3

A General Study on Genetic Fuzzy

Systems
OSCAR CORDO N, FRANCISCO HERRERA

3.1 INTRODUCTION
As it is known, a rule based system (production rule system) has been successfully used
to model human problem-solving activity and adaptive behavior, where a classic way
to represent the human knowledge is the use of IF/THEN rules. The satisfaction of the
rule antecedents gives rise to the execution of the consequent, one action is performed.
The conventional approaches to knowledge representation are based on bivalent logic.
A serious shortcoming of such approaches is their inability to come to grips with the
issue of uncertainty and imprecision. As a consequence, the conventional approaches do
not provide an adequate model for modes of reasoning and all commonsense reasoning
fall into this category.
Fuzzy Logic (FL) may be viewed as an extension of classical logical systems,
provides an eective conceptual framework for dealing with the problem of knowledge
representation in an environment of uncertainty and imprecision. FL, as its name
suggests, is the logic underlying modes of reasoning which are approximate rather than
exact. The importance of FL derives from the fact that most modes of human reasoning
-and especially commonsense reasoning- are approximate in nature. FL is concerned
in the main with imprecision and approximate reasoning.
The applications of FL to rule based systems have been widely developped. From
a very broad point of view a Fuzzy System (FS) is any Fuzzy Logic Based Sytems,
where FL can be used either as the basis for the representation of dierent forms of
knowledge systems, or to model the interactions and relationships among the system
variables. FS have been shown to be an important tool for modelling complex systems,
in which, due to the complexity or the imprecision, classical tools are unsuccessful.
Sample Contributed Book
Editor Jenny Smith c 1993 John Wiley & Sons Ltd.
cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
 2 OSCAR CORDO N, FRANCISCO HERRERA

DESIGN PROCESS

Genetic Algorithm Based

Learning Process

Knowledge Base

Input Interface Fuzzy System Output Interface

Environment Computation with Fuzzy Systems Environment

Figure 3.1 Genetic Fuzzy Sysems

Among the most successful applications of this systems has been the area of Fuzzy
logic controllers (FLCs). FLCs are rule based systems useful in the context of complex
ill-dened processes, especially those which can be controlled by a skilled human
operator without knowledge of their underlying dynamics. Recentely fuzzy control
techniques have been applied to many industrial processes, FLCs have been widthly
used in automation and engineering. The experience of skilled operators and the
knowledge of control engineers are expressed qualitetively by a set of fuzzy control
rules. In fact, one of the features of the FLCs is that the IF-THEN rules are described
on the base of the conventional control strategy and the experts' knowledge. Each fuzzy
rule has an antecedent, or IF, part containing several preconditions, and a consequent,
or THEN, part which prescribes the value.
Recentely, numerous papers and applications combining fuzzy concepts and genetic
algorithms (GAs) have become known, and there is an increasing concern in the
integration of these two topics. In particular, there are a great number of publications
exploring the use of GAs for developping fuzzy systems, the called genetic fuzzy systems
(GFSs). Figure 1 shows this idea.
This paper presents an overview of the GFSs, showing the use of the GAs in the
construction of the fuzzy logic controllers knowledge bases comprising the known
knowledge about the controlled system.
To achieve that, this paper is divided into 4 sections the rst being this introduction.
The section 2 introduces the fuzzy systems with a special attention to FLCs, while
section 3 presents the GFSs. Some nal remarks are made in section 4.
cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
 A GENERAL STUDY ON GENETIC FUZZY SYSTEMS 3
3.2 FUZZY SYSTEMS
Fuzzy logic and fuzzy sets in a wide interpretation of FL (in terms of which fuzzy logic
is coextensive with the theory of fuzzy sets, that is, classes of objects in which the
transition from membership to nonmembership is gradual rather than abrupt) have
placed modeling into a new and broader perspective by providing innovative tools to
cope with complex and ill-dened systems. The area of fuzzy sets has emerged following
some pioneering works of Zadeh [Zad65, Zad73] where the rst fundamentals of fuzzy
systems were established.
As we aforesaid, a rule based system has been successfully used to model human
problem-solving activity and adaptive behavior. The conventional approaches to
knowledge representation, are based on bivalent logic. A serious shortcoming of such
approaches is their inability to come to grips with the issue of uncertainty and
imprecision. As a consequence, the conventional approaches do not provide an adequate
model for modes of reasoning. Unfortunatelly, all commonsense reasoning fall into this
category.
The application of FL to rule based systems leads us to the fuzzy systems. The
main role of fuzzy sets is representing knowledge about the problem, or to model the
interactions and relationships among the system variables. There are two essential
advantages for the design of rule-based systems with fuzzy sets and logic:
 the key features of knowledge captured by fuzzy sets involve handling
uncertainty, and
 inference methods become more robust and exible with approximate
reasoning methods of fuzzy logic.
Knowledge representation is enhanced with the use of linguistic variables and their
linguistic values that are dened by context-dependent fuzzy sets whose meanings are
specied by graded membership functions. On other hand, inference methods such
as generalized modus ponens, tollens, etc., which are based on fuzzy logic form the
bases of approximate reasoning with pattern matching scores of similarity. Fuzzy logic
provides an unique computational base for inference in rule based systems. Unlike
traditional logical systems, fuzzy logic is aimed at providing modes of reasoning which
are approximate and analogical rather than exact.
Abording the fuzzy system modeling issue, it is essentially developed into two
dierent types of system models identied as acquisition of rules and their parameters:
i) fuzzy expert system models, and ii) fuzzy logic controllers.
Fuzzy expert system models are designed, developed and implemented with a direct
participation of a system's expert who is throughly familiar with the characteristic
behaviour of the system under investigation. The knowledge of the expert is extracted
from the expert through experimental methods of questionnaires, protocols and
interviews which may be conducted by people or by computers for the purpose of
identifying the form and the structure of the rules, i.e., structure identication as
well as the membership functions of the linguistic values of linguistic variables, i.e.,
parameter identication.
On the other hand, fuzzy control model design, development and implementation
are dependent on the availability of input-output data sets. The system structure
identication in terms of rules and specication of membership functions that dene
cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
 4 OSCAR CORDO N, FRANCISCO HERRERA

Knowledge Base

Fuzzification Defuzzification
Inference System
Interface Interface

State Variables Control Variables

Controlled System

Figure 3.2 Generic structure of a fuzzy logic controller

the meaning representation of linguistic values of linguistic variables are determined

by learning techniques. Also there is a third kind that is the combination of the others
two that may be called fuzzy expert-control.
Here, we will center in the second fuzzy system model, the FLCs, where the GAs
have been used for design the structure identication of the system.
FLCs, initiated by Mamdani and Assilian in the work [MA75], are now considered
as one of the most important applications of the fuzzy set theory. FLCs are knowledge
based controllers that make use of the known knowledge of the process, expressed in
form of fuzzy linguistic control rules collected in a knowledge base (KB), to control
it. The advantage of this approach with respect to the classical Control Theory is
that it has not necessity of expressing the relationships existing in the system by
means of a mathematical model, what constitutes a very dicult task in many real
situations presenting nonlinear characteristics or complex dynamic. In the following
two subsections we present an introduction to FLCs and to the FLC KBs.

3.2.1 Fuzzy Logic Controllers

An FLC is composed by a Knowledge Base, that comprises the information given by
the process operator in form of linguistic control rules, a Fuzzication Interface, which
has the eect of transforming crisp data into fuzzy sets, an Inference System, that
uses them joined with the Knowledge Base to make inference by means of a reasoning
method, and a Defuzzication Interface, that translates the fuzzy control action so
obtained to a real control action using a defuzzication method. The generic structure
of an FLC is shown in gure 2 [Lee90].
Several factors with a signicant in uence have to be analyzed in order to design an
FLC for a concrete process. Concretely, there are two main decisions to make in order
to design a FLC, to derive a KB for the system and to decide the reasoning method to
use. As can be viewed, only the rst one depends directly on the concrete application
altough several reasoning methods will perform better with some kind of systems than
others.
cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
 A GENERAL STUDY ON GENETIC FUZZY SYSTEMS 5
The Knowledge Base is the FLC component comprising the expert knowledge
known about the controlled system. So it is the only component of the FLC depending
on the concrete application and it makes the accuracy of the FLC depends directly on
its composition. It is composed by a set of fuzzy control rules with the form:
If X1 is Ai1 and X2 is Ai2 and ... and Xn is Ain then Y is Bi
being the Xi and Y linguistic system variables and the Ai and the Bi linguistic labels
associated with fuzzy sets specifying their meaning.
The Fuzzication Interface denes a mapping from an observed input space to
fuzzy sets in certain input universes of discourse, obtaining the membership function
associated to each one of the crisp system inputs.
The Inference System is based on the application of the Generalized Modus
Ponens, extension of the classical logic Modus Ponens proposed by Zadeh. It is done by
means of the Compositional Rule of Inference (CRI) given by the following expression:
n o
B (y) = Supx2X T (A (x); I(Ai (x); B (y)))
0
0
0 (3.1)
being T, T connectives, Ai (x) = T (Ai1 (x); : : :; Ain (x)) and I an implication
0

operator.
Since the input x corresponding to the state variables of the controlled system is
crisp, x = x0, the fuzzy set A is a singleton, that is, A (x) = 1 if x = x0 and
0
0

A (x) = 0 if x 6= x0 . Thus the CRI is reduced to:

0

B (y) = I(Ai (x0 ); B (y))

0 (3.2)
Since from each rule Ri is obtained a fuzzy set Bi from the inference process, the
0

Defuzzication Interface uses an aggregation operator G which composes them and
applies a defuzzication method D to translate the fuzzy sets obtained in this way into
values corresponding to the control variables of the system. So, calling S to the FLC,
x0 to the inputs value and y0 to the crisp value obtained from the defuzzication, we
have:
n o
B (y) = G B1 (y); B2 (y); : : :; Bn (y)
0 0 0 0 (3.3)
y0 = S(x0 ) = D(B (y)) 0 (3.4)
At present, the commonly used defuzzication methods may be described as the Max
Criterion, the Mean of Maximum (MOM) and the Center of Area (COA) [Lee90].
The design tasks that have to be developed in order to decide the FLC reasoning
method are the selection of the fuzzy operators I, T and G and the defuzzication
operator D [KKS85]. The problem of selection them have been analyzed in several
works such us [CCC+ 94, CCC+ 95, CHP95b, CHP95a, KKS85].
For more information about FLCs see [CHP95a, DHR93, HMB93, Lee90].

3.2.2 The Fuzzy Logic Controller Knowledge Base

The Knowledge Base is comprised of two components, a Data Base (DB), containing
the denitions of the fuzzy control rules linguistic labels, that is, the membership
cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
 6 OSCAR CORDO N, FRANCISCO HERRERA
functions of the fuzzy sets specifying the meaning of the linguistic terms, and a
Rule Base (RB), constituted by the collection of fuzzy control rules representing the
expert knowledge. We are going to analize more concretely this two components in the
following sections.

The Data Base

The concepts associated with a DB are used to characterize fuzzy control rules and
fuzzy data manipulation in an FLC [Lee90]. In this way, the main task to be done in
order to design an FLC DB is to associate a membership function to every one of the
linguistic terms that the system input and output variables can take as possible values.
There are two modes of DB denition:
 By means of a quantization or normalization process.
 By means of a tuning or learning process.
Every linguistic variable involved in the FLC KB forms a fuzzy space with respect to
a certain universe of discourse and have associated a label set containing the possible
linguistic values that it can take. A fuzzy partition determines how many terms should
exist in the label set. The choice of the term set is equivalent to nding the primary
fuzzy sets or linguistic labels (terms). This is a previous task for the rst mode and
for some of the methods included in the second one.
The rst denition mode makes use only of a little part of the a priori known
knowledge of the system. In order to dene the meaning of the linguistic term set,
a discretization of the process input and output variables continuous universes of
discourse have to be performed. This process is usually called quantization and is
done in a number of steps [DHR93, HMB93, Lee90]:
1. The continuous domain is discretized (quantized) into a certain nite
number of segments called quantization levels. Each segment is labeled
as a generic element and the set of all generic elements forms a discrete
universe of discourse.
2. Given a linguistic value from a certain term set, the fuzzy set dening
the meaning of this linguistic value is built by assigning a degree of
membership to each generic element. This is done for every linguistic
value in a term set.
The use of quantized or normalized domains requires a scale transformation which
maps the physical values of the process variables into the discretized universe. In both
cases, the mapping can be uniform (linear), non-uniform (nonlinear) or both. Either a
numerical or functional denition may be used to assign the degrees of membership to
the primary fuzzy sets [DHR93, HMB93, Lee90]. This choice is based on the subjective
criteria and it should be more convenient that the human processes operators could
represent the meaning of their usually employed linguistic terms in form of fuzzy sets.
Unfortunatelly, in many real cases it is not possible for him to make that, and it is
very common the use of an uniform fuzzy partition as the one proposed by Liaw and
Wang [LW91] shown in gure 3.
cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
 A GENERAL STUDY ON GENETIC FUZZY SYSTEMS 7

NB NM NS ZR PS PM PB

-6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6

Figure 3.3 Fuzzy partition proposed by Liaw

The second denition mode employs many of the known knowledge and leads to
a more application specic DB requiring less intervention of the controller designer
and presenting more automatically developed tasks. Belonging to this group, we can
distinguish two dierent subgroups:
 Denition by means of a tuning process
 Denition by means of a learning process
Methods included in the rst one make use of a primary DB denition developed
by means of a quantization or normalization process and then apply a process that
modies the meaning of the linguistic labels, that is, the denitions of the fuzzy sets
associated to them. On the other hand, it would be a learning process when there is
not an initial DB denition. Usually, this last process is carried out joined to a Rule
Base learning process.
Moreover, an important decision in order to dene the FLC is to determine the
nature of its DB. There are two dierent approximations for it depending on the scope
in what there is assigned the meaning to the linguistic labels belonging to the dierent
term sets. On one hand, an usual DB denition process in what the fuzzy sets giving
meaning to the linguistic labels are uniformly dened for all rules included in the RB
constitutes a descriptive approach since the linguistic labels represents a real world
semantics. On the other hand, it can be considered a KB whose rules present dierent
meaning for the same linguistic terms. The meaning associated to a concrete label
will depend on the concrete rule in what this label appears. In this case, the KB and
the FLC using it present a dierent philosofy. The approach is approximative and the
system is in the line of an Universal Approximator [Cas95].

The Rule Base

There are dierent kinds of fuzzy control rules proposed in the specialized literature
regarding to the expression of the consequent. Mamdani and Assilian employ rules in
cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
 8 OSCAR CORDO N, FRANCISCO HERRERA
which the consequent is another fuzzy variable [MA75], while Sugeno et al. use rules
whose conclusion is a polynomial function of the inputs (TSK rules) [TS85]. Another
kind of rules present also the consequent being a function of the input parameters.
The following three rules show respectively the generical expressions of the three types
commented:
If X1 is A1 and ::: and Xn is An then Y is B
If X1 is A1 and ::: and Xn is An then Y = p0 + p1X1 + ::: + pn Xn
If X1 is A1 and ::: and Xn is An then Y = f(X1 ; :::; Xn)
being the Xi and Y, linguistic system variables and the Ai and B, linguistic labels
associated with fuzzy sets specifying their meaning.
There are four modes of derivation of fuzzy control rules that are not mutually
exclusive [HMB93, Lee90]. These modes are the following:
1. Expert Experience and Control Engineering Knowledge.
2. Modeling of the Operator's Control Actions.
3. Based on the Fuzzy Model of a Process.
4. Based on Learning and Self-Organization.
The rst method has been widely used. This method is eective when expert human
operators can express what they use to control the system in terms of control rules.
The rules more usually obtained by means of this process are Mamdani type because
they present an adequate form to represent the expert knowledge. The second method
directly models the control actions of the process operator. Instead of interviewing the
operator, the types of control actions taken by him are modeled. The third approach
is based on developing a model of the plant and construct an FLC to control the fuzzy
model generating the fuzzy control rules of the RB by means of the fuzzy model of the
system. It makes this approach similar to that traditionally used in Control Theory.
Hence, structure and parameter identication are needed. Finally, the fourth method
is focused on learning. In this case, the ability to create fuzzy control rules and to
modify them based on experience in order to improve the controller performance is
considered.

3.3 DESIGNING GENETIC FUZZY SYSTEMS

As it has been shown in the above section, there are many tasks that have to be
performed in order to design an FLC to control a concrete system. We have commented
yet that the derivation of the KB is the only one directly depending on the controlled
system and it presents a signicative importance in the design process. It is known
that the more used method in order to perform this task is based directly on extracting
the expert experience from the human processes operators. The problem arises when
these are not able to express their knowledge in terms of fuzzy control rules. In order to
avoid this drawback, researches have been investigating automatic learning methods for
designing FLCs by deriving automatically an appropiate KB for the controlled system
without necessity of its human operator.
The genetic algorithms (GAs) have demostrated to be a powerful tool for automating
the denition of the KB since adaptative control, learning and self-organization can
cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
 A GENERAL STUDY ON GENETIC FUZZY SYSTEMS 9
be considered in a lot of cases as optimization or search processes. The fuzzy systems
making use of a GA in their design process are called generically GFSs.
These advantages have extended the use of the GAs in the development of a wide
range of approaches for designing FLCs in the last years. Some of these approaches of
genetic FLC design will be shown in the present chapter. It is possible to distinguish
three dierent groups of genetic FLC design processes according to the KB components
included in the learning process. These ones are the following:
1. Genetic denition of the Fuzzy Logic Controller Data Base.
2. Genetic derivation of the Fuzzy Logic Controller Rule Base.
3. Genetic learning of the Fuzzy Logic Controller Knowledge Base.
In the following subsections we are going to analyze each one of the approaches for
genetic design of FLCs.

3.3.1 Dening the Fuzzy Logic Controller Data Base using Genetic Algorithms
As we have commented already, one of the modes of denition of the FLC DB is
based on learning or tuning this FLC component. The dierence between these two
approaches depends on the existence of a previously primary DB denition. While
learning processes do not need this previous denition, tuning processes works over it
obtaining a more accurated one.
Several methods have proposed in order to dene the FLC DB using GAs
[BN95, BMU95, FTH94, HLV95b, HTS93, Kar91b]. All of them are based on the
existence of a previously dened RB, usually extracted from the process operator. Each
chromosome involved in the evolution process will represent dierent DB denitions,
that is, each one of the chromosomes will contain a coding of the whole membership
functions giving meaning to the linguistic terms. The degree of adaptation of an
individual is measured using a tness function that usually is based on the aplication of
the FLC to the controlled system, using a KB formed by the RB and the DB encoded
by the chromosome.
There are two dierent approaches for the genetic denition of FLC DBs depending
on the scope of the association of membership functions to linguistic labels in the KB,
either all fuzzy control rules using the same meaning for the system variables linguistic
terms or a dierent approximation in what each rule presents its own meaning for the
labels involved by it. In this subsection we analyze an example of each one of both
groups. The method proposed by Karr [Kar91b], belonging to the rst one, and the
method of Herrera et al. [HLV95b], belonging to the second one.

The DB denition method proposed by Karr

The approach of Karr [Kar91b] is based on the existence of primary fuzzy partitions of
the dierent system variables input and output spaces. The GA is applied for dening
the meaning of the dierent linguistic term sets, that is, for learning the fuzzy sets
associated to each one of the linguistic labels belonging to the fuzzy partitions. As it
can be viewed, this approach is a descriptive one because all the fuzzy terms appearing
in the fuzzy control rules will use the same meaning (that dened by means of the GA
learning process).
cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
 10 OSCAR CORDO N, FRANCISCO HERRERA

0.8

0.6

0.4 LOW MEDIUM HIGH

0.2

Figure 3.4 An example of fuzzy partition in the Karr's DB denition method

In order to develop this task, Karr uses a Simple GA with binary coding, proportional
selection mechanism, simple crossover and random mutation. The binary coding outline
will represent all the membership functions associated to the dierent linguistic labels
belonging to each one of the linguistic term sets into a single chromosome.
The membership functions selected by Karr to dene the meaning of the linguistic
labels are triangular-shaped. He consider only two points, extremes of the fuzzy sets
support, for dening every triangle, xing its central point which presents value 1 in
the membership function. In this way, the evolution performed by the GA can make the
triangles to be distorted (when the base width is altered) and translated (when these
two points are shifted along the x-axis) freely. The process is dierent for the extreme
triangles, which only requires a single point with value of the membership functions
equal to 0 to be dened. Their modal point is clearly placed in the correspondent
extreme of the concrete system variable universe of discurse. Only one operation can
be performed with these triangles, altering their base width in one of the two x-axis
directions, that is, making it bigger or smaller. Figure 4 shows an example.
The constraint placed over the membership functions is that the ones associated to
the extreme labels must remain right triangles while those associated to the interior
terms must remain isosceles triangles. It is clear that this constraint avoid the GA to
obtain incorrectlly dened membership functions.
A chromosome coding all the membership functions is built nally by joining all the
individual coding of these ones into a single string. As each point is represented by a
binary number with a xed number of bits, the chromosomes are of xed-length, that
is, all individuals in the population will present the same length.
Finally, Karr do not present a concrete tness function but introduces several
considerations in order to dene it. His idea for measuring the accuracy of a concrete
DB in the optimal control of the system is based on an application-dependent measure,
that is, any error or convergence measure (see [CCC+ 94, CHP95a]).
cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
 A GENERAL STUDY ON GENETIC FUZZY SYSTEMS 11
The DB denition method proposed by Herrera et al.
In [HLV95b] it was presented a DB denition process used to tune the DB parameters.
The process is based on the existence of a previous complete KB, that is, an initial
DB denition and a RB constituted by m control rules. The chromosomes will encode
a complete KB since each one of them contains the RB with a dierent DB associated.
The GA designed for the process present real coding issue and use the stochastic
universal sampling as selection procedure and the Michaelewicz's non-uniform mutation
operator. Regarding to the crossover, two dierent operators are employed: the simple
and the Max-Min-Arithmetical crossover. This last operator have been proposed by
the authors and makes use of fuzzy operators in order to improve the behavior of the
GA crossover operator.
The membership functions selected in order to dene the DB are trapezoidal- shaped.
They have associated a parametric representation based on a 4-tupla of real values.
Let the following rule be the ith rule of the previous RB:
If X1 is Ai1 and X2 is Ai2 and : : :and Xn is Ain then Y is Bi
Then the fuzzy sets giving meaning to the linguistic labels Aij associated to the input
variables Xi will be represented by the 4-tuple (cij ; aij ; bij ; dij ) and the ones associated
to the output variable linguistic labels Bi by (ci; ai ; bi; di). Thus each one of the rules
0 0 0 0

will be encoded in pieces of chromosome Cri , i = 1; : : :; m, in the following way:

Cri = (ci1 ; ai1; bi1; di1; :::; cin; ain; bin; din; ci ; ai; bi; di)
0 0 0
(3.5)
0

Therefore the complete RB with an associated DB is represented by a complete

chromosome Cr :
Cr = Cr1 Cr2 ::: Crm (3.6)
As it can be viewed, each individual of the population represents a complete KB.
More concretelly, all of them encode the derived system RB and the dierence between
them is the meaning associated to the linguistic variables taking part in the fuzzy
control rules, that is, the DB denition. As each rule is coded in a piece of chromosome,
the fuzzy set giving meaning to a linguistic term can be changed in one rule and not
in the other ones in which it appears or, in a more extreme case, it can be changed
in many rules presenting dierent forms in several ones. The meaning of the linguistic
terms will depend then on the rule in which they are involved. The KB so obtained
will present an approximative behavior.
The initial gene pool is created from the initial KB. This KB is encoded directly in
a chromosome, denoted as C1 . The remaining individuals are generated by associating
an interval of performance, [clh ; crh ] for every gene ch of C1, h = 1 : : :(n + 1)  m  4.
Each interval of performance will be the interval of adjustment for the correspondent
variable, ch 2 [clh ; crh ].
If (t mod 4) = 1 then ct is the left value of the support of a fuzzy number. The
fuzzy number is dened by the four parameters (ct, ct+1, ct+2 , ct+3 ) and the intervals
of performance are the following:
ct 2 [clt; crt ] = [ct , ct+12,ct ; ct + ct+12,ct ]
cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
 12 OSCAR CORDO N, FRANCISCO HERRERA

l r r r r
Ct Ct Ct C t+ 1 C t+ 1 C t+ 2 C t+ 2 C t+ 3 C t+ 3
l l l
C t+ 1 C t+ 2 C t+ 3

Figure 3.5 Intervals of performance

ct+1 2 [clt+1; crt+1] = [ct+1 , ct+12,ct ; ct+1 + ct+2 ,2 ct+1 ]

ct+2 2 [clt+2 ; crt+2] = [ct+2 , ct+2 ,2 ct+1 ; ct+2 + ct+3 ,2 ct+2 ]
ct+3 2 [clt+3 ; crt+3] = [ct+3 , ct+3 ,2 ct+2 ; ct+3 + ct+3 ,2 ct+2 ]
Figure 5 shows these intervals. Therefore, we create a population of chromosomes
containing C1 as its rst individual and the remaining ones initiated randomly, with
each gene being in its respective interval of performance.
By using a training input-output data set, ETDS , and a concrete error measure, the
square medium error, the tness function of a chromosome is dened. In this way,
the adaptation value associated to an individual is obtained by computing the error
between the outputs given by the FLC using the KB encoded in the chromosome and
those contained in the training data set. The tness function is represented by the
following expression:
X k
E(C) = 2jE1 j (ey , S(exk ))2 (3.7)
TDS ek 2ET DS

3.3.2 Deriving the Fuzzy Logic Controller Rule Base using Genetic Algorithms
The great majority of the approaches belonging to this group are based on learning the
consequents of the fuzzy control rules included in the FLC RB. In this way, many of
these genetic processes encode the complete system decision table in the chromosomes.
There are dierent methods developing this task [Bon93, HT94, Kar91a, KB93, Thr91]
but in many cases the only dierence existing among them is the cappability of learning
the number of rules forming the RB. This characteristic is presented when there exist
a possible alelle representing the absence of consequent for a rule with a concrete
antecedent, that is, the absence of the correspondent rule in the RB.
All methods belonging to this family suppose the existence of a dened DB. Thus,
a primary partition of the fuzzy spaces, that is, a term set for each one of them,
and a collection of fuzzy sets giving meaning to these primary labels are considered.
Other common characteristic for a majority of them is that they consider the FLC
RB represented in form of decision table (called too look-up table). As it is known,
cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
 A GENERAL STUDY ON GENETIC FUZZY SYSTEMS 13
an usual FLC RB constituted by control rules presenting n input variables and a
single output variable can be represented using an n-dimensional decision table, each
dimension corresponding to each one of the input variables. Every dimension will have
associated an array containing the labels of the concrete variable term set and the
cells of the decision table will contain the linguistic label that the output variable take
for the combination of antecedents represented in this cell. Therefore each table cell
represents a fuzzy control rule that can belong to the FLC RB.
This structure is encoded in the individuals forming the GA population. If there not
exist a value for the alleles representing the absence of value for the rule consequent,
there is not possible to derive a FLC RB with an optimal number of rules but all the
possible rules have to be considered belonging to it. This is the case of the methods
proposed by Karr [Kar91a], and Kropp and Baitinger [KB93] altough the rst of them
do not encode the complete decision table as we are going to see in the following. The
remaining ones commented are able to learn the number of fuzzy rules.
In this subsection we are going to study three dierent approaches. The methods
selected were proposed by Thrift [Thr91], Karr [Kar91a] and Bonarini [Bon93]. This
last one constitutes an original approach for learning FLC RB and diers a lot from
the others belonging to this family as we are going to see in the following.

The RB derivation method proposed by Thrift

This method, as many others belonging to these category, is based on encoding all the
cells of the complete decision table in the chromosomes. In this way, Thrift [Thr91]
establishes a mapping between the label set associated to the system output variable
and an increasing integer set (containing one element more and taking 0 as rst
element) representing the alelle set. An example is shown to clarify the concept.
Being fNB; NS; ZR; PS; PB g the term set associated to the output variable, it can
be noticed the absence of value for the output variable by the symbol -. The complete
set formed joining this symbol to the term set is mapped into the set f0; 1; 2; 3; 4;5g.
Hence the label NB is associated with the value 0, NS with 1, : : :, PB with 4 and the
blank symbol - with 5.
Therefore the GA emploies an integer coding. Each one of the chromosomes is
constituted by joining the partial coding associated to each one of the linguistic labels
contained in the decision table cells. A gen presenting the alelle - will represent the
absence of the control rule contained in the corresponding cell in the RB being the
chromosome phenotype.
The GA proposed emploies an elitist selection scheme and the genetic operators
used are of dierent nature. While the crossover operator is the standard two -point
crossover, the mutation operator is specically designed for the process. When it is
applied over an alelle dierent from the blank symbol, changes it either up or down
one level or to the blank code. When the previous gen value is the blank symbol, it
selects a new value at random.
Finally, the tness function is based too on an application specic measure. A
Measure of Convergence is considered for this task. The tness of an individual is
determined by applying the FLC employing the RB coded in its genotype to the
controlled system with several dierent starting points and computing the convergence
of it to the desired equilibrium point.
cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
 14 OSCAR CORDO N, FRANCISCO HERRERA
The RB derivation method proposed by Karr
Karr proposes a method for deriving FLC RBs [Kar91a] where the complete RB in
form of decision table is not coded in each chromosome. The approach of Karr is
based in the fact that the number of rules is provided by an expert, together with
many complete rules forming the RB and the antecedents for the reamining ones. He
presents an GA derivation method for learning the rule consequents of this last type
rules.
As it can be viewed, the method needs the existence of a deep knowledge about
the controlled system. As Karr comments, his approach is useful when the controller
human designer knows all the input states relevant for the problem (in this way, he
knows the number of rules needed to control it) and the control action that have to
be performed in order to obtain an optimal control from several of them (that is, he
knows several complete antecedent-consequent control rules), but the actions that have
to be associated to the remaining input states are not obvious for him. Usually, the
known rules are those describing the extreme conditions, which are easy to write in
many cases.
The coding outline will associate a binary number to each label belonging to the
output variable term set in the way that each of them will be mapped to a binary
number according to its order in the label set. Thus, considering seven linguistic
terms describing the control variable, the possible control actions for one rule will be
represented as a three-bit string (000 represents action 1 (NB), 001 represents action
2 (NM), and so on). A chromosome is obtained by joining the partial codings of the
rules for what there is not known the control action to apply. Therefore a chromosome
codes only a little part of the complete RB. In the commented example, calling n
the number of not determinated rules, a string of length 3n represents every possible
conguration for this partial FLC rule set.
The GA employed is the same used for the Karr's DB denition method that we
have commented in the previous subsection, that is, a Simple GA with binary coding,
proportional selection mechanism, simple crossover and random mutation. In the same
way, the tness function is dened by using an application specic measure.
Finally, it is important to note that the author proposes to combine their two FLC
genetic design methods, the DB denition method and the RB derivation method, in
order to design a complete KB. In this way, the RB derivation method is applied in
a rst step and, once an acceptable RB have been learned by the GA, the selection
of high-performance membership functions (an accurated DB) is carried out using the
above described tuning process.

The RB derivation method proposed by Bonarini

The method proposed by Bonarini [Bon93], called ELF (Evolutionary Learning of
Fuzzy Rules), is quite dierent to the other approaches developed into this group. The
author considers the high computationally cost needed to derive a RB by applying an
GA with a population of individuals coding the whole decision table and presents a
model based on learning only the control rules that the FLC will use.
In order to develop this task, the GA used by Bonarini maintains a population of
rules This new coding will allow to learn the complete structure of the control rules,

cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
 A GENERAL STUDY ON GENETIC FUZZY SYSTEMS 15
that is, their antecedent and consequent and the optimal number of rules forming the
FLC RB. The author wants to learn with ELF what states occur in the controlled
system and which are irrelevant for it, obtaining optimal RBs for the application.
The tness function judges the state reached at each activation of the rules. Each
individual of the population, that is, each rule, will have associated information about
several questions: how good it has been judged (its strength), when it has been
generated, when it has been triggered the last time and how much it contributed to past
control actions performed by the controller. ELF will modify the strength associated to
a rule according to the performance of the action it contributed to. This performance
is evaluated by the tness function.
Other important characteristic of this method is that it is designed in order to run
in a real enviroment. The process rst selects among the rules matching the current
state of the controlled system, the rules matching better than a degree given by the
FLC designer. The rules belonging to this set will compete between them in order to
propose the best control action for the current state.
Several genetic operations are applied over this set of individuals. Some of these
rules are considered tested enough since they contributed to past control actions more
than a given reference. If some of them have low strength, it means that the control
actions proposed by them do not perform well. These rules are substituted by others
performing better and the consequents of such rules are modied with a probability
proportional to their strength. This last step constitute the GA mutation operator and
makes ELF to look for new rules in a neighboorhood of the good ones previously
learned.
If much time have passed from the last rule modication, this means that the
population of rules matching well the current state have stucked and all of them have
almost the same strength. In this case, ELF selects the worst rule and mutates its
consequents in order to continue the search looking for a better conguration.
If there exists few rules matching the current state (the parameter representing the
enough number of rules is changed dinamically), ELF generates a new rule covering it
and proposes a control action at random. This is the only mechanism introducing new
antecedents and it is called cover-detector. It may introduce with a given probability
also "dont care" symbols as values for some variables in the new rule generated. Rules
whose antecedents contain these kind of symbols match dierent states and compete
with dierent groups of rules, one for each of the matching states.
The dual situation is in which there are too many rules matching a state. In this
case, the genetic operator applied is called rule-killer and it simply eliminate the worst
rule matching this state from the population.
One time performed the selection and applied the genetic operators, ELF uses the
FLC with the RB coded in the population taking the current state as input. The tness
function will then evaluate the new state obtained, giving it a reward. There exists a
sharing process in the distibution of this reward, each individual gets part of the total
reward according to the contribution of the rule that it codes gave to the control action
applied. This process results in a modication of the strenght associated to the rules
coded in the population.
Finally, several questions have to be noted. On one hand, Bonarini proposes
application specic performance measures to dene the tness function. On the other
hand, ELF gives dierent RBs as output of the learning process. It is due to each
cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
 16 OSCAR CORDO N, FRANCISCO HERRERA
time the performance of the system is higher than a "satisfactory" value given by the
controller designer, it saves the current RB and forces modications in the current
population in order to obtain a better solution.

3.3.3 Learning the Fuzzy Logic Controller Knowledge Base using Genetic
Algorithms
This last group is the one to with more contributions have been made in the last
years. There exists many approaches for genetic learning complete FLC KBs such as
[CV93, HLV94, HLV95a, LT93, LP94, LM94, NL94, NHW92, SK94, VM95] and all of
them present dierent characteristics. Belonging to these great group of works, we nd
approaches presenting variable cromosomal length [CV93, LP94], others making use of
many expert knowledge in order to improve the learning process [HLV95a, LT93] and
several working with chromosomes encoding single control rules instead of complete
KBs [HLV95a, VM95]. Many of them dene the tness function by means of a single
application specic measure, usually a measure of convergence, while others include
more objective to optimize for obtaining more robust KBs [HLV95a, LT93]. In the
following we are going to study ve dierent approaches belonging to this family.

The KB learning method proposed by Lee and Takagi

In [LT93] Lee and Takagi introduce an GFS design method that allows to learn
automatically a complete FLC KB. The method is based on an RB formed by TSK
control rules. In this way, the derivation of the rule consequents will consist on learning
the weights wi used for combining linearly the input values in order to obtain the
outputs.
For the linguistic terms in the rule antecedents, the authors consider triangular-
shaped membership functions although they note that the method can work with any
kind of parametrized membership functions, such as Gaussian, bell, trapezoidal or
sigmoidal ones. They suppose a fuzzy partition of the input space. Each triangular
membership function associated to each one of the primary fuzzy sets (linguistic terms)
is represented by three parameters. The rst one is the center point of the triangle,
that is, its modal point. Only the center of the triangle associated to the rst primary
fuzzy set is given as an absolute position, while the parameters associated to the other
ones represent the distance existing from the center point of the current triangle to
the previous one. The other two parameters represent the left and right point of the
triangle base respectively and they both present membership function value 0.
The GA used emploies a binary coding. First, the membership functions are encoded
by taking the binary values of the three associated parameters and joining them into
a binary substring. Eight bits are used to encode each parameter value. Then, the
complete DB is encoded by joining the partial codings of the membership functions
associated to each one of the input variables primary fuzzy sets one after other. The
last part of the chromosome is built by coding the parameters wi associated to each
combination of the input values and joining them into a new binary substring. Eight bits
are used again to encode the values of these parameters. In this way, each chromosome
represents a complete KB. The number of rules forming the KB will depend on the
number of primary fuzzy sets associated to each one of the input variables and it will
cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
 A GENERAL STUDY ON GENETIC FUZZY SYSTEMS 17
be equal to the product of them.
The coding used allows to decide the optimal number of control rules forming the RB
in the following way. Those linguistic terms in what the center point of their associated
triangular membership functions are out of a concrete bound, obtained from the known
knowledge about the system, are not considered to belong to any rule in the RB. Thus
all rules whose antecedent should be a combination of the valid input values coded in
the rst part of the chromosome, will belong to the FLC RB. As an example, let us
consider a chromosome coding the KB of a system with two input variables having
associated m and n valid linguistic labels respectively. The whole number of rules of
the RB encoded in it will be m  n.
The tness function is based on optimizing two dierent criteria. On one hand, an
application specic measure is considered (concretely, in the example proposed it is
used a measure of convergence and the FLC is applied from dierent initial states). On
the other hand, chromosomes coding RBs with a large number of rules are penalized
in order to obtain others with less rules.
Finally, it is very important to note that the authors propose two dierent ways
of incorporating previous known knowledge about the system to the proposed GFSs
design method in order to improve its behavior. They comment that this previous
knowledge will make the GA gain signicant speedup if the solutions designed by
means of it are approximatelly correct. On one hand, it is possible to incorporate
knowledge via the initial settings for the FLC KB parameters. This knowledge is used
to generate the GA initial population. Thus, the individuals forming it are not all
generated at random but several of them are obtained by equally partitioning the
input spaces into varying number of linguistic terms. This knowledge can also be used
to initially set appropriately the parameters wi of the rule consequents. On the other
hand, the previous knowledege can be incorporated via structural representation of the
KBs. For example, in a problem of symmetric nature as the Inverted Pendulum, the
FLC can be constrained in order to simetrically partition the input space about the
origin. This will reduce the size of the search space because the number of membership
functions is reduced by half.

The KB learning method proposed by Ng and Li

The method proposed by Ng and Li [NL94] is very specic since it is designed for
applying in two-inputs-one-output systems all whose input and output spaces are
partitioned into exactly seven primary fuzzy sets (the term set of the three variables
is the same: fzero; small; medium; largeg ). On the other hand, it is not able to
learn the number of rules that will constitute the optimal RB. As we are going to see
in the following, the whole decision table is represented in the part of chromosome
corresponding to the RB. Therefore, the way for derivating the RB is the same used by
the methods belonging to the second group commented in this work. The look-up table
will allways present dimmension 7  7 and all the rules contained in it will belong to
the nal RB.
The DB considered is designed by using symmetrical exponential membership
functions dened by the following expression:

cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
 18 OSCAR CORDO N, FRANCISCO HERRERA

i !
i(x) = exp , jx ij (3.8)
i
considering that +large = 1, if x > +large and ,large = 1, if x < ,large .
Hence, it is possible a parametric representation of the membership functions by
means of the set of parameters (; ; ), representing respectively the position, shape
and scaling parameters.
The most important characteristic of this approach is that the FLC design space
is coded in base-7 chromosomes. As in the great majority of the methods belonging
to this group, each chromosome represents a complete KB. Each chromosome is built
by joining ve substrings coding the dierent parameters of the problem. The rst
one of them represents the FLC RB and, as we have commented yet, the process is
similar to this developed in the methods deriving only the FLC RB. The fourty nine
rule consequents contained in the decision table cells are encoded one after other in a
substring presenting a base-7 value for each one of them.
The second substring is constituted by the scaling parameters () associated to the
seven linguistic labels of the two input variables. It presents fourteen base-7 values,
two for each one of the following parameters: large , medium , small and zero .
The next substring of eight base-7 values represent the positions () of the fuzzy sets
associated to the terms "small" and "medium" (each one requiring two bits), whilst
the positions of those giving meaning to "large" and "zero" are xed (,large = ,3,
+large = 3 and zero = 0). The fourth substring represents K1 and K2 as gains of the
two inputs variables, error and change of error, with three base-7 bits associated to each
and, nally, the last group of eigth integer characters encodes the shape coecients 
of the fuzzy sets associated to the term set of both variables, requiring a base-7 value
for each parameter.
The tness function employed by the GA is application specic. It is based directly on
computing the value obtained by the FLC in a measure of convergence when controlling
the system by using the concrete KB encoded in the individual.

The KB learning method proposed by Leitch and Probert

The more important characteristic presented by the method proposed by these authors
[LP94] is the specic coding outline that they design in order to represent FLC KBs into
chromosomes in an ecient way. The coding proposed will make the genetic variation
to be increased, reducing in this way premature convergence and avoiding the need for
complex crossover operators.
Leitch and Probert begin studying the previous approaches for designing GFSs and
comment that the coding schemes employed by the dierent GAs developed for this
task in the specialized literature represent the KB in a very xed form. In order to
avoid this shortcoming they develop a coding scheme more exible than the position
dependent ones commented. In their context depending coding the meaning of a piece
of chromosome is not determined by the absolute position that it presents into the
genotype, as usual, but it is determined by surrounding genes contained into it. In this
way, the chromosomal length is variable.
They consider a DB with spherical fuzzy sets each one determined by its concrete
cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
 A GENERAL STUDY ON GENETIC FUZZY SYSTEMS 19
center point presenting height 1, and all of them by a global radius value. All inputs and
output are scaled to the interval [0; 1] and the values of the commented parameters are
encoded by using binary substrings of variable length. The parameter value is decoded
in the following way. Let it be a binary substring of length n whose integer value is
m  n. The real number encoded in it is 2nm,1 .
The method is able to learn the optimal number of rules forming the RB, the input
variables involved in each rule and the shape of the membership functions associated to
the input and output variables. As it can be viewed, it does not use a previously dened
fuzzy partition of the input and output spaces but learn directly the membership
functions associated to each variable involved in each concrete rule. Therefore, the KB
obtained from the learning process clearly present an approximative behavior.
Each chromosome encodes a complete KB by using the context depending coding
commented. The alphabet of the chromosome consists of the integers 0, 1 and the letter
E, used to indicate the end of a number (the value E is associated to a surrounding
gen). The encoding of a rule is obtained by joining several substrings separated by
E symbols. The rst one presents a bit for each system input variable, taking value
1 if the concrete variable is involved in the rule and 0 otherwise. The surrounding
genes contained in this substring are ignored. Next substring encodes the center of the
membership functions associated to the input variables involved in the rule, each one
separated from each other by a surrounding gen. An empty value is associated to the
variables not taking part in the rule. The third substring presents only a binary number
coding the membership functions radius. The value encoded represents a percentage
from a maximum radius value given by the controller designer. The last substring
encodes the central point of the membership function associated to the output variable.
A chromosome is obtained by joining the partial codings of the rules contained in
the KB. The evolutionary process can make the end of a chromosome not coincide with
the end of a rule (it rarely does). In this case, the incomplete rule is ignored and the
genes coding it (called junk genes) are used to limit the eect of the disruption caused
by the genetic operators.
Regarding to these ones, single point crossover is employed with a little modication
in its usual form: the crossover point is randomly selected on each of the chromosomes
involved. On other hand, classical mutation operator is used. Due to the coding scheme
employed by the GA, both operators can not produce illegal chromosomes.
We nish noting several remarks about the method. The initial population formed
by chromosomes of random initial length is generated at random. The tness function
emploies a measure of convergence since the authors enunciate that this is the more
ecient way to design GFSs. Several noise is introduced in the simulations in order to
obtain more robust KBs from the learning process.

The KB learning method proposed by Cooper and Vidal

The underlying idea in the approach proposed by Cooper and Vidal [CV93] is that the
excesive length of the chromosomes encoding the KB can make the GA not able to
nd accurated solutions due to the high complexity of the search space. In this way,
they propose a novel encoding scheme which maintains only those rules necessary to
control the target system, alowing to obtain an RB with optimal number of rules.
In this case, the membership functions contained in the DB are triangles
cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
 20 OSCAR CORDO N, FRANCISCO HERRERA
characterized by the location of its center and the half-length of its base. A single rule,
therefore, consists of the concatenation of the one-byte unsigned characters (assuming
values from 0 to 255) specifying the centers and half-lengths of the membership
functions. The rule descriptions for a single KB are then concatenated into a single bit
string where the number of rules is not restricted. Hence, the GA employs a integer
coding and the chromosomes present variable-length.
This GA does not use the classical genetic mutation operator. In this approach, this
operator include the inversion of the copied bit and the addition or deletion of an entire
rule. These latter two mutations permit the size of a FLC KB to evolve. The cycle of
evaluation and reproduction continues for a predetermined number of generations or
until an acceptable performance level is achieved.
To be meaningful, the genetic paradigm requires that the rules in the two strings
should be aligned so that similar rules are combined with each other. Simply by
combining the strings in the order they appear it does not preserve much information
about either KB encoded and produces nearly random results, rather as a child KB that
performs in a manner similar to its parents. For example, this can make that the child
chromosome will present repeated rules. Therefore, before reproduction, both strings
must be aligned so that the centers of the input variables match as closely as possible.
The most closely matching rules are combined rst, followed by the next most closely
matching rules from those remaining and so on. Any rules forming a longer string that
is not matched are added at the end.
The tness function is designed by using a measure of convergence. The FLC is used
to control the system from twenty test cases and its accuracy is measured during sixty
seconds each trial.

The KB learning method proposed by Herrera et al.

The KB learning process proposed in [HLV95a] presents several important dierences
with respect to the other methods belonging to the same group as we are going to
see in the following. First, it can be considered the existence of previous control rules
derivated directly from the human process operator and include them in the learning
process, being able to combine them with other rules automatically learned and to
detect incorrect rules into this previous set. Moreover, the approach is based on several
steps and not in a single process such in the other methods. The chromosomes of the
main GA represent single rules and not complete KBs. On other hand, the tness
function is based not only in an FLC performance measure but it include several
criteria in order to obtain optimal KBs. Finally, the approach will obtain a KB with an
approximative nature even in the case in which there exists several descriptive fuzzy
control rules provided by the expert.
The approach is based on the use of GAs under the following hypotheses:
 There is some linguistic information from the experience of the human
controller but linguistic rules alone are usually not enough for designing
successfully a control system or could not be available.
 There is some numerical information from sampled input-output (state-
control) pairs that are recorded experimentally.
 The combination of these two kinds of information may be sucient for
cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
 A GENERAL STUDY ON GENETIC FUZZY SYSTEMS 21
a successful design of a FLC KB.
 The possibility of not having any linguistic information and having a
complete numerical information is considered.
Taking into account the aforementioned hypothesis a learning process is designed
according to the following goals:
 to develop a KB generating process from numerical data pairs; and
 to develop a general approach combining both kinds of information,
linguistic information and fuzzy control rules obtained by the generating
process, into a common framework using both simultaneously and
cooperatively to solve the control design problem.
In order to reach these goals, it is proposed a methodology based on the design of
the three following components:
1. a KB generating process of desirable fuzzy control rules able to include
the complete knowledge of the set of examples,
2. a combining information and simplifying rules process, which nds the
nal KB able to approximate the input-output behaviour of a real system,
3. a tuning process of the nal KB DB,
all of them developed by means of GAs.
As it is possible to have some linguistic IF , THEN rules given by an expert,
it is used a linguistic fuzzy rules structure to represent them. That is, a previously
dened DB representing fuzzy partitions with real-world meaning is considered. On
other hand, there are sampled input-output pairs and a free fuzzy rules structure to
generate the fuzzy rules covering these examples is used (the meaning presented by
alinguistic term is dierent when it belongs to dierent rules, that is, the example set
are used to generate rules with an approximative behavior). The membership functions
of the linguistic labels involved in the control rules are trapezoidal-shaped. Then both
kind of rules are combined, by using a simplication method based on a GA, and
nally a tuning method is applied over the simplied set of rules that will make all of
them nally present an approximate behavior.
The generating fuzzy rules process consists of a generating method of desirable fuzzy
rules from examples using GAs together with a covering method of the set of examples.
 The generating method of fuzzy rules is developed by means of a real
coded GA (RCGA) where a chromosome represents a fuzzy rule and it
is evaluated by means of a frequency method. The RCGA nds the best
rule in every running over the set of examples according to the following
features which will be included in the tness function of the GA.
 The covering method is developed as an iterative process. It permits to
obtain a set of fuzzy rules covering the set of examples. In each iteration, it
runs the generating method choosing the best chromosome (rule), assigns
to every example the relative covering value and removes the examples
with a covering value greater than a value  provided by the controller
designer.
cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
 22 OSCAR CORDO N, FRANCISCO HERRERA
Because we can obtain two similar rules in the generating process or one rule similar
to another given by an expert, it is necessary to combine and simplify the complete
KB obtained from the previous process for deriving the nal KB allowing to control
the system. Finally, the tuning method presented in [HLV95b] and commented in a
previous subsection is applied over the simplied KB for obtaining a more accurated
one.

3.4 FINAL REMARKS

In this chapter we have reviewed some approaches of GFSs in which GAs have been
used for designing FLCs KBs. Three dierent modes to cope this problem have been
attached and dierent methods developing each one of them have been analyzed.
There are many others researchs that have contributed to these three areas as we
have cited above. Moreover, there are other researchs that have contributed to various
aspects of GFSs, not included in the three GFS designing modes aforementioned, which
we have not a change to deal with but are worthy to mention. They are commented in
the following.
In [SKG93] it is proposed an automatic design method combining self-organizing
feature maps and GAs. In a rst step, the fuzzy control rules and linguistic variables
are extracted from a referential data set by using a self-organizing process. Then a
GA is used to nd optimal membership functions (that is, to tune the DB).
George et al. [GSR94] propose a method based on GAs and Neural Networks (NN)
for dening FLC DBs by learning the membership functions associated to the primary
fuzzy partitions provided by the expert. The combination of the two search process
allows to combine the best characteristics of each one of them in order to solve the
problem. Other papers combining both GA and NN are [IFSA95, SFH95].
Homan and Pster present two dierent genetic based methods for designing
hierarchical FLCs by learning the RB [HP94, HP95]. While the rst of them uses
a Simple GA with xed-length binary coding and usual genetic operators [HP94], the
second one employs a Messy GA with variable-length binary coding and the crossover
operator replaced by two simple ones: splice and cut.
Several approaches based on fuzzy classier systems (FCSs) have been developed
[BK94, CF94, PB93, VR91b, VR91a]. The FCS is a genetic based machine learning
system which integrates a fuzzy RB, the bucket-brigade learning algorithm (the
learning algorithm commonlly used by the classier systems) and a GA [GS92, VR91b,
VR91a]. Each classier represents a fuzzy rule and the FCS employs a GA to evolve
adequate rules running over the population of classiers searching new improved ones.
Finally, to point out that although the application of GA for designing fuzzy systems
is recent, it has an increasing concern over the last years that will allow to obtain
fruitful researches in the building of fuzzy logic based intelligent systems.
cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
References
[BK94] Back T. and Kursawe F. (July 1994) Evolutionary algorithms for fuzzy
logic: A brief overview. In Proc. Fifth International Conference on Information
Processing and Management of Uncertainty in Knowledge Based Systems
(IPMU'94), pages 659{664. Paris.
[BMU95] Braunstingl R., Mujika J., and Uribe J. P. (March 1995) A wall following
robot with a fuzzy logic controller optimized by a genetic algorithm. In Proc.
Fourth IEEE International Conference on Fuzzy Systems (FUZZ-IEEE'95),
pages 77{82. Yokohama.
[BN95] Bolata F. and Nowe A. (March 1995) From fuzzy linguistic specications to
fuzzy controllers using evolution strategies. In Proc. Fourth IEEE International
Conference on Fuzzy Systems (FUZZ-IEEE'95), pages 1089{1094. Yokohama.
[Bon93] Bonarini A. (September 1993) Elf: Learning incomplete fuzzy rule sets
for an autonomous robot. In Proc. First European Congress on Fuzzy and
Intelligent Technologies (EUFIT'93), pages 69{75. Aachen.
[Cas95] Castro J. L. (March 1995) Fuzzy logic controllers are universal
approximators. To appear in IEEE Transactions on Systems, Man and
Cybernetics.
[CCC+94] Cardenas E., Castillo J. C., Cordon O., Herrera F., and Peregrn
A. (January 1994) In uence of fuzzy implication functions and defuzzication
methods in fuzzy control. BUSEFAL 57: 69{79.
[CCC+95] Cardenas E., Castillo J. C., Cordon O., Herrera F., and Peregrn A.
(January 1995) Applicability of t-norms in fuzzy control. BUSEFAL 61: 28{37.
[CF94] Carse B. and Fogarty T. C. (1994) A fuzzy classier system using the
pittsburgh approach. In Davidor Y., Schwefel H. P., and Maanner R. (eds)
Parallel Problem Solving from Nature - PPSN III, pages 260{269. Springer-
Verlag, Berlin.
[CHP95a] Cordon O., Herrera F., and Peregrn A. (March 1995) Applicability of
the fuzzy operators in the design of fuzzy logic controllers. Technical Report
DECSAI-95111, University of Granada, Department of Computer Science and
Articial Intelligence, University of Granada, Granada, Spain.
[CHP95b] Cordon O., Herrera F., and Peregrn A. (July 1995) T-norms versus
implication functions as implication operators in fuzzy control. To appear
in Proc. Sixth International Fuzzy Systems Association World Congress
(IFSA'95).
[CV93] Cooper M. G. and Vidal J. J. (1993) Genetic design of fuzzy logic
controllers. In Proc. Second International Conference on Fuzzy Theory and
Technology (FTT'93). Durham.
[DHR93] Driankov D., Hellendoorn H., and Reinfrank M. (1993) An Introduction
to Fuzzy Control. Springer-Verlag.
[FTH94] Fathi-Torbaghan M. and Hildebrand L. (July 1994) Evolutionary
strategies for the optimization of fuzzy rules. In Proc. Fifth International
Conference on Information Processing and Management of Uncertainty in
Knowledge Based Systems (IPMU'94), pages 671{674. Paris.
Sample Contributed Book
Editor Jenny Smith c 1993 John Wiley & Sons Ltd.
cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
 24 REFERENCES
[GS92] Geyer-Schulz A. (1992) Fuzzy classier systems. In Lowen R. (ed) Fuzzy
Logic: State of the Art. Kluwer Academic Publishers, Dordretch.
[GSR94] George S. M., Saxena A., and Rambabu P. (September 1994) Genetic
algorithm in the aid of fuzzy rule deduction. In Proc. Second European
Conference on Soft Computing and Intelligent Technologies (EUFIT'94), pages
1130{1133. Aachen.
[HLV94] Herrera F., Lozano M., and Verdegay J. L. (July 1994) Generating fuzzy
rules from examples using genetic algorithms. In Proc. Fifth International
Conference on Information Processing and Management of Uncertainty in
Knowledge Based Systems (IPMU'94), pages 675{679. Paris.
[HLV95a] Herrera F., Lozano M., and Verdegay J. L. (February 1995) A learning
process for fuzzy control rules using genetic algorithms. Technical Report
DECSAI-95108, University of Granada, Department of Computer Science and
Articial Intelligence, University of Granada, Granada, Spain.
[HLV95b] Herrera F., Loz-
ano M., and Verdegay J. L. (1995) Tuning fuzzy logic controllers by genetic
algorithms. International Journal of Approximate Reasoning 12: 293{315.
[HMB93] Harris C. J., Moore C. G., and Brown M. (1993) Intelligent Control.
Aspects of Fuzzy Logic and Neural Nets. World Scientic Publishing.
[HP94] Homann F. and Pster G. (September 1994) Automatic design
of hierarchical fuzzy controllers using genetic algorithms. In Proc.
Second European Conference on Soft Computing and Intelligent Technologies
(EUFIT'94), pages 1516{1522. Aachen.
[HP95] Homann F. and Pster G. (July 1995) A new learning method for the
design of hierarchical fuzzy controllers using messy genetic algorithms. To
appear in Proc. Sixth International Fuzzy Systems Association World Congress
(IFSA'95).
[HT94] Hwang W. R. and Thompson W. E. (June 1994) Design of fuzzy logic
controllers using genetic algorithms. In Proc. Third IEEE International
Conference on Fuzzy Systems (FUZZ-IEEE'94), pages 1383{1388. Orlando.
[HTS93] Hu H.-T., Tai H.-M., and Shenoi S. (1993) Fuzzy controller design using
cell mappings and genetic algorithms. In Proc. Second International Conference
on Fuzzy Theory and Technology (FTT'93). Durham.
[IFSA95] Ishigami H., Fukuda T., Shibata T., and Arai F. (May 1995)
Structure optimization of fuzzy neural networks by genetic algorithm.
Fuzzy Sets and Systems 71(3): 257{264.
[Kar91a] Karr C. (March 1991) Applying genetics. AI Expert pages 38{43.
[Kar91b] Karr C. (February 1991) Genetic algorithms for fuzzy controllers.
AI Expert pages 26{33.
[KB93] Kropp K. and Baitinger U. G. (September 1993) Optimization of fuzzy
logic controller inference rules using a genetic algorithm. In Proc. First
European Congress on Fuzzy and Intelligent Technologies (EUFIT'93), pages
1090{1096. Aachen.
[KKS85] Kiszka J. B., Kochanska M. E., and Sliwomska D. S. (January 1985)
The in uence of some fuzzy implication operators on the accuracy of a fuzzy
model - parts i and ii. Fuzzy Sets and Systems 15: 111{128, 223{240.
[Lee90] Lee C. C. (March 1990) Fuzzy logic in control systems: Fuzzy logic con-
troller - parts i and ii. IEEE Transactions on Systems, Man and Cybernetics
20(2): 404{435.
[LM94] Liska J. and Melsheimer S. S. (June 1994) Complete design of fuzzy
logic systems using genetic algorithms. In Proc. Third IEEE International
Conference on Fuzzy Systems (FUZZ-IEEE'94), pages 1377{1382. Orlando.
[LP94] Leitch D. and Probert P. (1994) Context depending coding in genetic
algorithms for the design of fuzzy systems. In Proc. IEEE/Nagoya University
WWW on Fuzzy Logic and Neural Networks/Genetic Algorithms. Nagoya.
[LT93] Lee M. A. and Takagi H. (July 1993) Embedding apriori knowledge into

cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)
 REFERENCES 25
an integrated fuzzy system design method based on genetic algorithms. In Proc.
Fifth International Fuzzy Systems Association World Congress (IFSA'93),
pages 1293{1296. Seoul.
[LW91] Liaw C. M. and Wang J. B. (July 1991) Design and implementation
of a fuzzy controller for a high performance induction motor drive.
IEEE Transactions on Systems, Man and Cybernetics 21(4): 921{929.
[MA75] Mamdani E. H. and Assilian
S. (1975) An experiment in linguistic systhesis with a fuzzy logic controller.
International Journal of Man-Machine Studies 7: 1{13.
[NHW92] Nomura H., Hayashi I., and Wakami N. (1992) A learning method
of simplied fuzzy reasoning by genetic algorithm. In Proc. International
Fuzzy Systems and Intelligent Control Conference (FSIC'92), pages 236{245.
Louisville.
[NL94] Ng K. C. and Lee Y. (June 1994) Design of sophisticated fuzzy logic
controllers using genetic algorithms. In Proc. Third IEEE International
Conference on Fuzzy Systems (FUZZ-IEEE'94), pages 1708{1712. Orlando.
[PB93] Parodi A. and Bonelli P. (July 1993) A new approach to fuzzy classier
system. In Proc. Fifth International Conference on Genetic Algorithms
(ICGA'93), pages 223{230.
[SFH95] Shijojima K., Fukuda T., and Hasegawa Y. (May 1995) Self-tuning fuzzy
modeling with adaptive memebership function, rules, and hierachical structure
based on genetic algorithm. Fuzzy Sets and Systems 71(3): 295{309.
[SK94] Satyadas A. and Krishnakumar K. (June 1994) Ga-optimized fuzzy
controller for spacecraft attitude control. In Proc. Third IEEE International
Conference on Fuzzy Systems (FUZZ-IEEE'94), pages 1979{1984. Orlando.
[SKG93] Surmann H., Kanstein A., and Goser K. (September 1993) Self-
organizing and genetic algorithms for an automatic design of fuzzy control and
decision systems. In Proc. First European Congress on Fuzzy and Intelligent
Technologies (EUFIT'93), pages 1097{1104. Aachen.
[Thr91] Thrift P. (1991) Fuzzy logic synthesis with genetic algorithms. In Proc.
Fourth International Conference on Genetic Algorithms (ICGA'91), pages 509{
513.
[TS85] Takagi T. and Sugeno M. (January 1985)
Fuzzy identication of systems and its applications to modeling and control.
IEEE Transactions on Systems, Man and Cybernetics 15(1): 116{132.
[VM95] Velasco J. R. and Magdalena L. (July 1995) Genetic learning applied to
fuzzy rules and fuzzy knowledge bases. To appear in Proc. Sixth International
Fuzzy Systems Association World Congress (IFSA'95).
[VR91a] Valenzuela-Rendon M. (1991) The fuzzy classier system: A classier
system for continuously varing variables. In Proc. Fourth International
Conference on Genetic Algorithms (ICGA'91), pages 346{353.
[VR91b] Valenzuela-Rendon M. (1991) The fuzzy classier system: Motivations
and rst results. In Manner R. and Manderick B. (eds) Parallel Problem Solving
from Nature - PPSN II, pages 330{334. Springer-Verlag, Berlin.
[Zad65] Zadeh L. A. (1965) Fuzzy sets. Information and Control 8: 358{353.
[Zad73] Zadeh
L. A. (1973) Outline of a new approach to the analysis of complex systems
and decision processes. IEEE Transactions on Systems, Man and Cybernetics
3: 28{44.

cbook 2/9/1997 17:36|PAGE PROOFS for John Wiley & Sons Ltd (jwbook.sty v3.0, 12-1-1995)