1483. The number of rules
in the fuzzy model is only four. Normalized and denormalized
parameter matrices of the fuzzy model are also given in Table 2.
In Table 3, our fuzzy model is compared with other models
identified from the same data.
(b) Modeling of a static nonlinear function: In this subsection,
double input and single output static function was chosen as
 A. Bagis / ISA Transactions 47 (2008) 32–44 37

Fig. 4. Optimized fuzzy rule base for gas furnace.

Table 3
Comparison of our model with other models

Method Rule number MSE

Box and Jenkins (1970) [25] – 0.202
Tong (1978) [26] 19 0.469
Pedrycz (1984) [27] 81 0.320
Takagi and Sugeno (1985) [5] 6 0.190
Xu and Lu (1987) [28] 25 0.328
Sugeno and Tanaka (1991) [29] 2 0.068
Sugeno and Tanaka (1991) [29] 2 0.359
Wang and Langari (1996) [10] 5 0.158
Nakoula et al. (1997) [30] 90 0.175
Kim et al. (1997) [8] 2 0.055
Farag et al. (1998) [23] 37 0.111
Kang et al. (2000) [22] 5 0.161
Evsukoff et al. (2002) [16] 36 0.153
Evsukoff et al. (2002) [16] 90 0.090
Fig. 5. Comparison of model outputs and original output data for the gas Our model 4 0.148
furnace.

five. Normalized and denormalized parameter matrices of the

a target system. The function was taken from Narendra and
Parthasarathy [13] and represented as
y = (1 + x1−2 + x2−1.5 )2 , 1 ≤ x1 , x2 ≤ 5
from which 50 data was obtained. From this system equation,
50 input–output data are obtained. In this instance, two random
inputs x1 and x2 were inserted to test input identification.
Normalization intervals for input–output parameters x1 , x2 , y
were selected as [0, 6], [0, 6], [0, 10], respectively. Model
parameters were searched in the interval of [0, 4]. During the
evolution, the modeling error at each generation was calculated
by Eq. (5a). The optimized fuzzy model for nonlinear system
and the comparison of model outputs and the original output
data are shown in Figs. 6 and 7, respectively, and the modeling
error is 0.0019. The rule number of the fuzzy model is
fuzzy model are given in Table 4. Performance indices of the
presented method and the other methods in the literature are
(6) given in Table 5. To test the validity of the derived model,
the later 20 data that are not used in the model derivation are
considered by the model, where MSE calculated by (5a) is
0.0043 (Fig. 7(b)).
(c) Human operation at a chemical plant: This example
deals with the modeling of human operator control in a
chemical plant. The plant is for producing a polymer by the
polymerization of monomers. Since the start-up of the plant is
very complicated, a human operator is used to manually operate
the system in this circumstance. The structure of the human
operation is given by Sugeno and Yasukawa [6]. Five input
candidates, u 1 : the monomer concentration, u 2 : the change of
monomer concentration, u 3 : the monomer flow rate, u 4 and u 5 :
38 A. Bagis / ISA Transactions 47 (2008) 32–44

Fig. 6. Optimized fuzzy rule base for a static nonlinear function.

Fig. 7. (a) Comparison of model outputs and original output data for example (b). (b) Test outputs and model outputs for example (b).

the local temperature inside the plant are available, to which
the human operator may refer. The output y is the set point for
monomer flow rate. There are 70 data points for each of the
above six variables (five inputs, and one output).
There are different fuzzy models available for modeling this
plant. In [6], a six rule qualitative fuzzy model was given by
identifying u 1 , u 2 , and u 3 as input variables. In [10], u 1 , u 2 ,
and u 3 were selected as the input variables, and six clusters
from the given data were taken to give six rules and three
premise variables of the Takagi–Sugeno–Kang model. In [15],
a fuzzy-neural system was used for modeling the plant. The
performance measurement was not stated in [6] and [10].
In our approach, the input variables u 1 and u 3 were selected
to build the fuzzy model. Normalization intervals for all
variables (u 1 , u 3 , y) were selected as [4, 7], [0, 8000], and
[0, 8000], respectively. Model parameters were searched in the
solution area of [0, 2]. During the evolution, the modeling error
at each generation was calculated by Eq. (5b). Optimized model
parameters and optimum rule base are given in Table 6 and
Fig. 8, respectively. Optimum rule base has only three rules.
The comparison of model outputs and the original output data
are shown in Fig. 9. In Table 7, our fuzzy model is compared
with the other models in the literature.
 A. Bagis / ISA Transactions 47 (2008) 32–44 39

Table 4
Optimized parameter matrix for fuzzy rule base

Rules Input 1, x1 Input 2, x2 Output, y Adjustment parameters

ai1 ai2 ai3 bi1 bi2 bi3 ci di1 di2
Normalized parameters
(1) 0.5348 0.7867 0.7887 0.3438 0.7735 0.8282 0.1730 0.0625 1.4922
(2) 0.2285 0.2286 0.5411 0.0000 0.0728 0.3282 1.0000 0.1279 0.2500
(3) 0.1661 0.2247 3.1662 0.1641 0.3141 0.6661 0.2328 0.2731 1.0490
(4) 0.1661 0.2516 0.2657 0.0938 0.5781 1.9137 0.0313 0.3516 0.5742
(5) 0.0938 0.0956 0.2657 0.2130 0.3439 0.3594 0.9705 1.7887 0.0509
Denormalized parameters
(1) 3.2088 4.7202 4.7322 2.0628 4.6410 4.9692 1.7300 0.0625 1.4922
(2) 1.3710 1.3716 3.2466 0.0000 0.4368 1.9692 10.000 0.1279 0.2500
(3) 0.9966 1.3482 18.9972 0.9846 1.8846 3.9966 2.3280 0.2731 1.0490
(4) 0.9966 1.5096 1.5942 0.5628 3.4686 11.4822 0.3130 0.3516 0.5742
(5) 0.5628 0.5736 1.5942 1.2780 2.0634 2.1564 9.7050 1.7887 0.0509

Fig. 8. Optimized fuzzy rule base for human operation at a chemical plant.

Table 5
Comparison of different models for example (b)

Method Rule number MSE

Sugeno and Yasukawa (1993) [6] 6 0.0790
Lin and Cunningham (1995) [15] 3 0.0035
Kim et al. (1997) [8] 3 0.0197
Our model 5 0.0019

(d) Nonlinear differential equation: This example is taken

from Narendra and Parthasarathy [13] in which the plant to
be identified is given by the second-order highly nonlinear
difference equation

y(k − 1).y(k − 2).(y(k − 1) + 2.5)

y(k) = + u(k).
1 + y 2 (k − 1) + y 2 (k − 2)
Training data of 500 points are generated from the plant
model, assuming a random input signal u(k) uniformly
distributed in the interval [−2, 2]. This unpredictable input
“u(k)” is randomly generated and is inserted into the system.
The first 500 data points are produced from (7) for model
identification. To demonstrate the performance of the fuzzy
(7)
Fig. 9. Comparison of model outputs and original output data for the chemical
plant.
model, a different type sinusoidal input data u(k) is inserted
into the system y(k) (Fig. 10).
The model has three inputs u(k), y(k − 1), and y(k − 2),
and a single output y(k). Normalization intervals for the inputs
40 A. Bagis / ISA Transactions 47 (2008) 32–44

Table 6
Optimized parameter matrix for fuzzy rule base

Rules Input 1, u1 Input 2, u3 Output, y Adjustment parameters

ai1 ai2 ai3 bi1 bi2 bi3 ci di1 di2
Normalized parameters
(1) 0.0000 0.3130 1.1006 0.5938 1.4525 1.4642 0.8827 0.2544 1.4486
(2) 0.1719 0.4840 0.7141 0.4527 0.4844 0.8908 0.4172 0.0774 0.3157
(3) 0.2537 0.3387 0.6829 0.0430 0.0489 0.5266 0.0110 0.0351 0.8387
Denormalized parameters
(1) 4.0000 4.9390 7.3018 4750.40 11 620.00 11 713.60 7061.60 0.2544 1.4486
(2) 4.5157 5.4520 6.1423 3621.60 3 875.20 7 126.40 3337.60 0.0774 0.3157
(3) 4.7611 5.0161 6.0487 344.00 391.20 4 212.80 88.00 0.0351 0.8387

Table 7
Comparison of different models for human operation at a chemical plant

Method Model Rule PI

inputs number
Sugeno and Yasukawa u1, u2, u3 6 Not stated
(1993) [6]
Lin and Cunningham u1, u2, u3 7 0.002245
(1995) [15]
Wu and Yu (2000) [17] u3, u4 4 0.002000
Our model u1, u3 3 0.001700

were selected as [−2, 2], [−5, 5], and [−5, 5], respectively.
The search interval of all the variables is [0, 4]. After the
learning process is finished, the model is tested by applying a
sinusoidal input data u(k) = sin(2πk/25) to the fuzzy model.
The optimized fuzzy model is given in Fig. 11. The rule number
of the fuzzy model is four. Normalized and denormalized Fig. 10. Train and test inputs for a nonlinear differential equation.
parameter matrices of the fuzzy model are also given in Table 8.
The output of both the fuzzy model and the actual model are Table 9
shown in Fig. 12. The fuzzy model has a good match with Comparison of different models for the nonlinear differential equation
the actual model with a MSE of 0.0378. Table 9 compares our Method Rule Learning Testing MSE
modeling approach with the other approaches in the literature. number MSE
The models are obtained from the previously generated 500 Sugeno’s model (1991, 12 0.5072 0.2447
data pairs and tested by applying a sinusoidal input signal 1993) [6,29]
u(k) = sin(2πk/25). Farag’s model (1998) [23] 75 0.0374 0.0403
Wang’s model (1999) [7] 8 0.6184 0.2037
(e) Modeling of a real reservoir system: This example includes Evsukoff’s model (2002) [16] 100 0.1577 0.0185
the modeling of the nonlinear relationship among the main Our model 4 0.0341 0.0378
variables of the reservoir in a reservoir management system.

Table 8
Optimized parameter matrix for fuzzy rule base

Rules Input 1, u(k) Input 2, y(k − 1) Input 3, y(k − 2) Output, y(k) Adjustment parameters
ai1 ai2 ai3 bi1 bi2 bi3 ci1 ci2 ci3 di ei1 ei2 ei3
Normalized parameters
(1) 0.0000 0.0001 0.8363 0.1212 0.4045 0.6214 0.0899 0.2462 1.0005 0.5226 1.3420 2.5905 1.2435
(2) 0.0000 0.0000 0.3381 0.1250 0.1760 0.5904 0.0977 0.1055 1.9072 0.1055 0.7062 0.4260 1.4667
(3) 0.3752 1.0179 1.9853 0.5814 0.9993 1.1009 0.9223 0.9231 0.9243 1.0924 1.3952 1.0081 0.4905
(4) 0.1255 0.9849 3.8359 0.4385 0.8799 3.3353 0.5178 0.6566 0.9385 1.3333 0.4279 0.6614 0.3088
Denormalized parameters
(1) −2.000 −1.999 1.345 −3.788 −0.955 1.214 −4.101 −2.538 5.005 0.2260 1.3420 2.5905 1.2435
(2) −2.000 −2.000 −0.648 −3.750 −3.240 0.904 −4.023 −3.945 14.07 −3.9450 0.7062 0.4260 1.4667
(3) −0.499 2.072 5.941 0.814 4.993 6.009 4.223 4.231 4.243 5.9240 1.3952 1.0081 0.4905
(4) −1.498 1.940 13.344 −0.615 3.799 28.35 0.178 1.566 4.385 8.3330 0.4279 0.6614 0.3088
 A. Bagis / ISA Transactions 47 (2008) 32–44
Fig. 11. Optimized fuzzy rule base for the nonlinear differential equation.
Fig. 12. Comparison of model outputs and original output data for the
nonlinear differential equation.
Reservoir management system is a control system that manages
a spillway gate in a dam to increase or decrease the amount
of discharge water [3,4,31]. Reservoir management is a
complex, nonlinear, nonstationary control process because: (a)
The hydrological conditions have a nonlinear behavior, (b)
Determination of the inflow hydrograph can be extremely
difficult, (c) Construction of a precise model of this process is
very difficult, (d) Owing to human factor, accurate and efficient
control of the process is quite difficult.
The main variables of a reservoir are reservoir lake level (H ,
m), spillway gate opening (d, m) and released water amount
(Q, m3 /s) [3,4,31]. The fuzzy model is used to model the
nonlinear relationship among these variables. Lake level (H )
and spillway gate opening (d) are the inputs of the model. The
output variable of the fuzzy model is the released water amount
(Q). In this work, the reservoir system of Adana Catalan Dam
in Turkey is considered, and the real flood hydrograph data for
41
Catalan Dam are used [31]. To generate comparable results to
those already published; the 548 real data points were used
to generate a model using the TSA based method. At the end
of the training process, the resultant model was applied to the
validation dataset, which was the dataset of 25 unseen data. All
of the above information is obtained from the file containing the
final project details of Catalan Dam [31].
Normalization intervals for input–output parameters H, d, Q
were selected as [117, 131], [0, 12], [0, 10 000], respectively.
Model parameters were searched in the interval of [0, 10]. Op-
timized model parameters and optimum rule base are given in
Table 10 and Fig. 13, respectively. The PI was calculated for
the model predictions of the validation data and the results have
been given in Table 11. The optimum rule base has only five
rules. The comparison of model outputs and the original output
data is shown in Fig. 14 as three dimensions. In Fig. 15, process
outputs for train and test data are given as two dimensions. The
results show that the fuzzy rule structure optimized by the TSA
can yield a better model than the model which has the struc-
ture of a neural network, improving the validation data model
prediction PI.
Because it did not seem to be the most commonly used
parameter in the literature to be considered, the time parameter
was not selected as a comparison parameter. The values of the
computation time were not stated mostly in the literature to be
compared. However, to indicate the computational performance
of the TSA, the computation time per evaluation of the
algorithm for each problem is given in Table 12. In the
simulations, Matlab programming package and Pentium IV
3600 MHz computer were used.
It is clear from the figures that fuzzy modeling based TSA
is a powerful tool for modeling of nonlinear systems. The
approach is easy to use and there is no difficulty in writing its
algorithm based modeling procedure.
42 A. Bagis / ISA Transactions 47 (2008) 32–44

Table 10
Optimized parameter matrix for fuzzy rule base

Rules Input 1, H Input 2, d Output, Q Adjustment parameters

ai1 ai2 ai3 bi1 bi2 bi3 ci di1 di2
Normalized parameters
(1) 0.0566 0.0661 0.6299 0.0000 0.0000 0.5368 0.0049 1.7488 2.4610
(2) 1.3333 4.0000 4.3175 0.5402 0.8887 2.8233 0.8105 0.0000 1.2308
(3) 0.1604 0.1631 1.3449 0.0059 0.5761 0.9376 0.5102 0.0293 1.1965
(4) 0.4688 0.6421 0.9029 0.7512 0.8905 0.9922 0.0000 0.2344 0.0778
(5) 0.0802 0.2444 0.4727 0.0056 0.3711 0.8422 0.4859 0.0510 1.1429
Denormalized parameters
(1) 117.79 117.93 125.82 0.0000 0.0000 6.4416 49.00 1.7488 2.4610
(2) 135.67 173.00 177.45 6.4824 10.6644 33.8796 8105 0.0000 1.2308
(3) 119.25 119.28 135.83 0.0708 6.9132 11.2512 5102 0.0293 1.1965
(4) 123.56 125.99 129.64 9.0144 10.6860 11.9064 0.00 0.2344 0.0778
(5) 118.12 120.42 123.62 0.0672 4.4532 10.1064 4859 0.0510 1.1429

Fig. 13. Optimized fuzzy rule base for a reservoir system.

Fig. 14. Comparison with three dimensions of the outputs for (a) train operation and (b) test operation in the reservoir system.
 A. Bagis / ISA Transactions 47 (2008) 32–44 43

Fig. 15. Comparison with two dimensions of the outputs for (a) train operation and (b) test operation in the reservoir system.

Table 11 for multimodal problems in a reasonable time although the

Comparison of different models for a nonlinear reservoir system
Method Model inputs PI (train)
Neural model [31] H, d 0.00170
Our model H, d 0.00158
When the model performance and rule size are considered,
the resulting outcome is quite exciting. It is clear from the tables
that a striking reduction in the value of the performance index
(MSE or PI) is obtained without the need of a large number of
rules. The low rule size is a significant advantage of the model.
The studies in the literature show that TSA is an effective
optimization method in the designing of fuzzy systems [3, 4].
By using the control factors such as the initial solution, type of
move, size of neighbourhood, tabu list size, aspiration criterion
and stopping criterion, the performance of the algorithm can be
significantly improved. And, system models with high quality
can be obtained.
Three important conclusions can be made from the details of
the numerical results:
First, the requirement for using the large number of rules is
successfully removed by employing the TSA. Tabu algorithm
enables us to obtain better results when the number of
membership functions and rules is as small as possible to solve
the problem at hand, a realistic case from an industrial point of
view. With tabu search, it is indeed possible to optimally tune
the structure of the fuzzy model.
Second, even though the presented approach does not show
the best performance for all of the problem types considered, it
obviously exhibits a competitive performance for the modeling
of the nonlinear or complex systems. It has the ability of
getting out of local minima and finding global optimal solutions
traditional optimization algorithms fail to produce global
PI (test) optimal solution for such problems. Thus, it is verified that
the presented method is capable of defining the nonlinearities
0.0056
0.0065 or uncertainties of the system conditions without requiring
any expert knowledge or experience about the system to be
considered.
Third, the numerical results show that the modeling
approach based on the TSA has high accuracy (low error)
and good generalization ability. Moreover, it can start to
learn a process from a very low number of initial data
samples and improve the performance of the initial model.
Therefore, it can be used effectively for various engineering
problems in different areas with the nonlinear behaviors. And,
accurate, fast, and reliable fuzzy models can be developed from
measured/simulated data.
When the process is highly nonlinear, time varying, or when
it depends on a lot of parameters, the cost for getting an accurate
model is often estimated as being too high. In this case, fuzzy
modeling approach based on the use of TSA can be viewed as
a useful alternative method.
5. Conclusion
In this paper, an efficient approach to fuzzy rule base
design using the tabu search algorithm for nonlinear system
modeling is presented. To demonstrate the effectiveness of the
presented approach, several numerical examples given in the
literature are examined. Simulation results show that the fuzzy
modeling approach presented here is a powerful tool for the
modeling of nonlinear systems. As compared with the other
modeling methods, the described approach has the advantages
of simplicity, flexibility, and high accuracy. The method is

Table 12
Computational performance of the TSA for different problems

Problem Data Rule Parameter number Total parameter Neighbourhood Computation time per iteration (s)
number number per rule number number
Box–Jenkins 296 4 9 36 288 9.18 (for 288 evaluation)
Static nonlinear function 50 5 9 45 360 2.37 (for 360 evaluation)
Chemical plant 70 3 9 27 216 1.27 (for 216 evaluation)
Nonlinear differential 500 4 13 52 416 31.27 (for 416 evaluation)
equation
Reservoir operation 548 5 9 45 360 24.13 (for 360 evaluation)
44 A. Bagis / ISA Transactions 47 (2008) 32–44
easy to use and there is no difficulty in writing its automatic
modeling procedure. Especially, when the model performance
and rule size are considered, the method based on the use of a
tabu search algorithm performs an important and very effective
modeling procedure to fuzzy rule base design in the modeling
of nonlinear or complex systems.
Acknowledgements
The author is grateful to the reviewers, and Prof. Dervis
Karaboga at Erciyes University in Turkey for their valuable
comments and suggestions.
References
[1] Babuska R. Fuzzy modeling for control. Kluwer Academic Pub.; 1998.
[2] Nguyen HT, Sugeno M. Fuzzy systems: Modeling and control. Kluwer
Academic Pub.; 1998.
[3] Bagis A. Fuzzy and PD controller based intelligent control of spillway
gates of dams. J Intell Fuzzy Syst 2003;14(1):25–36.
[4] Bagis A. Determining fuzzy membership functions with tabu search—An
application to control. Fuzzy Sets and Systems 2003;139:209–25.
[5] Takagi T, Sugeno M. Fuzzy identification of systems and its applications
to modeling and control. IEEE Trans Syst Man Cybern 1985;15:116–32.
[6] Sugeno M, Yasukawa T. A fuzzy logic based approach to qualitative
modeling. IEEE Trans Fuzzy Syst 1993;1(1):7–31.
[7] Wang S, Lee C. Fuzzy system modeling using linear distance rules. Fuzzy
Sets and Systems 1999;108:179–91.
[8] Kim E, Park M, Ji S, Park M. A new approach to fuzzy modeling. IEEE
Trans Fuzzy Syst 1997;5(3):328–37.
[9] Mouzouris GC, Mendel JM. Dynamic non-singleton fuzzy logic systems
for nonlinear modeling. IEEE Trans Fuzzy Syst 1997;5(2):199–208.
[10] Wang L, Langari R. Complex systems modeling via fuzzy logic. IEEE
Trans Syst Man Cybern-Part B: Cybern 1996;26(1):100–6.
[11] Kukolj D. Design of adaptive Takagi–Sugeno–Kang fuzzy models. Appl
Soft Comput 2002;2:89–103.
[12] Jang JR. ANFIS: adaptive network based fuzzy inference system. IEEE
Trans Syst Man Cybern 1993;23(3):665–85.
[13] Narendra KS, Parthasarathy K. Identification and control of dynamical
systems using neural networks. IEEE Trans Neural Netw 1990;1(1):4–27.
[14] Babuska R, Verbruggen H. Neuro-fuzzy methods for nonlinear system
identification. Ann Rev Control 2003;27:73–85.
[15] Lin Y, Cunningham GA. A new approach to fuzzy-neural system
modeling. IEEE Trans Fuzzy Syst 1995;3(2):190–8.
[16] Evsukoff A, Branco ACS, Galichet S. Structure identification and
parameter optimization for non-linear fuzzy modeling. Fuzzy Sets and
Systems 2002;132:173–88.
[17] Wu B, Yu X. Fuzzy modeling and identification with genetic algorithm
based learning. Fuzzy Sets and Systems 2000;113:351–65.
[18] Pham DT, Castellani M. Evolutionary learning of fuzzy models. Eng Appl
Artif Intell 2006;19:583–92.
[19] Kilic K, Sproule BA, Turksen IB, Naranjo CA. A fuzzy system modeling
algorithm for data analysis and approximate reasoning. Robot Auton Syst
2004;49(3–4):173–80.
[20] Hanss M. Identification of enhanced fuzzy models with special
membership functions and fuzzy rule bases. Eng Appl Artif Intell 1999;
12:309–19.
[21] Cordon O, Herrera F, Hoffman F, Magdalena L. Genetic fuzzy systems:
Evolutionary tuning and learning of fuzzy knowledge bases. Singapore,
New Jersey, London, Hong Kong: World Scientific; 2001.
[22] Kang S, Woo C, Hwang H, Woo KB. Evolutionary design of fuzzy rule
base for nonlinear system modeling and control. IEEE Trans Fuzzy Syst
2000;8(1):37–45.
[23] Farag WA, Quintana VH, Torres GL. A genetic based neuro-fuzzy
approach for modeling and control of dynamical systems. IEEE Trans
Neural Netw 1998;9(5):756–67.
[24] Pham DT, Karaboga D. Intelligent optimisation techniques: Genetic
algorithms, tabu search, simulated annealing and neural networks.
Springer-Verlag; 2000.
[25] Box GEP, Jenkins GM. Time series analysis, forecasting, and control. San
Francisco (CA): Holden Day; 1970.
[26] Tong RM. Synthesis of fuzzy models for industrial processes-some recent
results. Int J General Syst 1978;(4):143–62.
[27] Pedrycz W. An identification algorithm in fuzzy relational systems. Fuzzy
Sets and Systems 1984;(13):153–67.
[28] Xu CW, Lu YZ. Fuzzy model identification and self-learning for dynamic
systems. IEEE Trans Syst Man Cybern 1987;(17):683–9.
[29] Sugeno M, Tanaka K. Successive identification of a fuzzy model and its
application to prediction of a complex system. Fuzzy Sets and Systems
1991;(42):315–34.
[30] Nakoula Y, Galichet S, Foulloy L. A learning method for structure and
parameter identification of fuzzy linguistic models. In: Hellendoorn H,
Driankov D, editors. Selected approaches to fuzzy model identification.
Berlin: Springer; 1997. p. 281–319.
[31] Bagis A, Karaboga D. Artificial neural networks and fuzzy logic based
control of spillway gates of dams. Hydrol Process 2004;18:2485–501.