Adaptation of Genetic Algorithm

Parameters ?Based on Fuzzy Logic

Controllers
Francisco Herrera and Manuel Lozano
Department of Computer Science and Arti cial Intelligence
University of Granada
18071 - Granada, Spain

Abstract. The genetic algorithm behaviour is determined by the exploitation

and exploration relationship kept throughout the run. Adaptive genetic algo-
rithms have been built for inducing exploitation/exploration relationships that
avoid the premature convergence problem and improve the nal results. One of
the most widely studied adaptive approaches are the adaptive parameter setting
techniques. In this paper, we study these techniques in depth, based on the use
of fuzzy logic controllers. Furthermore, we design and discuss an adaptive real-
coded genetic algorithm based on the use of fuzzy logic controllers. Although
suitable results have been obtained by using this type of adaptive technique, we
report some re ections on open problems that still remain.

Keywords. Exploitation/exploration relationship, adaptive genetic algorithms,

fuzzy logic controllers.

1 Introduction
GA behaviour is strongly determined by the balance between exploiting what
already works best and exploring possibilities that might eventually evolve into
something even better. The loss of critical alleles due to selection pressure, the
selection noise, the schemata disruption due to crossover operator, and poor
parameter setting may make this exploitation/exploration relationship (EER)
disproportionate and produce the lack of diversity in the population ([38, 43,
40]). Under these circumstances a preeminent problem appears: the premature
convergence problem.
Some tools for monitoring the EER were proposed in order to avoid the pre-
mature convergence problem and improve GA performance. These tools include
modi ed selection and crossover operators (see [43] and [40]) and optimization
of parameter settings ([24, 49]). However, nding robust genetic operators or
parameter settings that allow the premature convergence problem to be avoided
in any problem is not a trivial task, since their interaction with GA performance
? This research has been supported by DGICYT PB92-0933
is complex and the optimal ones are problem dependent ([3]). Furthermore, dif-
ferent operators or parameter con gurations may be necessary during the course
of a run for inducing an optimal EER. For these reasons, many adaptive tech-
niques were suggested for changing the GA con guration based on the current
information about the search space in order to o er the most appropriate explo-
ration/exploitation behaviour. One of the these techniques lies in the application
of fuzzy logic controllers for adjusting the control parameter of GAs. The goal
of this paper is to report on an extensive study of the parameter setting adap-
tation, based on the use of fuzzy logic controllers. For doing this, it is set up as
follows: in Section 2, we review di erent parameter setting adaptive techniques
presented in the literature, in Section 3, we tackle the study of adaptive GAs
based on fuzzy logic controllers, in Section 4, we present di erent diversity mea-
sures that may be used by these FLCs, in Section 5, we review the applications
of the FLCs for controlling GAs, in Section 6, we design an adaptive real-coded
GA, in Section 7, open problems that still remain about the topic are discussed,
nally some conclusions are presented in Section 8.

2 Adaptive Genetic Algorithms

In this section we brie y study the classi cation of the di erent adaptive methods
proposed in the literature and we go into the parameter setting adaptation in
depth.
2.1 Classi cations of Adaptive Genetic Algorithms
Adaptive methods may be classi ed following two di erent ways. The rst one
takes into account the GA issues that are adapted throughout the GA run:
{ adaptive parameter settings ([32, 41, 7, 63, 11, 20, 14, 59, 60, 12, 3, 5, 38, 53,
1, 55, 34]),
{ adaptive genetic operators ([42, 30, 31, 44]),
{ adaptive genetic operator selection ([48, 54, 52, 62, 51]),
{ adaptive representation ([46, 61, 50]) and
{ adaptive tness function ([22, 42, 35, 45]).
The other one, suggested by Spears ([54]), concerns the interaction between
the adaptive method and the evolutive process:
{ Tightly Coupled. The adaptation is driven by the internal forces of evolution
itself. There are three levels where the adaptation may be carried out ([51]):
the level of individuals, the level of subpopulations and the level of popula-
tion. The level implicitly de nes the time interval when the adaptation takes
place. If, for instance, populations are used for adaptation, then a minimum
number of generations is needed for evaluating the population. In contrast,
individuals are evaluated after each generation. This approach is elegant and
straightforward, since no new adaptive mechanism is needed, however, two
possible problems may arise ([54]):
  The coupling of the two search spaces could hinder the adaptive mecha-
nism. As the population evolves, diversity will be lost in both spaces. If
diversity has been lost, adaptation may be dicult to accomplish.
 This mechanism may only work well with large population sizes, since
the population is being used to sample the space of GAs.
{ Uncoupled. A totally separate adaptive mechanism adjusts the GA. An un-
coupled approach does not rely upon the GA for the adaptive mechanism.
While this may alleviate the problems associated with coupling, such mech-
anisms appear to be dicult to construct, and involve a lot of separate
bookkeeping. In [23], Goldberg wrote:
"Nature acts in a distributed fashion without central coordination
or control. Such distributed processing gives nature and genetic al-
gorithms a large measure of their global perspective by permitting
di erent solutions to be explored in parallel".
With these words he pointed out his disagreement with the uncoupled mech-
anisms, since they break the original GA philosophy. Moreover, he suggested
that the loss of explicit parallelism often implied by central control may re-
duce some portion of the search to serial hillclimbing.
{ Loosely Coupled. The GA is partially used for the adaptive mechanism, i.e.,
either the population or the genetic operators are used in some fashion.
Next, we review the adaptive GAs based on adaptive parameter settings.
This type of adaptive GAs have been widely studied.
2.2 Parameter Setting Adaptation
The GA control parameter settings such as mutation probability (pm ), crossover
probability (pc ) and population size (N ) are key factors in the determination of
the exploitation versus exploration tradeo . It has long been acknowledged that
they have a signi cant impact on GA performance [24]. If poor settings are used,
the EER may not be reached in a pro table way; the GAs' performance shall
be severely a ected [14]. Next, we report on some adaptive approaches that are
described in the literature for adjusting these parameters.

2.2.1. Adapting the Mutation Probability

At rst glance, it would seem that premature convergence is best treated by
an increase in the mutation probability to restore the lost alleles ([11]). However,
there are some problems associated with this solution ([8]): 1) the high in ux
of material would disrupt the ongoing genetic search and 2) this task attempts
to recover from convergence rather than to avoid it, and hence, may su er from
the loss of potentially valuable, genetic combinations. More suitable proposals
for enforcing diversity through pm have been proposed. Next, we review these.
A direction followed by GA research for the variation of pm lies in the spec-
i cation of an externally speci ed schedule which modi es it depending on the
time, measured by the number of generations. A time-dependency of pm was rst
suggested by Holland ([32]) himself, although he did not give a detailed choice
of the parameter for the time-dependent reduction of pm . Later on, in [20] and
[12], the variation of pm was proposed by decreasing it exponentially during the
GA run. A theoretical argument for introducing a time-dependent schedule of
pm is presented in [3].
Mauldin ([41]) de ned a uniqueness value as the minimumHamming distance
allowed between any o spring and all existing chromosomes in the population. If
a new chromosome is closer than this to an existing chromosome, the allele values
which match are randomly changed in the o spring until the uniqueness value
is achieved. This implicitly represents an increase in the mutation probability.
To ensure that the GA is allowed to converge, the uniqueness value is decreased
as the search proceeds from a value of k to 1.
The idea of sustaining genetic diversity through mutation is presented in [59,
60] as well. Mutation is invoked in response to increasing genetic homogeneity
in the population. Population homogeneity is indirectly monitored by measuring
the Hamming distance between the two parents during reproduction. The more
similar the two parents, the more likely that mutation would be applied to bits
in the o spring created by recombination.
In [3, 5], a coupled adaptive method for mutation probabilities was adopted.
The principal features of this model are 1) each position of each chromosome
has associated a mutation probability, 2) these probabilities are incorporated
into the genetic representation of the chromosomes encoded as bitstrings and 3)
they are also subject to mutation and selection, i.e., they undergo evolution as
well as the chromosomes.

2.2.2 Adapting the Population Size

The population size is one of the most important choices faced by any user of
GAs and may be critical in many applications. If the population size is too small,
the GA may converge too quickly, whereas if it is too large, the GA may waste
computational resources. Next, we present some approaches for the adaptation
of N .
One of the rst attempts for controlling N was made by Baker in [7]. Baker
emphasized the importance of preventing rapid convergence rather than recover-
ing from it. He observed that rapid convergence often occurs after a chromosome
or small group of chromosomes contributes a large number of o spring through
selection. Since population is nite, a large number of o spring for one chro-
mosome means fewer o spring for the rest of the population. When too many
chromosomes have no o spring at all, the result is a rapid loss of diversity and
premature convergence. In order to recognize rapid convergence before the ge-
netic material is lost, Baker presented a measure, called percent involvement, as
the percentage of the current population which contributes o spring for the next
generation. Baker proposed a dynamic population size method that enforces a
lower bound on the percent involvement. This is done by adding chromosomes
to the population until the acceptable value for the percentage involvement is
reached. The population size decreases during periods of slow convergence.
A di erent proposal is attempted in [1]. In this paper a GA with varying pop-
ulation size (GAVaPS) was discussed. This algorithm introduces the concept of
age of a chromosome, which is equivalent to the number of generations that such a
chromosome stays alive. GAVaPS does not operates with a selection mechanism;
the age of the chromosomes replaces the concept of selection. When a chromo-
some is born a lifetime is assigned to it. The death of this chromosome occurs
when its age exceeds its lifetime value. In the calculation of the lifetime value the
current state of the GA is taken into consideration. This state is described by
the average, maximal and minimal tness values in the current population and
the maximal and minimal found so far. Any lifetime calculation strategy should
assign the higher lifetime values to chromosomes having above-average tness
values. Di erent proposals were implemented and compared. None of them stood
out as the best with regards to the results and its computational cost. Hence, it
was pointed out that the choice of the optimal strategy of lifetime calculation is
still an open problem and it deserves further investigation.
Finally, we point out that in [53], an algorithm was presented, which adjusts
the population size with respect to the probability of selection error.

2.2.3 Adapting the Crossover Probability

In a paper by Wilson ([63]) on GAs for classi er systems, he claimed that
dynamically varying the crossover probability may be bene cial. His approach
is based on the use of an entropy measure over the population. Crossover prob-
ability is adjusted up slightly if entropy is falling and down if entropy is at or
rising. Another attempt for adapting the crossover probability is presented by
Booker in [11]. He pointed out that the crossover operator is the key point for
solving the premature convergence problem. Furthermore, he proposed varying
the crossover probability depending on the percent involvement presented by
Baker ([7]) (see previous subsection). Every percentage change in percent in-
volvement will be countered with an equal and opposite percentage change in
the crossover probability. Experiments carried out showed that oine and online
measures were improved with this model.

2.2.4 Adapting both the Crossover and Mutation Probabilities

In [38], two diversity functions were de ned for describing GA behaviour and
were used for handling the pc and pm parameters in order to induce an ideal EER.
One of these diversity functions measures the diversity between chromosomes
(BC) in the population and the other measures the diversity between the alleles
(BA) in all the chromosomes. The former is concerned with the eciency of
genetic operators and the latter expresses the state of convergence. Li et al.
([38]) proposed a dynamic GA that can approximate an ideal genetic procedure,
which is an approach so that the EER can be automatically adjusted and the
genetic operators will be most ecient. In order to construct a dynamic GA, the
whole procedure is divided into three stages: initiation, search and re nement. In
the initial stage, the values of the diversity measures are dependent on the initial
conditions of the GA, so the diversity measure's behaviour is unpredictable. The
control parameters are kept at their initial values at this stage. In the search
stage, the control parameters are designed to vary and to guarantee a broad
search and ecient exploitation. If the BC value is less than a minimum critical
limit, the c is adjusted upwards slightly and if the BC value is larger than a
p

maximum critical limit, the c is forced downwards slightly. According to the

p

BA value, the m was varied from 0.001 to 0.2 to have sucient alleles in the
p

population. This stage was designated as the main part of the GA, i.e, half of
the generation time. Finally, in the last stage, re nement, the balance must be
designed towards exploitation and the search is forced in the local regions where
there are higher probabilities of nding the optimum solutions. To do so, m p

decreases from the current value to a minimum value of 0 001 and c is help
: p

at a constant value of 0 6. The experiments showed that dynamic GA was near

:

an ideal behaviour de ned by Li et al. with respect to the level of diversity

generated.
In [55], the use of adaptive probabilities of crossover and mutation was rec-
ommended to achieve the twin goals of maintaining diversity in the population
and sustaining the convergence capacity of the GA. AGA was proposed, an
adaptive GA where c and m are varied depending on the tness values of the
p p

solutions. Each chromosome has its own c and m . Both are proportional to
p p

( best ? ) ( best ? ), where is the chromosome's tness, best is the population
f f = f f f f

maximum tness and  is the mean tness. In this way, high- tness solutions are
f

protected, whilst solutions with subaverage tnesses are totally disrupted. AGA
increases c and m when the population tends to get stuck at a local optimum
p p

and decreases them when the population is scattered in the solution space. The
factor best ?  is used by AGA as a measure for detecting convergence.
f f

In [14], a real-coded GA is implemented by using two types of crossover oper-

ators, guaranteed uniform crossover and guaranteed average, and three types of
mutation operators, guaranteed mutation, guaranteed big creep and guaranteed
little creep. These operators o er a wide range of exploration and exploitation
levels. Each operator is given an initial application probability. For each repro-
duction event, a single operator is selected probabilistically according to the
set of operator probabilities, as opposed to the technique used more commonly,
in which both crossover and mutation could be applied to the same individual
during reproduction. An adaptive process provides for the alteration of opera-
tor probabilities in proportion to the tnesses of chromosomes created by the
operators during the course of a run. Operators which create and cause the gen-
eration of better chromosomes are allotted higher probabilities, i.e., they should
be used more frequently. On the other hand, operators producing o spring with
a tness which is lower than that of the parents should be used less frequently.
In [34], a similar model is presented for a steady-state GA. This method adjusts
the operators' probabilities after the generation of each o spring, rather than at
longer intervals.

3 Adaptive GAs based on Fuzzy Logic Controllers

Currently, a certain body of expertise, experience and knowledge on GAs has be-
come available as a result of empirical studies conducted over a number of years.
This human expertise and knowledge would be very useful for increasing the
capabilities of GAs to reach a suitable EER for avoiding premature convergence
and improve GA behaviour. However, generally, too much of this information
is vague, incomplete, or ill-structured to a certain degree, which causes it to be
rarely applied. This latter feature suggests the use of fuzzy logic-based tools for
dealing with this type of knowledge. One application of fuzzy logic (FL) that is
useful for controlling GAs following control strategies underlying in the human
expertise and knowledge are fuzzy logic controllers (FLCs). FLCs implement an
expert operator's approximate reasoning process in the selection of a control
action. An FLC allows one to qualitatively express the control strategies based
on experience as well as intuition. These control strategies may be expressed in
a form that permits both computers and humans to share them eciently.
Next, we study the application of the FLCs for controlling GAs, before brie y
describing the generic structure of FLCs.

3.1 Description of the FLCs

An FLC is composed by a knowledge base, that includes the information given

by the expert in the form of linguistic control rules, a fuzzi cation interface,
which has the e ect of transforming crisp data into fuzzy sets, an Inference
System, that uses them together with the knowledge base to make inference by
means of a reasoning method, and a defuzzi cation interface, that translates the
fuzzy control action thus obtained to a real control action using a defuzzi cation
method. The generic structure of an FLC is shown in Figure 1 ([36]).

Knowledge Base

Fuzzification Defuzzification
Inference System
Interface Interface

State Variables Control Variables

Controlled System

Fig. 1. Generic structure of an FLC.

 The knowledge base encodes the expert knowledge by means of a set of fuzzy
control rules. A fuzzy control rule is a conditional statement with the form
"if a set of conditions are satis ed then a set of consequences can be inferred"
in which the antecedent is a condition in its application domain, the consequent
is a control action to be applied in the controlled system (notion of control rule)
and both antecedent and consequent are associated with fuzzy concepts, that is,
linguistic terms (notion of fuzzy rule).
The knowledge base is composed of two components, a data base, containing
the de nitions of the fuzzy control rules linguistic labels, i.e., the membership
functions of the fuzzy sets specifying the meaning of the linguistic terms, and
a rule base, constituted by the collection of fuzzy control rules representing the
expert knowledge.
3.2 Application of the FLCs for Controlling GAs
Applications of FLCs for controlling GAs are to be found in [37, 64, 65, 2, 27,
10, 56]. The main idea is to use an FLC whose inputs are any combination of
GA performance measures or current control parameters and whose outputs are
GA control parameters. Current performance measures of the GA are sent to
the FLC, which computes new control parameters values that will be used by
the GA. Figure 2 shows this process. Clearly, under Spears' classi cation, this
adaptive technique is uncoupled.

Adaptive GA

Performance measures
GA control parameters APPLICATION
FLC GA
TASK
GA control parameters

Fig. 2. Structure of an Adaptive GA based on FLCs.

For designing a FLC that controls a GA, the following steps are needed:
1. De ne the inputs and outputs
Inputs should be robust measures that describe GA behaviour and the e ects
of the genetic setting parameters and genetic operators. In [56], some possible
inputs were cited: diversity indices, maximum, average and minimum tness,
etc. In [37] and [64, 65], it is suggested that current control parameters may
also be considered as inputs. Since an adaptive GA based on FLCs should
maintain an adequate EER for avoiding premature convergence, we think
that diversity measures stand out as important input candidates due to the
 fact that they describe the convergence state of the population. In Section
4, a broad study of this type of measures is carried out.
Outputs indicate values of control parameters or changes in these parame-
ters ([37]). In [56], the following outputs were reported: pm , pc, N , selective
pressure, the time the controller must spend in a target state in order to be
considered successful, the degree to which a satisfactory solution has been
obtained, etc.
2. De ne the data base
Each input and output should have an associated set of linguistic labels. The
meaning of these labels is speci ed through membership functions of fuzzy
sets. So, it is necessary that every input and output have a bounded range
of values in order to de ne these membership functions over it.
3. Obtain the rule base
After selecting the inputs and outputs and de ning the data base, the fuzzy
rules describing the relations between them should be de ned. There are
di erent ways to do so:
(a) using the experience and the knowledge of the GA experts or
(b) using an automatic learning technique for those cases where knowledge
or expertise are not available.
In [37], an automatic fuzzy design technique is used for obtaining the rule
base (and the data base as well). The technique is very similar to the meta-
GA of Grefentette [24]. By using an automatic technique, relevant relations
and membership functions may be automatically determined and may of-
fer insight for understanding the complex interaction between GA control
parameters and GA performance.

4 Diversity Measures
There are two types of diversity measures that describe the state of the popu-
lation: genotypic diversity measures (GDMs) and phenotypic diversity measures
(PDMs). GDMs involve the genetic material held in the population whereas
PDMs concern the tness of the chromosomes. Next, we review each one of
these measure types.

4.1 Genotypic Diversity Measures

GDMs describe variation or lack of similarity between the genetic material,

e.g., alleles, chromosomes, etc. Di erent ways have been followed for de ning
them, such as using allele frequency measures, statistical dispersion measures,
histograms over the action interval of the genes, Hamming distance, Euclidean
distance, entropy measures, etc. Next, we summarize some of these.
4.1.1 GDMs Based on Allele Frecuencies
Diversity measures based on allele frequencies are common for binary-coded
GAs (BCGAs). De Jong, in [15], proposed two diversity measures based on allele
frequencies, the lost alleles, a count of the gene positions which are completely
converged, i.e., the number of gene positions for which the entire population has
the same allele, and the converged alleles, a count of the gene positions which
are at least 95% converged, i.e., the number of gene positions for which 95% of
the population has the same allele. Baker ([8]) claimed that these measures do
not indicate the overall convergence in the population, so he presented the com-
mitment, which is the average convergence of each gene position. For example, if
the individuals are composed of 5 genes, and each gene has converged by 100%,
80%, 95%, 50%, and 70%, respectively, then the commitment is 100+80+95+50+70
5 .
Collins et al. ([13]) measure the allele diversity of a BCGA population by com-
paring the current allele frequencies with the expected allele frequencies of a
maximally diverse population. Maximum diversity occurs when each allele fre-
quency is 0.5, since each gene position has two possible alleles. The diversity of
alleles at position i is de ned as
D = 1 ? 4(0:5 ? frec )
i i

where frec is the current frequency of the 0 allele at position i. D can range
i i
from 0, which indicates complete convergence, to 1, which indicates the maxi-
mum possible genetic diversity. The diversity of the population is calculated as
the average value of the D . Mauhfoud ([40]) presented a general model of diver-
i
sity measures based on the idea of comparing a frequency distribution associated
with a particular feature of the elements of the population and a distribution
representing the full diversity with respect to this feature. The diversity mea-
sures of Mauhfoud take on real values in the range from 0 to MAX , with 0
indicating no diversity, higher values indicating higher levels of diversity, and
MAX indicating maximal diversity. Mauhfoud demostrated that the proposals
of De Jong and Collins et al. are instances of his model.

4.1.2 GDMs Based on Histograms

In [33], a measure of diversity, called total allele lost, for a population with
integer-coded chromosomes, was presented. The action interval of each gene,
[?10; 10], was divided into seven subintervals, I1 = [?10; ?8], I2 = [?7; ?5],...,I7 =
[8; 10]. A histogram, hist(
;
), was computed such as for each position i of the
chromosomes and each interval I , hist(i; j ) stores the number of alleles at po-
j
sition i that belong to the subinterval I . For each position i, a measure, AL(
),
j
called allele lost, was de ned as
X
7
AL(i) = number of zeros in hist(j, i):
j =1
 So, the total allele lost was de ned as the sum of the AL(
) measures. This
measure has a bounded range of values, therefore it is adequate as the input for
an FLC. A similar measure is de ned in [50] for detecting convergence of binary-
coded GAs that works on continuous spaces. The action interval associated with
each coded parameter in the chromosomes is divided into four zones, which
have the same length, called quarters. A histogram is computed on the quarters,
with the parameter's available values in the population. Then, the intersections
of the quarters are considered and the associated values of the histogram are
calculated. With these intervals, statistics are created. If the sampling rate of the
most sampled intersection interval exceeds a trigger threshold, then convergence
in such an interval is assumed.

4.1.3 GDMs Based on Hamming Distance

Other approaches for measuring the diversity of BCGAs are based on the
Hamming distance. Louis et al. ([39]) calculated the Hamming distance between
all non-redundant combinations of two strings and take an average. Whitley et
al. ([61]) proposed measuring the Hamming distance between the worst and the
best chromosomes in the population for monitoring population diversity.

4.1.4 GDMs Based on Euclidean Distance

For real-coded GAs, it seems interesting to dispose of GDMs based on the
Euclidean distance between chromosomes. We propose one of these measures,
which is based on the Euclidean distances of the chromosomes in the population
from the best one. It is denoted as ED.

ED = d d ? ?dmin
max dmin
 P N
where d = N i=1 d(Cbest; Ci), dmax = maxfd(Cbest; Ci)jCi 2 P g and dmin =
1
minfd(Cbest; Ci)jCi 2 P g. P denotes the population of real-coded chromosomes.
The range of ED is [0; 1]. If ED is low, most chromosomes in the population
are concentrated around the best chromosome and so convergence is achieved. If
DE is high most chromosomes are not biased toward the current best element.

4.1.5 GDMs Based on Dispersion Statistical Measures

Diversity measures for BCGAs based on dispersion statistical measures have
been considered as well. This is the case of Li et al ([38]), which proposed two
diversity measures: the diversity between the chromosomes (BC) and the diversity
between the alleles (BA). If it is considered that the population is a matrix, in
which a row expresses a chromosome and a column an allele, then BC and BA
are de ned as follows:
 P Si 2

? diversity between the chromosomes BC

i ? LS2N
P= j Sj S2
2
L
N ?1
? diversity between the alleles BA = L?1 N ? LN
where N is the population size, L the length of the chromosomes, S is the sum
of genes '1', Si expresses the sum over a row i and Sj expresses the sum over a
column j . Li et al. showed that BC and BA show the following properties:
1. If the population is homogeneous in either 0 or 1, BC and BA are zero.
2. When all the chromosomes in the population are identical, BC is zero and
BA holds a constant value that is dependent on how many genes 1 are in a
chromosome.
3. BA is independent from the crossover operator.
4. BC and BA are always less than NL and NL , respectively..
These measures show clear limits and their ideal behaviour throughout the
GA run was studied in depth by Li et al. in [38].
In [27], other instances of diversity measures of the same type were considered
for real coding. They were based on the the variance average of the chromosomes
(VAC) and the average variances of the alleles P (AVA):
N (S ?S)2
i
? The variance average chromosomes V AC PNj=1 PLi=1i=1(SNij ?Sj )2
=
? The average variances alleles AV A = LN
where Si , i = 1, ..., N, denotes the chromosome i, Sj , j = 1, ..., L, denotes the
vector of genes with position j in the population, Sij , denotes the gene with
position j in the chromosome i, and
PL PL PN PNi
Si = j =1 Sij , S = i=1 j =1 Sij =1 Sij :
and Sj =
L LN N
The properties of these diversity measures are the following: they are indif-
ferent to mutual exchange of two chromosomes in a population; and when all the
chromosomes in a population are almost identical, the VAC and AVA measures
take low values.
In [9], four diversity measures were de ned in a similar fashion for real coding
as well. They were called: the within-individual diversity (Wb ), the between-
individual diversity (Bb ), the within-gene diversity (Wg ) and the between-gene
diversity (Bg ). A total diversity measure was de ned as the sum of the previous
four measures.

4.1.6 GDMs Based on Entropy Measures

Finally, we point out that some authors used entropy measures for describing
the convergence state of the population ([63, 25]).
4.2 Phenotypic Diversity Measures

The rst measures based on tness function presented for describing the conver-
gence state of the population are the percent involvement ([7]) (See Subsection
2.2.2) and the reproduction rate ([8]) (de ned in the same way). The principal
feature of this type of measures is that they allow premature convergence to be
predicted rather than being detected.
Other phenotypic diversity measures have been proposed. Most of them are
de ned through the combinations of some of the following measures: average
tness ( ), the best tness ( best) and the worst tness ( worst ). Srinivas et
f f f

al. ([55]) used best ?  for detecting convergence. This measure is likely to be
f f

less for a population that has converged to an optimum solution than that for
a population scattered across the solution space. Lee et al. ([37]) propose two
performance measures that may be considered as phenotypic diversity measures:

P DM1 = best
 and
f
P DM2 = f
:
f fworst

P DM1 andP DM 2 belong to the interval [0 1]. If they are near to 1, convergence
;

has been reached, whereas if they are near to 0, the population shows a high
level of diversity.
In [2], a measure called tness range was de ned as the di erence between
the tness values of the best and worst chromosomes.
Another phenotypic diversity measure is the Online average ([15]), which is
the average tness of all the chromosomes tested during the search. This measure
would be appropriate in situations where the testing of every chromosome must
be paid for and penalizes GAs that must test many poor chromosomes before
locating good ones. To do well in online measures, a GA must quickly decide
where the best values lie and concentrate its search there ([49]).
Finally, in [42], another phenotypic diversity measure called was de ned
span

as follows:
1
q P
N (f ? f)2
i=1 i
span =
N? 1
1 N P :

n i=1 fi
p
The maximum value of is ? 1.
span N span was used in [42] as a measure
of the diculty of the tness function.

5 FLC Applications for Controlling GAs

Next, we review di erent models of application of FLC to the control of GAs
described in the literature.
5.1 Dynamic Parametric GAs
In [37], the dynamic parametric GA (DPGA) was proposed, a GA that uses an
FLC for controlling GA parameters. The inputs to the FLC are any combination
of GA performance measures or current control settings, and outputs may be
any of the GA control parameters. An automatic fuzzy design technique was
proposed for obtaining both the data base and the rule base for those cases
where knowledge or expertise are not available. An example was developed by
Lee et al. ([37]); they proposed to build three FLCs. All of them use the measures
PDM1, PDM2 (see Subsection 4.2) and the change in the best tness since
the last control action as inputs. The outputs of the three FLCs are variables
that control variations in the current mutation probability, crossover probability
and population size, respectively. For example, such output for the control of
the population size, represents the degree to which the current population size
should vary. The new population size is computed by multiplying the output
value by the current population size.
The automatic fuzzy design mechanism was executed for learning both the
data base and the rule base for this example. A DPGA based on the resultant
FLCs were used in a particular problem: the inverted pendulum control task.
The results obtained exhibited better performance than a simple, static GA.

5.2 Fuzzy GAs

In [64, 65], the use of FLCs for control GAs is considered for solving two prob-
lems to which a standard GA may be subjected: very slow search speed and
premature convergence. For Xu et al. ([64, 65]) these problems are due to: 1)
control parameters not well chosen initially for a given task, 2) parameters al-
ways being xed even though the environment in which the GA operates may
be variable and 3) diculties resulting from selection of other parameters such
as population size and in understanding their in uence, both individually and in
combination, on the GA's performance. They proposed fuzzy GAs (FGA), i.e.,
GAs integrated with FLCs. FGAs may 1) choose control parameters before GA's
run, 2) adjust the control parameters on-line to dynamically adapt to new situa-
tions and 3) assist the user in accessing, designing, implementing and validating
the GA for a given task.
Experiments were carried out with an FGA, in which the adaptation of the
parameters was performed by two FLCs. Both of them had the same inputs:
generation and population size. The outputs were pc and pm . They proposed
using the rule bases shown in Table 1.
The FGA stood out as the most ecient algorithm against a standar GA in
solving the travelling salesman and other optimization problems.

5.3 Fuzzy Government

In [2], it is claimed that evolutionary algorithms require human supervision dur-
ing their routine use as practical tools for the following reasons: 1) for detecting
Table 1. Rule Bases for the Control of pc and pm , respectively
Population Size Population Size
Generation Small Medium Big Generation Small Medium Big
Short Medium Small Small Short Large Medium Small
Medium Large Large Medium Medium Medium Small Very Small
Long Very Large Very Large Large Long Small Very Small Very Small

the emergence of a solution, 2) for tuning algorithm parameter and 3) for mon-
itoring the evolution process in order to avoid undesiderable behaviour such as
premature convergence. Arnone et al. ([2]) advised that any attempt to develop
arti cial intelligence tools based on evolutionary algorithms should take these
issues into account . They proposed using FLCs for this task. They called the
collection of fuzzy rules and routines in charge of controlling the evolution of
the GA population fuzzy government. In [2], fuzzy government was applied to
the symbolic inference of formulae problem. Genetic programming was used to
solve the problem along with di erent FLCs, which dynamically adjusted the
maximum length for genotypes, acted on the mutation probability, detected the
emergence of a solution and stopped the process. The results showed that the
performance of the fuzzy governed GA was almost impossible to distinguish from
the performance of the same algorithm operated directly with human supervi-
sion.

5.4 Other Attempts

In [27], the use of FLCst was propose for monitoring the population diversity
and the EER of a real-coded GA using the GDMs based on dispersion statistical
measures, VAC and AVA. It was proposed to build two FLCs that control pc
and pm , depending on VAC and AVA, respectively. The range for the diversity
measures, VAC and AVA, was established as follows: V ACmin = 0 and V ACmax
as the maximun value obtained at the rst stage, for some value higher than it,
we automatically interchange the maximun value; and similarly for AVA, with
AV Amin = 0. The fuzzy rules suggested were the following:
If VAC is low then pc should be adjusted upwards slightly.
If VAC is high then pc should be forced downwards slightly.
and
If AVA is low then pm should be be adjusted upwards slightly.
If AVA is high then pm should be forced downwards slightly.
Another attempt to apply FLCs for improving GA behaviour is to be found
in [10].

6 An Adaptive Real-Coded GA Based on FLCs

In this Section, we propose and study an adaptive real-coded GA based on FLCs,
called ARGAF. Its principal features are the following:
 1. It applies two di erent crossover operators; one with exploitation properties
and another with exploration properties. A parameter, denoted as pe , de nes
the frequency of application of the exploitative operator.
2. It uses the linear ranking selection mechanism ([7]).
3. The pe parameter and the min parameter of the linear ranking mechanism
that determines the selective pressure produced by the selection, are ad-
justed using two FLCs. The inputs of the FLCs are a genotypic diversity
measure and a phenotypic diversity measure. The rst measure determines
the quantity of diversity in the population and the second the quality of this
diversity.
Next, we describe the conceptual foundations of ARGAF in greater detail,
then, we formulate the FLCs used in ARGAF and nally we test it through
empirical study.

6.1 Conceptual Foundations of ARGAF

In this subsection, we describe the types of crossover operators that are used by
ARGAF. Furthermore, we justify the adaptation of the pe and min parameters
and nally, we report some considerations about the e ects of their integration.

6.1.1 Crossover Operators Used by ARGAF

Di erent proposals of crossover operators for RCGAs are available in the lit-
erature (see [28]). We have built di erent versions of ARGAF with the following
types:
{ fuzzy connectives-based crossovers (FCB-crossovers) ([27, 29]),
{ dynamic FCB-crossovers ([31, 30]),
{ heuristic FCB-crossovers ([30]) and
{ dynamic heuristic FCB-crossovers ([30]).
Next, we comment on some of the main features of these crossover operators
and describe how ARGAF uses them.

FCB-crossovers These crossovers include explorative and exploitative crossovers.

The rst ones are the F -crossover and S -crossover operators, whereas the second
ones are the M -crossover operators. Di erent familes of F -, S - and M -crossover
operators were presented in [27, 29]: the Logical, Hamacher, Algebraic and Ein-
stein ones. Logical F - and S -crossover operators show the lowest degree of explo-
ration. The crossover strategy for ARGAF using FCB-crossovers is the following:
? Strategy ST1 :
For each pair of chromosomes from a total of c Do p
N

If a random number [0 1] is lower than e Then

r 2 ; p

Generate two o spring, the result of applying two -crossovers.

M

Else
Generate two o spring, the result of applying an -crossover and an
F

-crossover.
S

The two o spring substitute their parents in the population.

Dynamic FCB-crossovers These operators keep a suitable sequence between

the exploration and the exploitation throughout the GA run: "to protect the
exploration in the initial stages and the exploitation later". This sequence is
suitable for avoiding the premature convergence problem and for improving GA
eciency [42]. The -crossovers and the -crossovers vary the exploration degree
F S

throughout the GA run: rstly, high diversity is induced, later on, their behaviour
is similar to the one in the Logical - and -crossovers and so convergence
F S

is caused. The -crossovers vary the exploitation degree throughout the GA

M

run. Di erent families of dynamic FCB-crossover were presented in [31, 30]: the
Dubois, Dombi and Frank ones. Dubois operators supply the lowest levels of
diversity. The crossover strategy for ARGAF using dynamic FCB-crossovers,
ST 2 , is similar to 1 , but now -, - and
ST F S -crossover operators are applied.
M

Heuristic FCB-crossovers Heuristic FCB-crossover operators use the tness

parents for leading the search process towards the most promising zones. Two
types were presented in [30]: one with heuristic exploration properties, called
dominated heuristic FCB-crossovers, that attempts to introduce a useful di-
versity into the GA population; another with heuristic exploitation properties,
called biased heuristic FCB-crossovers, that induces a biased convergence led
toward the best elements. The crossover strategy for ARGAF using heuristic
FCB-crossovers, 3 , is the following:
ST

? Strategy 3 :
ST

Repeat p
c times

N

Choose a pair of parents

If a random number [0 1] is lower than e Then
r 2 ; p

Generate an o spring, the result of applying a biased heuristic FCB-

crossover.
Else
Generate a o spring, the result of applying an dominated heuristic
FCB-crossover.
The o spring generated substitute one of its parents. It shall never be con-
sidered as a parent.
Dynamic Heuristic FCB-crossovers Dynamic heuristic FCB-crossovers put
together the heuristic properties and the features of the dynamic FCB-crossover
operators. Again, two types were presented in [30]: the dynamic dominated
heuristic FCB-crossovers and the dynamic biased heuristic FCB-crossovers. The
crossover strategy for ARGAF using dynamic heuristic FCB-crossovers, ST4 , is
analogous to ST3 .
Next, we justify the reasons for the adaptation of the parameter pe.

6.1.2 Adaptation of the pe parameter

The role of the pe parameter is to establish the frequency of application
of each crossover type (the exploitative and explorative ones). It is a control
parameter for the selection of crossover operators.
The crossover operator plays an important role in the RCGA's performance
([19, 58, 29, 30]). The value of pe strongly in uences the EER induced by
crossover; if pe is low, ARGAF shall generate diversity and so exploration takes
e ect, whereas if it is high, the current diversity shall be used for generating
best elements and so exploitation comes into force. With the adaptation of pe ,
ARGAF may handle the EER produced by crossover in a suitable way.
Many adaptive GAs include the crossover operator as a key element in the
adaptation mechanism ([48, 14, 54, 52, 62, 30, 31]). The idea of an adaptive
selection between crossover operators with di erent explorative or exploitation
properties appears in [54] and [14]. In the rst paper, a tightly coupled adaptive
BCGA was de ned that considers only two forms of crossover: the two-point
and uniform ones. It was claimed that the use of these operators is reasonable
since they represent two extremes regarding their associated disruption level: the
lower one is for two-point crossover whilst the higher one is for uniform crossover
([16]). Thus, it is natural to allow the GA to explore a relative mixture of these
two operators.
The model in the second reference was described in Subsection 2.2.4; an
uncoupled adaptive RCGA adjusted parameters similar to pe . These parame-
ters de ned the frequency of crossover and mutation operators with di erent
exploitation and exploration properties.
We should point out that although the adaptation of pe may be considered
as an instance of adaptive parameter settings, it may also be considered to be
an instance of adaptive genetic operator selection.
Now, we explain the reasons for tuning the selective pressure (min ) along
with pe .

6.1.3 Adaptation of the min parameter

The nal results for GAs depend on the interactions between their com-
ponents (genetic operators, parameter settings, etc). If we want to control a
particular GA element, these interactions should be taken into account, and so
other elements need to be controlled as well. Since pe is controlled, we consider
that selection should be controlled as well; one reason for doing so is explained
in [17]:
Although the amount of exploration provided by crossover will be a ected
by the crossover rate, it will also be a ected by selection. As selection
pressure is increased, variability is decreased in the population, and con-
sequently there are fewer di erences that crossover can explore by recom-
bination. To the extent that the population has converged, there is less
information that can be shared between individuals. In short, crossover is
severely a ected by the exploration-exploitation tradeo . The amount of
exploration performed by crossover is limited by the amount of exploita-
tion performed by selection. Increased exploitation by selection leads to
decreased exploration by crossover.
These words highlight the important relations between the crossover operator
and the selection mechanism. These relations have been considered for improv-
ing the GA search by some authors; in [18], for example, it is advised that a
disruptive crossover operator accompanied by a conservative selection strategy
provides an e ective search.
So, a feedback between crossover and selection must be established, allowing
a suitable balance between their actions to be reached through the GA run.
For doing so, in ARGAF, the selective pressure of the linear ranking selection
mechanism is adjusted along with pe . In [6], the following properties of selection
mechanisms are pointed out which seem particularly desirable in order to easily
control the characteristics of the search process by means of varying the control
parameters of selection mechanisms:
1. The impact of the control parameters on selective pressure should be simple
and predictable.
2. One single control parameter for selective pressure is preferable.
3. The range of selective pressure that can be made by varying the control
parameter should be as large as possible.
Linear ranking mechanism matches these requeriments well. In this selection
mechanism the selective pressure is determined by the parameter min 2 [0; 1].
If min is low, high pressure is achieved, whereas if it is high, the pressure is low.
Finally, we point out that Baker suggested, in [7], adapting selection as well;
he proposed switching between two selection mechanisms with di erent proper-
ties when a rapid convergence is predicted.

6.1.4 An Integrated Frame

Using the pe parameter, ARGAF controls the e ects of crossovers, i.e., either
generating diversity or using diversity, whereas using the min parameter, it con-
trols the e ects of selection, i.e., either keeping diversity or eliminating diversity.
The joint management of these parameters allows ARGAF to administer the
diversity in a suitable way. For example, if useful diversity is detected by AR-
GAF, then it sets selection for keeping diversity and crossover for using it. If the
level of diversity is high and its quality is not good, then ARGAF increases the
selective pression and tries to obtain better elements by increasing exploitation
by means of crossover. In the rule base presented below, all these considerations
are highlighted.

6.2 Design of the FLCs Used by ARGAF

In this subsection, we discuss the di erent steps in the design of the FLCs used
by ARGAF. First, we consider the inputs and outputs by establishing their
associated ranges and linguistic labels, then we de ne the data base and the rule
base.

6.2.1 Inputs
Two diversity measures are considered as inputs. One is the genotypic diver-
sity measure ED described in Subsection 4.1.4 and the other is the phenotypic
diversity measure PDM1 presented in Subsection 4.2. The rst measure repre-
sents the quantity of diversity of the genetic material in the population and the
second the quality of this diversity. Most adaptive approaches based on diversity
measures use only one of these measures, i.e. either a genotypic diversity mea-
sure or a phenotypic diversity measure (see Subsection 2.2). We think that it is
more adequete to control the EER of the RCGAs using two diversity measures;
one of each type. The main goal of the FLCs used by ARGAF is to maintain
genotypic diversity, but with good properties, i.e, with an apropriate phenotypic
diversity. In [21] it is stated that maintaining diversity for its own sake is not the
issue; by itself, it doesn't guarantee the improvement of GA behaviour. Instead,
we need to maintain appropriate diversity, i.e., diversity that in some way helps
cause good chromosomes, i.e., useful diversity ([40]). Using a genotypic diversity
measure and a phenotypic diversity measure we may measure the quantity and
quality of the diversity. In this way, useful diversity may be rst detected and
later handled.
The range of ED and PDM1 is [0; 1]. The linguistic label set of these inputs
is fLow; Medium; Highg. Table 2. shows the correspondences between the ED
measure and the level of genotypic diversity and the PDM1 measure and the
level of phenotypic diversity.

6.2.2 Outputs
The outputs are variables that control the variation on the current pe and
min parameters, which are kept within the range [0:25; 0:75]. These variables,
noted as pe and min , represent the degree to which the current pe and min
 Correspondences between the diversity measures and the diversity levels
Table 2.

ED Genotypic Diversity PDM1 Phenotypic Diversity

Low Low Low High
Medium Medium Medium Medium
High High High Low

should vary, respectively. The variations shall be carried out by multiplying the
pe and min values, obtained by the FLC, by the current pe and min values,
respectively. The action interval of pe and min is [0:5; 1:5] and its associated
linguistic labels shall be fSmall; Medium; Bigg as well.

6.2.3 Data Base

The data base is shown in Figure 3. The meaning of the linguistic terms
associated with inputs ED is depicted in (a), the ones for PDM1 in (b) and
nally, the ones for pe and min in (c). For each linguistic term, there is a
triangular fuzzy set that de nes its semantic, i.e, its meaning.

Low Medium High Low Medium High

0.0 0.2 0.5 0.8 1.0 0.0 0.2 1.0

(a) ED (b) PDM 1

Small Medium Big

0.0 1.0 1.5

(c) δ p e and δ ηmin

Fig. 3. Meaning of the linguistic terms associated with the inputs and the outputs
6.2.4 Rule Bases
The rules describe the relation between the inputs and outputs. Table 3.
shows the rule bases used by the FLCs of ARGAF.

Table 3. Rule Bases for the Control of pe and min , respectively

P DM1 P DM1
ED Low Medium High ED Small Medium Large
Low Small Small Medium Low Small Medium Big
Medium Big Big Medium Medium Small Big Big
High Big Big Medium High Small Small Big

6.3 Experiments
Minimization experiments on four functions shown in Table 4., were carried out
in order to study the eciency of ARGAF. The number of variables, n, is 25. f1
is a continuous, strictly convex and unimodal function. f2 is a continuous and
unimodal function, with the optimum located in a steep parabolic valley with a
at bottom. This feature will probably cause slow progress in many algorithms,
since they must permanently change their search direction to reach the optimum.
f3 is a scalable, continuous and multimodal function which is produced from f1
by modulating it with a cos(! x ). f4 is a continuous and multimodal function.


i
This function is very dicult to optimize because it is non-separable.

Table 4. Test functions

 Sphere model ([15, 47])  Generalized Rosenbrock's function ([15])
P Pn?1
f1 (x) = n i=1 xi
2 f2 (x) = 22 2
i=1 (100
(xi+1 ? xi ) + (xi ? 1) )
?5:12  xi  5:12 ?5:12  xi  5:12
f1 = f1 (0; : : : ; 0) = 0 f2 = f2 (1; : : : ; 1) = 0

 Generalized Rastringin's  Griewangk's function ([26])

function ([57, 5])
Pn Pn Qn  
f3 (x) = a
n + i=1 x2i ? a
cos(!
xi ) f4(x) = d1 2 xi
i=1 xi ? i=1 cos pi + 1
a = 10, ! = 2 d = 4000
? 5:12  xi  5:12 ?600:0  xi  600:0
f3 = f3 (0; : : : ; 0) = 0 f4 = f4 (0; : : : ; 0) = 0
6.3.1 Algorithms
Table 5. shows the di erent ARGAF variations built for the experiments.

ARGAF variations
Table 5.

ARGAFs Crossover Strategy Operator Families

ARGAF1Lo ST1 Logical
ARGAF1Ha ST1 Hamacher
ARGAF2Du ST2 Dubois
ARGAF2Do ST2 Dombi
ARGAF3Lo ST3 Logical
ARGAF3Ha ST3 Hamacher
ARGAF4Du ST4 Dubois
ARGAF4Do ST4 Dombi

For studying the adaptive behaviour of ARGAF, we have executed some al-
gorithms like ARGAF, but with xed pe and min values. Di erent combinations
of these parameters were considered. For denoting these algorithms, we add to
each ARGAF name a sux such as those shown in Table 6.

Table 6. Sux used for denoting the algorithms with xed pe and min
Sux pe min
a 0.25 0.75
b 0.5 0.75
c 0.25 0.25
d 0.75 0.25

The mutation operator used by all the algorithms is the non-uniform muta-
tion ([42]). We carried out the experiments using the following parameters: the
population size is 61 individuals, the crossover probability pc = 0:6, the muta-
tion probability pm = 0:005, the sampling model used was stochastic universal
sampling ([8]). The FLCs of ARGAF are red every 5 generations. The initial
values for pe and min are 0:25 and 0:75, respectively. We executed all the al-
gorithms 5 times, each one with 5000 generations, and presented their average
value.

6.3.2 Results
Table 7. shows the average values of the results obtained. For each function
the Best from 5000 generations is shown and the nal Online measure (see Sec-
tion 4.2), which is used for measuring the phenotypic diversity measure produced
throughout the GA run.
6.4 Discussions of the Results

Looking over the results, we may report the following considerations about the
behaviour of the ARGAFs compared with their corresponding xed con gura-
tions.
In general, the results of ARGAF are similar to the ones of the most successful
xed con guration. So, this shows that ARGAF is a very robust GA since it
adapts the pe and min parameters to the setting that returns the best results.
ARGAFs based on crossover operators of the Dombi and Hamacher types
improve the results obtained by the remaining algorithms (see ARGAF2Do for
example). These operators supply good levels of diversity and so ARGAF may
administer it. This does not occur with the operators of the Logical and Dubois
types; since the diversity produced by them is very low, the reactions of ARGAF
don't produce any positive e ects. For example, let's suppose that the population
shows a low level of diversity, then ARGAF will try to produce more diversity
through crossover by means of the decrease in pe . However, this reaction does not
have a preeminent e ect, due to probability the diversity produced by crossover
shall not be sucient. On the other hand, with crossover operators with high
associated diversity levels, ARGAF produces sucient diversity for subsequent
treatment. So, we conclude that the ARGAF operation is better using crossover
operators with high associated diversity levels.
The Online measure associated with the ARGAFs show average values. This
is reasonable, since ARGAF changes settings frequentely.

7 Open Problems and Extensions

Although ARGAF has shown a good behaviour and other applications of FLCs
to the control of GAs have achieved suitable results, we think that there are still
many open problems. Next, we discuss some of them.
1. The behaviour of GAs and the interelations between the genetic operators
are very complex, although there are many possible inputs and outputs for
the FLCs, frequently rule bases are not easily available: " nding good rule
bases is not an easy task".
2. Automatic learning mechanisms for obtaining the rule bases may be used for
avoiding the previous problem. As we reported in Subsection 3.2, in [37], this
kind of automatic learning technique was proposed for obtaining suitable rule
bases (and the data bases as well). This technique is very similar to the meta-
GA of Grefenstette, ([24]). This implies that the robustness of the returned
rule bases depends strongly on the test function set and the performance
measures used by the technique. Therefore, the rule base de nition is still
a challenging feature in the fuzzy control of the GAs; Tettamanzi, in [56],
quoted an interesting re exion on this theme:
Table 7. Results
f1 f2 f3 f4
Algorithms 5000 Online 5000 Online 5000 Online 5000 Online
Lo
ARGAF1 7.98e-16 8.77e-01 2.10e+01 7.54e+02 3.98e-01 1.86e+01 2.84e-02 3.74e+00
ARGAF1Lo a 3.20e-11 1.13e+00 2.20e+01 1.07e+03 3.98e-01 2.00e+01 5.91e-03 4.44e+00
ARGAF1Lo b 1.90e-11 8.91e-01 2.21e+01 8.18e+02 1.59e+00 2.22e+01 1.33e-02 3.67e+00
ARGAF1Lo c 6.46e-17 6.45e-01 2.84e+01 6.71e+02 3.98e-14 5.57e+00 3.04e-02 2.56e+00
ARGAF1Lo d 8.55e-16 5.15e-01 2.10e+01 5.45e+02 6.03e-13 8.92e+00 2.60e-02 2.12e+00
ARGAF1Ha 1.75e-09 2.34e+01 8.29e+00 8.38e+03 2.23e+01 3.08e+02 4.48e-02 8.56e+01
ARGAF1Haa 2.36e-02 1.15e+02 1.01e+01 1.01e+05 3.32e+01 3.87e+02 1.12e+00 3.97e+02
ARGAF1Hab 2.79e-02 7.56e+01 1.53e+01 5.05e+04 3.72e+01 3.51e+02 1.09e+00 2.61e+02
ARGAF1Hac 8.78e-03 7.33e+01 6.73e+00 3.55e+04 2.24e+01 3.46e+02 9.50e-01 2.53e+02
ARGAF1Had 1.12e-09 2.16e+01 1.24e+01 7.81e+03 7.76e+00 1.87e+02 4.63e-02 7.51e+01
ARGAF2Du 7.70e-16 1.30e+00 3.68e+01 1.23e+03 7.96e-01 2.04e+01 3.29e-02 5.54e+00
ARGAF2Dua 2.57e-11 1.56e+00 2.07e+01 1.59e+03 5.97e-01 2.13e+01 1.97e-02 5.93e+00
ARGAF2Dub 2.88e-11 1.22e+00 2.19e+01 1.21e+03 3.38e+00 2.86e+01 1.53e-02 4.85e+00
ARGAF2Duc 6.37e-16 8.44e-01 2.43e+01 9.01e+02 3.41e-14 7.06e+00 5.05e-02 3.27e+00
ARGAF2Dud 4.03e-15 6.77e-01 7.06e+00 6.70e+02 2.57e-12 8.95e+00 2.85e-02 2.71e+00
ARGAF2Do 2.46e-09 2.80e+01 9.04e-04 2.93e+04 1.59e+00 1.94e+02 1.78e-02 1.02e+02
ARGAF2Doa 1.01e-04 1.06e+02 2.66e-02 1.41e+05 4.53e+00 3.41e+02 5.35e-01 3.65e+02
ARGAF2Dob 1.15e-04 7.36e+01 1.61e-02 8.85e+04 1.51e+00 3.01e+02 1.77e-01 2.54e+02
ARGAF2Doc 7.31e-04 8.19e+01 1.36e-01 9.20e+04 6.58e+00 3.18e+02 5.69e-01 2.82e+02
ARGAF2Dod 6.60e-10 2.68e+01 1.29e-03 2.74e+04 4.18e+00 1.58e+02 4.72e-02 9.29e+01
ARGAF3Lo 2.06e-18 3.87e-01 3.23e+01 4.24e+02 2.27e-14 4.85e+00 3.10e-02 1.56e+00
ARGAF3Lo a 6.79e-16 3.88e-01 3.21e+01 4.28e+02 1.19e-13 4.69e+00 2.21e-02 1.55e+00
ARGAF3Lo b 2.11e-15 3.30e-01 2.13e+01 3.71e+02 3.69e-13 4.86e+00 7.89e-03 1.31e+00
ARGAF3Lo c 1.04e-17 3.75e-01 1.98e+01 3.91e+02 2.27e-14 4.69e+00 4.51e-02 1.57e+00
ARGAF3Lo d 1.40e-15 2.92e-01 2.15e+01 3.36e+02 5.97e-01 6.60e+00 1.33e-02 1.17e+00
ARGAF3Ha 2.00e-14 1.94e+01 2.05e+01 7.37e+03 1.95e+02 4.02e+02 1.08e-02 6.78e+01
ARGAF3Haa 1.26e-02 7.59e+01 2.32e+01 4.93e+04 2.27e+02 4.21e+02 8.62e-01 2.63e+02
ARGAF3Hab 1.85e-11 3.51e+01 5.59e+00 1.25e+04 1.76e+02 3.87e+02 2.60e-01 1.22e+02
ARGAF3Hac 5.43e-05 5.62e+01 8.97e+00 2.48e+04 1.95e+02 3.98e+02 5.19e-01 1.94e+02
ARGAF3Had 1.29e-14 1.72e+01 1.70e+01 7.21e+03 2.07e+01 1.92e+02 1.28e-02 5.95e+01
ARGAF4Du 1.37e-17 4.83e-01 3.22e+01 5.14e+02 3.41e-14 5.54e+00 3.78e-02 1.98e+00
ARGAF4Dua 3.47e-19 4.90e-01 2.07e+01 5.16e+02 5.68e-15 5.06e+00 2.55e-02 1.99e+00
ARGAF4Dub 2.53e-22 4.06e-01 2.14e+01 4.30e+02 0.00e+00 5.93e+00 2.11e-02 1.68e+00
ARGAF4Duc 2.94e-18 4.71e-01 3.14e+01 4.94e+02 1.14e-14 4.81e+00 2.41e-02 1.93e+00
ARGAF4Dud 5.43e-21 3.26e-01 2.15e+01 3.55e+02 0.00e+00 5.15e+00 2.75e-02 1.38e+00
ARGAF4Do 1.65e-13 2.44e+01 1.33e+01 2.51e+04 1.48e+01 2.62e+02 8.97e-03 8.48e+01
ARGAF4Doa 3.18e-05 7.54e+01 2.28e+01 9.32e+04 1.02e+01 3.14e+02 3.03e-01 2.60e+02
ARGAF4Dob 1.44e-06 4.80e+01 2.35e+01 5.07e+04 1.18e+01 2.78e+02 1.71e-01 1.66e+02
ARGAF4Doc 9.53e-06 6.80e+01 4.92e+00 7.90e+04 1.30e+01 3.08e+02 1.93e-01 2.34e+02
ARGAF4Dod 2.21e-13 2.36e+01 9.02e+00 2.34e+04 8.76e+00 1.91e+02 2.67e-02 8.19e+01
 Statistics and parameters are in part universal to any evolutionary
algorithm and in part speci c to a particular application. Therefore
it is hard to state general rules to control the evolutionary process.
These are to be discovered by the implementer by means of experi-
ments and empirical observations.
3. Another important issue to consider is the frequency of application of the
fuzzy control. Usually, a xed scheduling for ring the FLC has been followed,
i.e. after each xed number of generations. However, the choice of the time-
interval between controls is a parameter that has an in uence on the nal
controller performance. If the controller is red with a low time-interval the
e ects of previous controls may not be achieved, whereas if the controller is
red with a high time-interval,the search process may be misled by particular
parameter values. A possible solution is to re the controller when certain
conditions relating to some performance measures are reached.
Next, we set out some possible extensions that may improve the behaviour
of adaptive GAs based on FLCs.
1. Design FLCs taking into account the action of each genetic operator in re-
lation to the behaviour of each one of the remaining ones. The future action
of an operator may be tuned depending on the repercussions of the actions
of other operators (even itself). In this way, a feedback between operators
must be established, allowing a suitable balance between their actions to be
reached throughout the GA run.
2. Try a distributed application of FLCs. In Subsection 2.1, it was stated that
uncoupled adaptive GAs show certain problems due to their central coordi-
nation or control.
In [55], a central control of the pc and pm parameter was sustituted by a
model where pc and pm are computed for each chromosome depending on the
convergence state of the population and the performance of such chromosome
(see Subsection 2.2.4). Simple ideas were used for implementing this model,
such as high- tness chromosomes should be protected, i.e., the pm for them
should be low, whereas subaverage tnesses should be disrupted, i.e. the pm
for them should be high. These ideas may be easily implemented by means
of fuzzy rules.
For RCGAs, for example, it would be interesting to control which changes
produced by mutation operators on real-coded chromosomes are adequate:
genes belonging to 'not good' chromosomes should receive large changes,
whereas these changes should be small for genes belonging to good chromo-
somes. This issue may be easily tackled by an FLC.
These examples show that a distributed application of FLC to the control
of GAs is possible. Inputs to the FLC may be central measures along with
measures associated with particular chromosomes or sets of chromosomes.
The outputs would be control parameters associated with genetic operators
that are applied on those chromosomes.
3. Build adaptive representation and adaptive tness function models based on
FLCs. In this paper, we have seen that FLCs are used for adaptive parameter
 setting and adaptive genetic operator selection (ARGAF is an example of
this). Extensions of the application of FLCs may be made for designing
adaptive representation and adaptive tness function techniques.

8 Conclusions
In this paper we have reviewed many aspects of the adaptation of GA parame-
ters based on FLCs; we have stressed the di erent steps for the design of such
controllers and we have highlighted the diversity measures as suitable candidate
inputs for them. Moreover, we presented ARGAF, an adaptive RCGA based on
the use of FLC. An important feature of ARGAF is that it applies two types
of crossover operators, one with exploration properties and the other with ex-
ploitation properties. The results of the experiment carried out with ARGAF
showed that the ARGAF operation is better using crossover operators with high
associated diversity levels .
An important conclusion arising from our study is that: "designing good FLCs
for controlling GAs is not an easy task". This is due to the great complexity of
GA behaviour and the interelations between the genetic operators. Therefore, we
have explained that many open problems still remain and that some extensions
may be taken into account for improving the technique studied.

References
[1] J. Arabas, Z. Michalewicz, J. Mulawka, GAVaPS - a Genetic Algorithm with
varying population size, in: Proc. of The First IEEE Conference on Evolutionary
Computation (1994) 73-78.
[2] S. Arnone, M. Dell'Orto, A. Tettamanzi, Toward a fuzzy government of genetic
populations, in: Proc. of the 6th IEEE Conference on Tools with Arti cial Intel-
ligence (IEEE Computer Socity Press, Los Alamitos, CA, 1994).
[3] T. Back, The interaction of mutation rate, selection, and self-adaptation within
genetic algorithm, in: R. Manner, B. Manderick, Eds., Parallel Problem Solving
from Nature 2 (Elsevier Science Publishers, Amsterdam, 1992) 85-94.
[4] T. Back, F. Ho meister, Genetic algorithms and evolution strategies: similari-
ties and di erences, Technical Report No. SYS-1/92. Dept. of Computer Science,
University of Dortmund, Germany (1992).
[5] T. Back, Self-adaptation in genetic algorithms, in: F.J. Varela, P. Bourgine, Eds.,
Proc. of the First European Conference on Arti cial Life (The MIT Press, Cam-
bridge, MA, 1992) 263-271.
[6] T. Back. Selective pressure in evolutionary algorithms: a characterization of selec-
tion mechanisms, in: Proc. of The First IEEE Conference on Evolutionary Com-
putation (1994) 57-62.
[7] J.E. Baker, Adaptive selection methods for genetic algorithms, in: Proc. First
Int. Conf. on Genetic Algorithms (L. Erlbaum Associates, Hillsdale, MA, 1985)
101-111.
[8] J.E. Baker, An analysis of the e ects of selection in genetic algorithms, Doctoral
Dissertation, University of Vanderbilt (1989).
[9] M.A. Bedau, F. Ronneburg, M. Zwick, Dynamic of diversity in an evolving popu-
lation, in: R. Manner, B. Manderick, Eds., Parallel Problem Solving from Nature
2 (Elsevier Science Publichers, Amsterdam, 1992) 95-104.
[10] A. Bergmann, W. Burgard, A. Hemker, Adjusting paremeters of genetic algorithms
by fuzzy control rules, in K.-H. Becks, D. Perret-Gallix, Eds., New Computing
Techniques in Physics Research III (World Scienti c Press, Singapore, 1994).
[11] L. Booker, Improving search in genetic algorithms, in: L. Davis, Ed., Genetic
Algorithms and Simulated Annealing (Morgan Kaufmann Publishers, Los Altos,
1987) 61-73.
[12] M.F. Bramlette, Initialization, mutation and selection methods in genetic algo-
rithms for function optimization, in: R. Belew, L.B. Booker, Eds., Proc. of the
Fourth Int. Conf. on Genetic Algorithms (Morgan Kaufmmann, San Mateo, 1991)
100-107.
[13] R.J. Collins, D.R. Je erson, Selection in massively parallel genetic algorithms, in:
R. Belew, L.B. Booker, Eds., Proc. of the Fourth Int. Conf. on Genetic Algorithms
(Morgan Kaufmmann, San Mateo, 1991) 7249-256.
[14] L. Davis, Adapting operator probabilities in genetic algorithms, in: J. David Schaf-
fer, Ed., Proc. of the Third Int. Conf. on Genetic Algorithms (Morgan Kaufmann
Publishers, San Mateo, 1989) 61-69.
[15] K.A. De Jong, An analysis of the behavior of a class of genetic adaptive systems,
Doctoral Dissertation, University of Michigan (1975).
[16] K.A. De Jong, W.M. Spears, A formal analysis of the role of multi-point crossover
in genetic algorithms, Annals of Mathematics and Arti cial Intelligence Journal
5(1) (1992) 1-26.
[17] L.J. Eshelman, R.A. Caruana, J.D. Scha er, Biases in the crossover landscape,
in: J. David Scha er, Ed., Proc. of the Third Int. Conf. on Genetic Algorithms
(Morgan Kaufmann Publishers, San Mateo, 1989) 10-19.
[18] L.J. Eshelman, The CHC adaptive search algorithm: how to have safe search
when engaging in nontraditional genetic recombination, in: G.J.E Rawlin, Ed.,
Foundations of Genetic Algorithms 1 (Morgan Kaufmann, San Mateo) 265-283.
[19] L.J. Eshelman, J.D. Scha er J.D., Real-coded genetic algorithms and interval-
schemata, in: L.Darrell Whitley, Ed., Foundation of Genetic Algorithms-2 (Morgan
Kaufmann Publishers, San Mateo, 1993) 187-202.
[20] T.C. Fogarty, Varying the probability of mutation in the genetic algorithm, in: J.
David Scha er, Ed., Proc. of the Third Int. Conf. on Genetic Algorithms (Morgan
Kaufmann Publishers, San Mateo, 1989) 104-109.
[21] D.E. Goldberg, J. Richardson, Genetic algorithms with sharing for multimodal
function optimization, in: Genetic Algorithms and their Applications: Proc. of the
Second Int. Conf. on Genetic Algorithms (L. Erlbaum Associates, Hillsdale, MA,
1987) 41-49.
[22] D.E. Goldberg, Genetic Algorithms in Search, Optimization, and Machine Learn-
ing, (Addison-Wesley, New York, 1989).
[23] D.E. Goldberg, Zen and the art of genetic algorithms, in: J. David Scha er, Ed.,
Proc. of the Third Int. Conf. on Genetic Algorithms (Morgan Kaufmann Publish-
ers, San Mateo, 1989) 61-69.
[24] J.J. Grefenstette, Optimization of control parameters for genetic algorithms, IEEE
Trans. on Systems, Man, and Cybernetics 16 (1986) 122-128.
[25] J.J. Grefenstette, Incorporating problem speci c knowledge into genetic algo-
rithms, in: L. Davis, Ed., Genetic Algorithms and Simulated Annealing (Morgan
Kaufmann Publishers, Los Altos, 1987) 42-60.
[26] A.O. Griewangk, Generalized descent of global optimization, JOTA 34 (1981) 11-
39.
[27] F. Herrera, E. Herrera-Viedma, M. Lozano, J.L. Verdegay, Fuzzy tools to improve
genetic algorithms, in: Proc. of the Second European Congress on Intelligent Tech-
niques and Soft Computing (1994) 1532-1539.
[28] F. Herrera, M. Lozano, J.L. Verdegay, Tackling real-coded genetic algorithms:
operators and tools for the behavioural analysis, Arti cial Intelligence Reviews
(1996). To appear.
[29] F. Herrera, M. Lozano, J.L. Verdegay, Fuzzy connectives based crossover operators
to model genetic algorithms population diversity, Technical Report #DECSAI-
95110, Dept. of Computer Science and A. I., University of Granada (Spain) (1995).
[30] F. Herrera, M. Lozano, J.L. Verdegay, Dynamic and heuristic fuzzy connectives
based crossover operators for controlling the diversity and convergence of real-
coded genetic algorithms. International Journal of Intelligent Systems (1996). To
appear.
[31] F. Herrera, M. Lozano, J.L. Verdegay, The use of fuzzy connectives to design
real-coded genetic algorithms, Mathware & Soft Computing 1(3) (1995) 239-251.
[32] J.H. Holland, Adaptation in Natural and Arti cial Systems, (The University of
Michigan Press, 1975). (The MIT Press, London, 1992).
[33] Y. Ichikawa, Y. Ishiiolland, Retainig diversity of genetic algorithms for multivari-
able optimization and neural network learning, in: Proc. IEEE Int. Conf. on Neural
Networks (San Francisco, California, 1993) 1110-1114.
[34] B. A. Julstrom. What have you done for me lately? adapting operator probabilities
in a steady-state genetic algorithm, in: L. Eshelman, Ed., Proc. of the Sixth Int.
Conf. on Genetic Algorithms (Morgan Kaufmann Publishers, San Francisco, 1995)
81-87.
[35] Y. Kakazu, H. Sakanashi, K. Suzuki, Adaptive search strategy for genetic algo-
rithms with additional genetic algorithms, in: R. Manner, B. Manderick, Ed.,
Parallel Problem Solving from Nature 2 (Elsevier Science Publishers, Amsterdam,
1992) 311-320.
[36] C.C. Lee, Fuzzy logic in control systems: fuzzy logic controller - Parts I and II,
IEEE Transactions on Systems, Man and Cybernetics 20(2) (1990) 404-435.
[37] M.A. Lee, H. Takagi, Dynamic control of genetic algorithms using fuzzy logic
techniques, in: S. Forrest, Ed., Proc. of the Fifth Int. Conf. on Genetic Algorithms
(Morgan Kaufmmann, San Mateo, 1993) 76-83.
[38] T-H. Li, C.B. Lucasius, G. Kateman, Optimization of calibration data with the
dynamic genetic algorithm, Analytica Chimica Acta 2768 (1992) 123-134.
[39] S.J. Louis, G.J.E Rawlins. Syntactic analysis of convergence in genetic algorithms,
in: L.Darrell Whitley, Ed., Foundations of Genetic Algorithms-2 (Morgan Kauf-
mann Publishers, San Mateo, 1993) 141-151.
[40] S.W. Mahfoud, Niching methods for genetic algorithms, IlliGAL Report No. 95001,
Urbana: University of Illinois, Illinois Genetic Algorithms Laboratory (1995).
[41] M.L. Mauldin, Maintaining diversity in genetic search, in: Proc. of the Nat. Conf.
on Art. Int. (Austin, TX, 1984) 247-250.
[42] Z. Michalewicz, Genetic Algorithms + Data Structures = Evolution Programs,
(Springer-Verlag, New York, 1992).
[43] J.C. Potts, T.D. Giddens, S.B. Yadav, The development and evaluation of an im-
proved genetic algorithm based on migration and arti cial selection. IEEE Trans.
on Systems, Man, and Cybernetics 24 (1994) 73-86.
[44] X. Qi. Analysis and applications of darwinian optimization algorithms in multi-
dimensional spaces. Doctoral Dissertation, University of Connecticut, 1992.
[45] H. Sakanashi, K. Suzuki, Y. Kakazu, Controlling dynamics of GA through ltered
evaluation function, in: Y. Davidor, H.-P. Schwefel, R. Maenner, Eds., Parallel
Problem Solving from Nature{PPSN III, International Conf. on Evolutionary
Computation, Proceedings, volume 866 of Lecture Notes in Computer Science
(Berlin, Germany, Springer-Verlag, 1994) 239-248.
[46] C.G. Shaefer, The ARGOT strategy: adaptive representation genetic optimizer
technique, in: Proc. of the Second International Conference (Hillsdale, NJ:
Lawrence Erlbaum Associates, 1987).
[47] H-P. Schwefel, Numerical Optimization of Computer Models, (Wiley, Chichester,
1981).
[48] J.D. Scha er, A. Morishima. An adaptive crossover distribution mechanism for
genetic algorithms, in: Proc. Second Int. Conf. on Genetic Algorithms (L. Erlbaum
Associates, Hillsdale, MA, 1987) 36-40.
[49] J.D. Scha er, R.A Caruana, L.J. Eshelman, R. Das, A study of control parameters
a ecting online performance of genetic algorithms for function optimization, in: J.
David Scha er, Ed., Proc. of the Third Int. Conf. on Genetic Algorithms (Morgan
Kaufmann Publishers, San Mateo, 1989) 51-60.
[50] N.N. Schraudolph, R.K Belew, Dynamic parameter encoding for genetic algo-
rithms. Machine Learning 9 (1992) 9-21.
[51] D. Schlierkamp-Voosen, H. Muhlenbein, Strategy adaptation by competing sub-
populations, in: Y. Davidor, H.-P. Schwefel, R. Maenner, Eds., Parallel Problem
Solving from Nature{PPSN III, International Conf. on Evolutionary Computation,
Proceedings, volume 866 of Lecture Notes in Computer Science (Berlin, Germany,
Springer-Verlag, 1994) 199-208.
[52] M. Sebag, M. Schownauer, Controlling crossover through inductive learning, in: Y.
Davidor, H.-P. Schwefel, R. Maenner, Eds., Parallel Problem Solving from Nature{
PPSN III, International Conf. on Evolutionary Computation, Proceedings, volume
866 of Lecture Notes in Computer Science (Berlin, Germany, Springer-Verlag,
1994) 209-218.
[53] R.E. Smith, Adaptively resizing populations: an algorithm and analysis, in: S.
Forrest, Ed., Proc. of the Fifth Int. Conf. on Genetic Algorithms (Morgan Kaufm-
mann, San Mateo, 1993) 653-653.
[54] W.M. Spears, Adapting crossover in a genetic algorithm, Laboratory Report,
#AIC-92-025, Navy Center for Applied Research in Arti cial Intelligence, USA
(1992).
[55] M. Srinivas, L.M. Patnaik, Adaptive probabilities of crossover and mutation in
genetic algorithms. IEEE Trans. on Systems, Man, and Cybernetics 24(4) (1994)
656-667.
[56] A.G. Tettamanzi, Evolutionary algorithms and fuzzy logic: a two-way integration,
in: 2nd Joint Conference on Information Sciences (Wrightsville Beach, NC, 1995).
[57] A. Torn, Z Antanas, Global Optimization, (Lecture Notes in Computer Science,
Vol 350, Springer, Berlin, 1989).
[58] Voigt, H.M., Muhlenbein, H., Cvetkovic, Fuzzy recombination for the breeder
genetic algorithm, in: L. Eshelman, Ed., Proc. of the Sixth Int. Conf. on Genetic
Algorithms (Morgan Kaufmann Publishers, San Francisco, 1995) 104-111.
[59] D. Whitley, T. Hanson, Optimizing neural networks using faster, more accurate
genetic search. Proc. of the Third Int. Conf. on Genetic Algorithms, J. David
Scha er (Ed.), (Morgan Kaufmann Publishers, San Mateo, 1989) 391-396.
[60] D. Whitley, D. Starkweather, Genitor-II: a distributed genetic algorithm, J. Expt.
Theor. Artif Int. 2 (1990) 189-214.
[61] D. Whitley, K. Mathias, P. Fitzhorn, Delta coding: an iterative search strategy
for genetic algorithms, in: R. Belew, L.B. Booker, Ed., Proc. of the Fourth Int.
Conf. on Genetic Algorithms (Morgan Kaufmmann, San Mateo, 1991) 77-84.
[62] T. White, Adaptive crossover using automata, in: Y. Davidor, H.-P. Schwefel, R.
Maenner, Eds., Parallel Problem Solving from Nature{PPSN III, International
Conf. on Evolutionary Computation, Proceedings, volume 866 of Lecture Notes
in Computer Science (Berlin, Germany, Springer-Verlag, 1994) 228-237.
[63] S. W. Wilson, Classi er system learning of a boolean function. Research Memo
RIS-27r, Rowland Institute for Science, Cambridge Mass (1986).
[64] H.Y. Xu, G. Vukovich, A fuzzy genetic algorithm with e ective search and op-
timization, in: Proc. of 1993 International Joint Conference on Neural Networks
(1993) 2967-2970.
[65] H.Y. Xu, G. Vukovich, Y. Ichikawa, Y. Ishii, Fuzzy evolutionary algorithms and
automatic robot trajectory generation, in: Proc. of The First IEEE Conference on
Evolutionary Computation (1994) 595-600.