You are on page 1of 16

Genetic algorithms for fuzzy control

Part 1 : Offline system development and application


D.A. Linkens
H.O. Nyongesa

Indexing terms: Adaptive control systems, Biomedical control, Evolutionary adaptation, Furzy logic control, Genetic algorithms, Machine learning

Abstract: Although fuzzy logic controllers and


expert systems have been successfully applied in
many complex industrial processes, they experience a deficiency in knowledge acquisition and
rely to a great extent on empirical and heuristic
knowledge which, in many cases, cannot be objectively elicited. Among the problems to be resolved
in fuzzy controller design are the determination of
the linguistic state space, definition of the membership functions of each linguistic term and the
derivation of the control rules. Some of these
problems can be solved by application of machine
learning. First, it is desirable to simplify and
automate the specification of linguistic rules. Secondly, it is also desirable that modification of
control rules is possible in order to cope with previously unknown or changes in process dynamics.
Machine learning methods have, in recent years,
emerged from the use of learning algorithms modelled on natural and biological systems. These
methods attempt to abstract the advanced mechanisms of learning exhibited by such systems, which
can, consequently, be applied to intelligent
control. One of these new algorithms is the genetic
algorithm which is modelled on the processes of
natural evolution. The paper develops the application of genetic algorithm techniques for fuzzy controller design. Genetic algorithms are used to
automate and introduce objective criteria in defining fuzzy controller parameters.

There is to-date no generalised method for the formulation of fuzzy control strategies, and design remains an ad
hoc trial and error exercise.
This paper discusses a new, objective approach to the
design of fuzzy controllers. The approach uses genetic
algorithms, a recent search and optimisation technique
[4,51, to optimise the parameters of a fuzzy controller. A
similar approach was recently reported in References 6
and 7 in which the method was used for the optimisation
of the membership functions of a fuzzy control rule-base.
The present study, however, extends the method to the
complete process of design of a multivariable controller
for a nonlinear biomedical process, namely, control of
anaesthesia.
2.1 Fuzzy design methodology

A typical design methodology for fuzzy controllers usually


follows the iterative steps shown in Fig. 1 [SI. I t is noted
define an operational
]model of the Drocess

define fuzzification and


defuzzif icot ion
scaling factors

tune rules. scaling factors


and membership to
achieve performonce levels

introduction

Several studies have shown fuzzy logic control to be an


appropriate method for the control of complex continuous unidentified or partially identified processes [1-33,
many of which cannot easily be modelled in a mathematical way. This is because, unlike a conventional
process controller such as a PID controller, no rigorous
mathematical model is required to design a good fuzzy
controller and in many cases they can also be implemented more easily. However, this simplicity with which they
can be implemented also presents a bottleneck in their
design. Fuzzy controllers rely on heuristic knowledge
that is subject to the designers interpretation and choice.
Q IEE, 1995
Paper 1766D (C8). received 10th March 1994
The authors are with the Department of Automatic Control and
Systems Engineering, University of Shefield, PO Box 600 Mappin
Street, Shefield SI 4 D U , United Kingdom

IEE Proc.-Control Theory Appl., Vol. 142, N o . 3, M a y 1995

Simulate process
cesign

Fig. 1

Fuzzy design methodology

that the first step in the fuzzy design procedure is to


obtain an understanding of the process dynamics. This is
necessary because, as is the case with other types of controllers, it is not possible to design a controller without
assuming certain characteristics about its environment.
However, in the case of fuzzy controllers this requires a
less rigorous model of the process that may be expressed
merely as estimates of the gain sensitivity, the system
delays and an estimate of the order of the system.
161

The second stage of the design process is to define the


boundaries of the fuzzy universe of discourse and the
number of partitions within it. Most studies have used
universes of discourse utilising fuzzified units that range
from + 6 to 10, and partitions of between 3 and 9 fuzzy
sets [9-111. In addition, the universes may be discrete or
continuous functions. Modification of a fuzzy universe of
discourse and its partitions alters the control surface and
thus can be used as a means of tuning the controller. On
the one hand it has been argued, by Daley and Gill [9],
that an even number of partitions should be used because
this provides control adjustment on either side of the setpoint to achieve a zero mean steady state error. On the
other hand, Cox [SI suggested that an odd number of
partitions should generally be used. Either of these points
of view, however, have not been strongly supported by
other published studies. The genetic algorithm technique,
applied in this study, uses binary coded strings which
naturally suggests an even number of coded parameters.
After specifying the fuzzy universe and the partitions
therein, the membership functions of the fuzzy sets can
then be defined. The shapes of membership functions
used in most studies are continuous triangular, or gaussian functions, although discretely defined functions have
also been used in other studies [9, 1 1 , 123. Gaussian functions are considered suitable when the controller rulebase has very few rules [13], since the function has
nonzero membership grades over the whole fuzzy universe. However, when the number of rules is large, gaussian functions result in the firing of several rules, some of
which may be antagonistic. In this case, it is necessary to
limit the number of rules which fire by defining a membership cut-off level, cc-cut. Membership functions used in
the present study are continuous triangular functions.
Triangular membership functions are simpler to use since
it is only necessary to specify where the peaks of fuzzy
sets are located in the universe, and how wide are the
fuzzy sets. It is also generally agreed that a certain
amount of overlap of the fuzzy sets is desirable to provide
continuous and smooth transitions on the control surface
[14]. However, since the fuzzy controller is nonlinear the
amount of overlap and positioning of the fuzzy sets is
much dependent on the process under consideration, and
hence subject to tuning and modification, in order to
obtain the required performance.
The third stage in the design process is to decide a
fuzzification and defuzzification strategy, to convert real
measurements and data to the fuzzy domain and vice
versa. Fuzzification and defuzzification can be viewed, in
a simplified way, as choosing scaling factors for each of
the predicates of a fuzzy relation. This step is critical in
the design process because the scaling factors have a very
profound effect on the other parameters of the fuzzy controller 1151. In fact, it is usually easier to find a good set
of sensible rules and membership functions to apply as a
starting point in an iterative design scheme, than to find
appropriate scale factors. In the fourth and final stage,
the control rules are derived. This is the set of IF ...
THEN . . . linguistic relations that form the expert knowledge of the controller. The traditional method of obtaining these rules is a heuristic trial-and-error approach
based on analysing process behaviour, and consequent
iterative modification to obtain acceptable performance.

2.2 Problems in fuzzy design


Although they have been applied in many complex industrial processes, fuzzy logic controllers and expert systems
experience a deficiency in knowledge acquisition, and rely
I62

to a great extent on empirical and heuristic knowledge


which in many cases cannot be elicited objectively.
Among the problems to be resolved in fuzzy design are
the determination of the linguistic state space, definition
of the membership grades of each linguistic term and the
derivation of the control rules. The traditional methods
through which this sort of information has been gathered
include interviews with experienced process operators,
process knowledge experts, or other sources of domain
knowledge and theory. More recently, fuzzy modelling
techniques have been studied by Cellier and Alvarez [16]
and Linkens and Shieh [17].
The traditional approach to fuzzy design, as outlined
above, is laborious, time consuming and in most cases
specific to each application. Other shortcomings that
result from the approach include the following:
(i) Process operators usually cannot easily translate
their knowledge and experience into an algorithmic or
rule-based form necessary for conversion to an automatic
control strategy.
(ii) The expert knowledge is not always available.
(iii) Multivariable fuzzy control remains very difficult
in terms of eliciting the rule-base.
2.3 Learning in fuzzy controller design
Some of the problems encountered in fuzzy design can be
solved by application of machine learning, and in particular the use of genetic algorithms. First, it i s desirable to
simplify and automate the specification of linguistic rules,
their membership functions, and the scaling factors. Secondly, it is also desirable that modification of control
rules or their membership functions be possible in order
to cope with previously unknown or changes in process
dynamics. This would also mean, for example, that
knowledge-bases can be applied more generally, only
requiting adjustments to either the rules, membership
functions or scaling factors in order to be suited to a different control problem.
The choice of a learning method depends on the
nature of the task domain and the available information.
Two types of learning methods have hitherto been used
to generate rules, depending on the available information.
If sufficient input/output data can be obtained then
inductive learning methods [16, 181 can be used. On the
other hand, if extensive domain theory or sources of
expert behaviour exist then explanation based methods
are applied [19]. There are however, situations in which
neither of the above information is available. One
method of developing rules in such cases involves testing
hypothetical trials on a model of the task, or indeed the
actual task itself if it can tolerate trials. Better trials can
then be evolved based on assessment of previous results.
Thus, two learning approaches can be envisaged, ofiline
learning when a simulation model is available and online
learning, when no suitable model can be obtained. This
paper considers the offline learning technique, using a
genetic algorithm as the learning medium. The online
approach is discussed in Part I1 of this paper. In both
approaches the genetic algorithm is applied to automate
and introduce objective criteria in defining fuzzy controller parameters. In previous studies, Karr el al. [SI, Karr
[20] and Thrift [21], used the offline techique to derive
membership functions for predefined sets of rules. In the
present study, the genetic learning method i s extended to
the complete process of fuzzy design, as outlined in Fig.
1, which includes elicitation of control rules and optimisation of their membership functions. Furthermore, the
genetic algorithm method is able to derive the fuzzy
IEE Proc.-Control Theory Appl., Vol. I42, No. 3, May I995

knowledge-base without following the step-by-step iterative process.


The incorporation of genetic learning into a fuzzy
design process adds an intelligentdimension to the fuzzy
controller enabling it to create and modify its rules.
Genetic algorithms give the possibility of adjusting membership functions down to the level of individual rules.
This is unlike usual fuzzy knowledge-based systems in
which the arguments of rules are constants or variables
that have been previously bounded, and which retain
these bounds whenever the rules are used. Using a
genetic algorithm, the rules can be allowed to develop
new bounds for their individual predicates, such that the
same variable may have different bounds in another rule
at the same time. or in the same rule at a different time.
3.1 An overview of genetic algorithms

Genetic algorithms (GA) are exploratory search and optimisation procedures that were devised on the principles
of natural evolution and population genetics. The basic
concepts of GA were developed by Holland [22,23], and
subsequently in several research studies. Goldberg [5]
and Davis [4] provide recent comprehensive overviews
and introductions to GA.
There are several differences between the functioning
of C A and traditional optimisation techniques, especially
those based on gradient methods. These differences
include the following:
(i) GA work on a coding of the parameters to be optimised, rather than the parameters themselves. Binary
coding is normally used and has been suggested to be
optimal in certain cases, however this is neither a
restriction nor a requirement of GA.
(ii) GA search a space using a population of trials,
representing possible solutions to the problem. The initial
population, usually, consists of randomly generated
individuals.
(iii) GA use an objective function assessment, to guide
the search of the problem space.
(iv) GA use probabilistic rules to make decisions.
Typically, the GA starts with little or no knowledge of
the correct solution depending entirely on responses from
an interacting environment and its evolution operators to
arrive at good solutions. By dealing with several independent points the GA samples the search space in parallel,
and hence is less susceptible to getting trapped on suboptimal solutions. In this way, GA have been shown to
be capable of locating high performance areas in complex
domains without experiencing problems with high
dimensionality, or false peaks, as would occur with hill
climbing optimisation techniques.
As has been stated, GA imitate natural evolution, and
hence include operations such as reproduction, crossover, mutation and inversion. A simple GA has four
features; Population size, Reproduction, Crossover, and
Mutation. Reproduction is a process in which a new generation of population is formed by randomly selecting
strings from an existing population, according to their
fitness. This process results in individuals with higher
fitness values obtaining one or more copies in the next
generation, while low fitness individuals may have none;
it is for this reason called a survival of the fittest test.
Crossover is the most dominant operator in a CA, and is
responsible for producing new trial solutions. Under this
operation, two strings are selected to produce new offspring by exchanging portions of their structures. The
offspring may then replace weaker individuals in the
I E E Proc.-Control Theory Appl., Vol. 142, N o . 3, M a y 1995

population. Mutation is a local operator, which is applied


with a very low probability of occurrence, typically 0.001
per bit or less. Its function is to alter the value of a
random position in a gene-string. When used in this way,
together with reproduction and crossover operators,
mutation acts as an insurance against total loss of any
gene in a particular position throughout the population,
using its ability to introduce a gene which may not,
initially, have existed due to limited population size or
was lost through application of the other operators. For
a given population of trials and set of operators together
with procedures for evaluating each trial, a GA proceeds
as follows:
(a) An initial random population of trials. n(0)=
A,(O), rn = 1, ..., M , where M is the number of trials in
the population, is generated.
(b)For successive sample instances:
(i) The performance of each trial, p(A,,,(T)).T = 0, 1,
2, .. .,is evaluated and stored
(ii) One or more trials are selected by taking a
sample of n(T)using the probability distribution
lhf

(iii) One or more genetic operators is applied to the


selected trials in order to produce new offspring,
A i ( T ) , m = 1, ..., N , where N is the number of offspring which is usually equal to the number of selected
trials (parents).
(iv) The next generation of population, IIl T + 1) is
formed by selecting AAT) E n(T),j = 1, . . . , N to be
replaced by the offspring, A J T ) . The criterion for
selecting which trials should be replaced may be
random, on the basis of the least fit or some other
fitness basis.
(c) The GA process is terminated after a prespecified
number of generations or on the basis of a criterion
which determines convergence of the population.
One of the attractions for researchers in C A as a search
and optimisation technique was that it could obtain
near-optimal solutions to different types of problem
without knowledge of their specific task domain, by
simply manipulating bit strings. GA were in most of these
cases applied to function optimisation over real R object
spaces. However, it was soon noted that the choice of a
number of control parameters and even the representation itself could severely affect the performance of the
GA, when applied to other types of problems. Running a
successful C A involves having to find settings for a
number of control parameters, which is not a trivial task.
The control parameters include: population size, and the
nature and rates of the recombination operators; crossover, mutation and reproduction. A number of studies
have investigated the selection of these parameters, and
there have been attempts to overcome the representation
difficulties by taking advantage of any available a priori
knowledge about the task problem, together with the
design of specialised coding and genetic operators. Representation issues have been studied by Antonisse and
Keller [24], Caruana and Schaffer [25] and more recently by Schraudolph and Belew [26] who investigated
online dynamic encoding of parameters. Grefenstette
[27], Davis [28], Schaffer et al. [29], among others, have
studied control of the parameters of a GA. Indeed, it has
been demonstrated in many studies that a specially
adapted GA will outperform other GAS on a specific
problem. Specialised genetic plans are derived by controlI63

ling the rates of the tripartite processes; reproduction,


crossover and mutation.
In learning applications, as opposed to function optimisation, one is not necessarily interested in finding
optimal solutions to problems. This is partly because the
global optimal solution is not even known. In such cases,
the CA is used as a learning heuristic that finds good
knowledge structures quickly, and which may or may not
be optimal [30]. This is the course followed in this study,
that is, to find a C A that derives knowledge-bases for
satisfactory fuzzy control, as quickly as possible. To
improve the performance of a GA there is always a need
to maintain a balance between exploration and exploitation of the search space. Exploration is related to the
number of new trials carried out, while exploitation is
related to the number of good candidates retained in a
population to produce better trials. An appropriate selection of the various genetic operators forms a suitable
genetic or evolution plan.

Fig. 2 illustrates a typical coding format for a population of fuzzy rule-bases comprising two conditions E
and CE, and one action U.This is the structure of rules
adopted in many fuzzy control applications, where E

3.2.1 Coding the genetic algorithm: The type of coding

used in this research is the classical concatenated binary


mapping. This coding joins together segment codes of all
parameters into one composite string. Let the code of the
ith parameter in the n-dimensional 9search space be
designated U,. When the usual binary alphabet is used U
is an [,-tuple string of Is and Os; ui++{0, 1}li, where Ii is
the length of the code. Thus, each trial structure using
concatenated binary mapping is given by
A

I
rule 8

31-29

28-24

Fig. 2

23-M 19-15 14-11 10-6 5-2

A typrcal coding offuzzy-rule base

represents the control error, C E the change-in-error and


U the process control input. In this work, it was decided
to contain the coding of a fuzzy rule within the basic data
structure of the particular computer hardware used,
which is 32 bits. This simplifies the manipulation of the
codes. Eight linguistic fuzzy sets, namely [Negative Big.
N B , Negative Medium N M , Negative Small N S , Negative Zero NZ, Positive Zero P Z , Positive Small PS,Positive Medium PM, Positive Big P E } , were specified for
either of the antecedents of a fuzzy rule. Hence, a rulebase consists of a maximum of 64 rules. By representing
the total coding for the control strategy in the usual form
of a two-dimensional rules table, the coding of the rule
antecedents, E and CE, can be made implicit by the positions of rules in the table. For example. all rows are
associated with one value of E and all columns are
associated with one value of CE. Therefore, explicit codes
will only be required for the peaks and the widths of
these fuzzy variables.
To obtain near-optimal attributes of a rule, its parameters are allowed to change. Thus, each of the linguistic
values, the positions of the peaks and widths of fuzzy sets
of the rule can change. Fig. 3 illustrates the representa-

En = ,y ui = [d,U * , . . ., U]

The structure of a fuzzy rule is made up of the codings of


the fuzzy sets of the linguistic predicates of the rule. Furthermore, a linguistic variable is defined by its membership function which, in the case of the chosen triangular
shape, is determined by the position of its peak and the
spread of its nonzero grades. Therefore, typically, each
linguistic fuzzy variable can be represented by three partcodes; one for its linguistic value (or name) l, one for the
position of the peak p and the third for the width of the
set w
CUI up
(3)
Considering a fuzzy rule with n conditions and rn actions
the composite coding for a trial rule, A E n is given by

A =
. . . uF, U:. . . urn]
(4)
where U;, or U;. are the codes of the ith condition or
action fuzzy sets.
164

3.2 Genetic algorithm for fuzzy controller design


A genetic adaptive plan can be defined as a quadruple
A = {Z, nN,
@, R}, where Z is the coding format, nNis a
population of size N , @ is a fitness rescaling algorithm
and R = [U,, 02,
. . ., U,] is the set of genetic operators.
In this research, three CA operators are used: reproduction U , , crossover w, and mutation om.By genetic
plan is meant the processes through which successive
populations are generated using evaluation, selection,
mating and deletion. Let Y be a probability distribution
over II which is derived from the fitness of each trial,
p(A E n). A genetic plan can then be formally expressed
as the mapping A : (Y x n x Q) + W. This paper investigates the efficacy of some of the more common GA
adaptive techniques, with a view to deriving a genetic
plan suited for fuzzy learning.

rule 1

Fig. 3

Linear partition fuzzy universe

tion of typical membership functions when the universe is


linearly partitioned. These are fuzzy sets of a static rulebase, as found in a traditional fuzzy controller. Fig. 4
shows alternative membership functions of the same
fuzzy set, in this case N M , when the peaks and widths are
allowed to change. The coding of the jth parameter in a
string of codes represents a value, l j between predetermined minimum and maximum constraints given by
Lj

= Ami,

+ (Amx

A,JA(u)

(5)

I E E Proc.-Control Theory Appl., Vol. 142, No. 3, M a y 1995

where I,, is the minimum parameter constraint, I,, is


the maximum constraint and A(uJ) is a real number in
[0 . . . 11 that indicates the relative magnitude of the jth

where 0,is the initial number of schemata in a population, 0,is number of schemata in a population at convergence, and T, the time taken to reach convergence.
The number of different schemata in a converged population is the number of different schemata represented by
a single string, and is given by 2, where I is the string
length. For example, a 2-bit binary string {lo} can represent the schemata { 10, *0, 1*, **}. Thus, the number of
schemata in a population of N different individuals is
equal to 2. However, this is true for small size populations and the number increases to 3 for very large
populations when nearly all possible strings are represented in the population. Hence, the figure of merit
proposed above can be expressed as

[/A,
NM

Fig. 4

Alternative membership functionsfor N M

parameter within its bounding constraints. A ( d ) is dependent on the decimal value of U, Dec (U) E [O, 1, . . ., 2
- 11 and the cardinality of the coding, equal to 2 - 1
where, lj is the codes length. The decimal value of a code
is given by
D e c ( d )= ( i l u i 2 i - 1 )

(6)

where ai are binary code bits, numbered starting with the


least significant. Consequently, A ( d ) is equal to
(7)

3.2.2 Population size: The choice of an appropriate


population size is a fundamental decision to be taken in
all C A implementations. Many studies have investigated
population sizing including, the work of Goldberg [31]
and De Jong and Spears [32]. On the one hand, if the
population size is too small the C A will usually converge
too quickly, and in many cases to a poor solution due to
insufficient information in the population. Too large a
population, on the other hand, will take a very long time
to evaluate and in addition results in slow progression
towards its final solution. Therefore, when the evaluation
of a trial structure is computationally expensive, there is
usually a need to impose a limitation on the size of population that the C A can reasonably support. This means
obtaining a balance between the requirement for a large
information capacity in the population and the need to
produce a solution within a limited amount of time.
A theoretical basis for the sizing of populations has
been Hollands 0(n3) schemata estimate [23]. Simply
stated, this says that the number of schemata processed
by a C A is proportional to the cube of the population
size. This has been used in many studies to justify different, usually larger, population sizes. Grefenstette [27]
used population numbers between 10 and 160 in an
experimental design to optimise control parameters for a
GA, but also demonstrated the interacting nonlinear relations between population size and the recombination
rates. Other studies on population sizing include:
Odetayo [33] who used numbers between 100 and 400,
and Robertson [34] who used population sizes of up to
8000 for classifier systems.
A detailed study was carried out by Goldberg [31]
who analysed appropriate sizing of populations for serial
and parallel GAS. He devised a figure of merit which
compared the population sizing and convergence times
for different lengths of coding. This figure of merit was
expressed as

eo- 0,

A=-

T,
IEE Proc.-Control Theory AppL, Vol. 142, No. 3, May 1995

A=-

2N - 2
nc

(9)

where N is the population size, nc is the number of generation to reach convergence, and t, is the processing time
per generation. A computer program was developed in
the study to maximise the figure of merit for different
values of 1. A surprising result was that although the
required population setting increased linearly when the
time to reach convergence was fixed, it was found to be
constant regardless of the string length when the time to
reach convergence is allowed to vary. Perhaps more surprising was the fact that this population size turns out to
be equal to 3.
The above study provides a sound basis for selection
of a suitable population size, although it ought to be said
that the suggested population size of 3 should not be
taken rigidly. It happens to be the minimum number that
can allow selection of different pairs of individuals for
recombination. However, a population size of 3 does not
provide any meaningful diversity in the population.
Hence, there is a need to increase the population size to
achieve a good compromise between diversity and convergence time. In the current study, population sizes
between 20 and 40 were used. Although these could be
regarded as very small sizes in view of the string lengths
(typically 2080 bits) it has been found that even the smallest size achieved good results with suitable recombination operators. This result is supported by the work of
De Jong and Spears [32] and Syswerda [35] who established that recombination operators that are highly disruptive to schemata do actually achieve better
performance than the traditional one-point crossover in
GAS using small population sizes. In the study by Goldberg [31] it was suggested that the rapid convergence of
small populations can be used to advantage to improve
ofline performance, by reinitialisation of the population
when it becomes uniform. The GA in the present study
does not reinitialise populations but permits high rates of
disruption to existing populations, through increased
mutation.

3.2.3 Reproduction: Reproduction is the process


through which parent structures are selected to form
new offspring, by applying genetic operators, which can
then replace members of the old generation. Reproduction itself is also an operator which produces ofispring identical to its parents. The method of selecting an
individual to produce offspring (or to be deleted from the
population) determines its lifespan and the number of its
offspring. For example, if p , is the probability that an
individual, A E ll is selected to produce offspring during
a sample step and p2 is the probability that it will be
deleted during that sample step, then the expected
165

number of offspring of A is p , / p 2 [23]. The most


common reproduction techniques are generational
replacement, generational gap and steady state schemes
[4, 51. In this paper, a reproductive technique called
selective breeding is introduced. As suggested by the
name, this technique deterministically selects which ONspring can become members of a new generation. Selective breeding is proposed because it is envisaged that the
determinstic replacement of population members will
permit maximum exploration and exploitation within a
population, and also that sampling errors can be reduced
through the deterministic elitism. The four reproduction
techiques, generational replacement, steady state, generational gap and selective breeding are critically considered and compared as follows.
Generational replacement replaces an entire generation
by newly formed offspring. The offspring are created by
selecting parents on the basis of their strength, that is, a
fitter individual has a higher chance of being selected to
be a parent. Genetic operators are then applied to form
new offspring. The process is continued until the number
of offspring is equal to the population size, and the population of offspring becomes the new population of trials
for the next generation. This method maximises exploration of the search space through creation of an entire
generation of new trials. Its main disadvantage, however,
is that it often loses some of the best individuals in previous generations without a chance of passing on their
genes. Generational replacement is one of the earliest
reproduction techniques and many of the pioneering
studies in GAS have used the technique. Its performance
is, however, usually inferior to other common techniques.
Generational gap method is similar to generational
replacement except that it only replaces proportions of
the existing population by the newly created offspring. A
generational gap of 0.5, for example, implies that 50% of
the present generation is replaced by newly formed offspring. The individuals to be replaced may be chosen
randomly or some form of elitism (so that the worse
members of the population are the ones replaced) may be
incorporated. The replacement method, however, does
not guarantee that the offspring are any better than the
individuals they replace. Thus, at large generational gaps
this method is similar to generational replacement.
Steady state reproduction [35] overcomes one of the
problems in the generational methods, being the loss of
good individuals through sampling errors. In this
method, typically, only one or two new offspring are
introduced into each new generation to replace the worst
members. Syswerda [36] compared generational replacement and steady state reproduction methods and found
the steady state scheme to perform better. The method
encourages more exploitation of the population species
and hence tends to be very cautious, requiring several
generations to improve the population mean fitness. The
steady state reproduction method is elitist which means
that the best individuals in the population are always
passed into succeeding generations. It can be made to
behave similarly to the generational methods by introducing more offspring and randomly selecting the
individuals to be deleted.
Selective breeding is a reproduction method that is
designed to overcome some of the deficiencies in the
other methods. It is noted that even the elitist techniques
still experience sampling errors in the selection of parents
and deletion of individuals from a population, and often
good individuals can appear in a population and then be
lost without a chance to recombine. Selective breeding
166

introduces determinism in order to eliminate stochastic


sampling errors in deletion of candidates. The method
functions as follows:
(a) An initial population, II(0) is created in the usual
manner.
(b) The population is evaluated to determine the performance of each individual, p ( A , , m = 1, . . . , M ) .
(c)for successive generations, thereafter :
(i) An entire population of offspring, IIo(T) is
produced by selecting parents and applying genetic
operators
(ii) The offspring population is then evaluated
(iii) The next generation of population is obtained
by choosing the best M individuals from both n(r)
and no(
T).
It can be noted that selective breeding differs from the
other techniques in the manner of forming the next generation of population. It is a generational method in the
sense that an entire population of individuals is formed at
each instant. The replacement procedure which is likened
to pedigree selection in livestock breeding. however, is a
sorting procedure carried out to determine the best individuals from both the parents and the offspring populations, and is done in a way to ensure that no parent or
offspring is retained while a better individual dies. This
guarantees that the best individuals are always passed on
from generation to generation and will not be lost due to
sampling error or application of crossover and mutation
until they become the worst in the population. This
encourages exploitation of the good individuals. In addition, since high performing individuals never die, temporal diversity of trials can be maintained by continuously
disrupting the offspring when created, so that they are
significantly different from their parents. Temporal diversity is, diversity of populations from one generation to
another, as opposed to spatial diversity which is diversity
between population individuals in one generation. Temporal diversity enforces exploration especially in latter
generations as the population becomes more homogeneous. However, the disruption of populations (e.g.,
through high mutation rates) does not affect the oflline
(best-so-far) or online (running population mean) performance since no individual enters the population unless
it is better than the worst member. Selective breeding,
therefore, maximises both exploitation and exploration of
the population space.
It can be demonstrated that selective breeding, to
qualify as a genuine C A reproduction technique, allocates schemata as provided by the schema theorem [22,
231. The expected number of copies which will be allocated to any schema, according to the schema theorem, is
given by

where m is the number of individuals in the schema;


is
the mean fitness of the schema; 7i;; is the mean fitness of
the population. Since, under selective breeding the best
individuals are always retained in the population, the
number of copies of a high performing schema after one
generation is given by

Thus, the number of copies of the best schemata


increases, replacing lesser performing schemata until the
population is filled up with the best. The rate of increase
I E E Proc.-Control Theory Appl., Vol. 142, No. 3, May 1995

of schemata is, by contrast, faster than in the traditional


GA.

Traditionally, the
number of cross-over points used in a GA, which determines the number of segments of the parent codes that
are exchanged, has been set to 1. The supporting basis for
this choice is the theoretical result expressed by the
schema theorem. Specifically. the one-point crossover
produces the minimal disruption to schema allocation.
However, there have been experimental studies which
have suggested the use of more crossovers and higher
rates of mutation. The study carried out by Syswerda
[35], for example, indicated that more disruptive forms of
mating can improve the performance of a CA. A disruptive mating process is one which results in offspring that
are significantly different from their parents. Disruption
is considered advantageous when the population
becomes homogeneous, when it can improve the productivity of a C A by continued creation of new trials.
Under these circumstances, more conservative operators,
such as the one-point crossover, produce offspring that
are identical to !he parents and, hence, the search stagnates. A latter study by De Jong and Spears [32] established that the improvement in performance was more
with smaller population sizes than larger ones.
There are two types of multiple crossover, namely
n-point crossover and uniform crossover [37]. In n-point
crossover, n breakpoints are chosen at random at which
points the parents bits are alternately passed on to the
offspring. Offspring take genes from one of the parents
until a crossover is encountered, at which point they
switch and take genes from the other parent. Another
crossover operator, called punctuated equilibria. functions in a very similar way except that the crossover
points are themselves coded in a separate chromosome,
which is also subject to genetic operators. In uniform
crossover there is a probability p , that the offspring will
take a bit from the corresponding positions of either
parent. For example, when pc is 0.7 an offspring will
acquire the genes of one of the parents with a probability
of 0.7 (and the genes of the second parent with a probability of 0.3). A value of p , equal to 0.5 thus implies
random crossover since the offspring take on bits from
the parents with equal likelihood, while p , equal to 1.0 or
0.0 creates offspring that are clones of the parents. The
mechanisms of both types of crossover are illustrated in
Fig. 5.

3.2.4 Crossover and mutation:

breok-points

t = e x c F n g e c bits
Fig. 5
il

n-Point and uniform crossover

(n = 3tpoint crossover

h uniiarm crossover

Mutation is a local operator that transforms the bits


of a G A construct, replacing 1 with 0 and vice versa.
Traditionally, the operator has been used as a random
search process that was infrequently applied. Recent
I E E Proc.-Control Theory Appl., Vol. 142, No. 3, M a y 199.5

studies, however, have shown the mutation operator to


play a more crucial role in genetic search. Studies by
Muhlenbein [38] and Back [39], for example, have
showed that the optimal rate of mutation is proportional
to the length of the code and the modality of the
problem. For example, it was shown that the optimal
rate of mutation for a k-order function is k/l, where I is
the length of the coding. Hence, in this study some determinism was introduced into the mutation operator to
ensure that m bits selected at random were altered. For
small rates of mutation (when the probability of mutation
p,
1/1) it is possible that no mutaton of the string will
have occurred after testing all 1 loci. This approach also
speeds up the computation involved in the mutation
process, especially when I is large because instead of
testing all loci for the chance to mutate only m (m < I )
positions are selected at random and flipped. Mutation
was used in this study to create disruption of the population, rather than as a local improvement operator or to
delay convergence, as is usual.
Crossover and mutation in context sensitive structures
may require careful application to avoid generating obviously unsuitable trials. For example, exchanging the linguistic values of different fuzzy rules can mean that the
new offspring acquire membership functions that are
obviously not suitable for them. Similarly, since the linguistic codes identify the premises of rules, mutation and
crossover within these codes will create new premises
that are merely duplicates of other rules. In this research,
crossover is restricted to rules with the same goals, and
no mutation is carried out on rule premises.

3.2.5Fitness scaling: The schema theorem suggests that


copies are allocated to competing trials in a population
according to their mean fitness. Thus, if two similar individuals in a population were to be assigned different
fitness values, then the number of copies each receives in
a new generation of population would be proportional to
their assigned fitnesses. Scaling or transformation of the
population fitness can, therefore, be used to exert selective pressure between competing trials in a population.
Two scenarios are usually encountered. In the early generations of the search there are only a few good individuals in a population, which may then tend to dominate
the selection and mating processes. They may rapidly
reproduce among themselves, leading to a premature
convergence. Later on in the search, most individuals in a
population tend to be very similar and the search
becomes nearly random because there is no preference of
any one individual over the other. These two conflicting
interests are examined in this paper; that is, the need to
encourage diverse interbreeding in the early stages of the
search, and the improvement of the mean fitness of the
population, by exerting selective pressure to favour the
best individuals.
There is no universal agreement on optimal fitness
scaling strategies and few studies have considered a theoretical treatment of the problem. However, many studies
use either the ranking method [40] or the window
method [27]. In an attempt to improve the GA used in
this study, another fitness method is introduced based on
a desire to overcome the difficulties experienced with the
other methods. The new method called functional normalisation is a dynamic rescaling technique which continuously adjusts the fitness values to maintain a desired
selective pressure between the best and the worst trials.
Investigative comparisons of the three fitness methods
are carried out.
167

'Ranking' was introduced into GA practice by Baker


[40], and is a technique in which the actual objective
evaluations are replaced by new values assigned on the
basis of their ranks in a population. Thus, after objective
evaluations have been carried out, the individuals are
sorted in terms of the magnitudes of evaluations, and
then the corresponding members of the population are
assigned new fitness values that are only a function of
their ranks. The assignment is usually a contant decrement (or increment) from an upper (or lower) bound. For
example, if the best individual may take a value of 100,
the succeeding ranked individuals can have values of 90,
80, and so on. Alternatively, the ranking may be on the
basis of a ratio between close individuals, such that in the
present example the values taken may be 100, 90, 81 and
so on, where the chosen ratio is 0.9. Minimum fitness
values are in general restricted. Ranking assigns fitness
values according to the equation
i j =
+ imin
(12)
where y: is the rank of an individual in a population, or
the number of individuals which have a worse evaluation
than the individual j at instant t, A is the fitness differential between neighbouring individuals and imin
is the
minimum restricted fitness.
One of the problems with a ranking technique is that
because it uses only the perceived rank of the individual
in a population it tends to exaggerate small and perhaps
insignificant differences in evaluations, while treating
larger differences in exactly the same way. This means
that in the early stages of the GA the poorer individuals,
which there will be more of in the population, will tend
to dominate the recombination process.
Windowiny methods assign fitness values by defining a
minimum value of fitness for each generation. For
example, new fitness values may be given by

[ .J = & . -J
&m m
(13)
where cminis a base-line value of evaluation. Often E,,,;" is
not known a priori and the popular strategy is to continually assign it values which have been observed in previous generations. For example, if all individuals in a
population have evaluations in the range C7.1 ... 9.23 a
fitness base-line could be set at 6.0, and hence, by subtracting this number from all evaluations, new fitness
values will be in the range [1.1 ... 3.21. The best individual now has a three times chance of selection over the
worst.
Functional normalisation is a new technique considered
in this study. One of the most common fitness normalisation techniques is linear normalisation [SI, which
scales the objective evaluations according to the equation

i j = ai+ b
(14)
The coefficients a and b are chosen to ensure that both
the original objective evaluation and the scaled fitness
have the same mean, and that the maximum value of the
scaled fitness is a certain multiple of this mean. One
shortcoming of the technique is that it does sometimes
result in negative fitness values and it is, therefore, necessary to take precautions to deal with this. Functional
normalisation, on the other hand, 'stretches' or 'compresses' the set of evaluations to maintain a given ratio of
minimum to maximum fitness values, which is called the
selection pressure K. This is done by calculating new
maximum and minimum fitness values to give the desired
ratio in each generation. The intermediate objective
evaluations can then be interpolated linearly, or nonlin168

early. Linear transformation of a set of objective evaluations, for example, is carried out according to the
following steps:
1 Calculate the new range of fitness values to obtain
the required pressure, K,

where
and E,, are observed minimum and maximum
evaluations and K is a specified value of the selection
pressure required
Kimin
(16)
2 Calculate the slope of the (linear) fitness transformation
imax

3 Assign new fitness to each individual by


= imin
+ ATEj - Emin)
(18)
Functional normalisation reduces the preference between
very good individuals and the poorer ones in the early
generations when the population is still diverse, in order
to encourage interbreeding. However, it will exhibit
marked differences between similar individuals in a nearhomogeneous population. This behaves similarly to
ranking, except that it preserves the proportionality of
the objective evaluations. Whereas ranking assigns predetermined fitness values based only on the perceived rank
of an individual, functional normalisation dynamically
determines new ranges necessary to exert a constant
selection pressure, and assigns fitness at points proportional to an individual's objective evaluation. This
removes the bias created by a ranking technique: thus,
individuals that are not significantly different in objective
evaluations will also not be significantly different after
rescaling, and vice versa.
The genetic processes, reproduction, crossover and
mutation, and the fitness rescaling techniques can be
combined in different ways to suit a GA to a particular
application, and because of the interdependence of
these processes the derived genetic plans may perform
differently.
ij

3.3 Evaluation of trials


Each individual (genotype) in a population is a hypothetical knowledge-base for a fuzzy logic controller,
containing a complete set of linguistic rules, their membership functions and a fuzzification/defuzzification strategy. The procedure of evaluating these knowledge-based
consists of submitting each to a simulation model, and
returning an assessment value according to a given cost
function. A controlled process is defined by a set of state
variables X = {xl, x2, . . . , xN} which are controlled by a
set of control variables C = {cl, c 2 , ..., c M } .The genotypes are trial 'control policies', for selecting C as a function of Xi.The role of the adaptive plan is to derive an
optimal policy A,,, which minimises a given performance
function, J. In many cases, J is determined as a cumulation over time of some instantaneous cost rate, Q[X(t),
C(t)].i.e., J =
Q[X(t), C(t)]. As an example, a trial
knowledge-base can be made to control the model of a
process and then sum the errors over the response trajectory. The sum of errors is then directly related to the
objective fitness of the trial. The configuration of such a
learning scheme is shown in Fig. 6.

Er=,

I E E Proc.-Control Theory Appl., Vol. 142. No. 3, M a y I995

The fitness of a trial is a measure of the overall worth


of a solution which takes into account factors of an
objective criterion, in this case, the performance of a

(fuzzy controller)

(cost function)

learning model
(senetic alqorithm)

I
Fig. 6

populotion model
(trials)

All these add up to a very taxing computational load in


terms of execution times. Consider, for example, the
anaesthetic muscle relaxation process which has a time
constant of approximately 2000 s. If the sample interval
is set to 20 s, 500 samples are taken for each set of initial
conditions. A simple calculation shows that the evaluation of each individual takes 25000 iterations. Thus, a
population of 20 individuals evolved over lo00 generations will have executed 500000000 iterations of the
Runge-Kutta routines alone. In terms of execution time,
if it takes 10 s to evaluate a trial knowledge base, a population of 20 individuals evolving over lo00 generations
will have taken more than 55 h to evaluate. This demonstrates the need to derive an algorithm that obtains a
suitable solution in fewer generations.

Ofline learning model

3.4 improvement of a genetic algorithm

fuzzy controller implementable with the trial knowledgebase. The basic control objective is simply stated as the
ability to follow a setpoint with minimal error. This
objective can thus be expressed in terms of minimisation
of controller performance indices which are in common
use. These include integral of absolute errors (IAE),
integral of square errors (ISE) and integral of time
multiplied absolute errors (ITAE). Each of these indices
has its own merits. For example, ITAE penalises errors at
large values of time and leads to reduction in steady state
errors at the expense of transient errors, while ISE is a
more suitable mathematical analysis criterion. Furthermore, the GA is only able to optimise the characteristics
explicit in the cost function. In this work, it was found
necessary to incorporate a penalty of excessive control
effort into the evaluation cost function, in particular to
prevent relay-type control. Thus, an index called integral
of absolute control rate and error (IACRE) was derived.
A general mathematical equation for such an evaluation
cost function is
7

=
k = l

[n(y(k)- r(k)) + @(u(k) - u(k - l))]

(19)

where y(k) is the process output, r(k) the reference input,


u(k) the control signal, and II and 0 are simple scaling
functions. 1was in this case, the sum-of-absolute values
operator. Because this cost function is minimised, the
least value returned is the best. Hence, to obtain an
objective fitness from the trial evaluations it is required
to reverse the orders of magnitude of the assessments
while maintaining their relative ratios. Objective fitness
assignment is performed according to the equation
N

p . = i=

6;

(20)

where p is the objective fitness, E is the cost function


assessment given by J , and N is the size of the population
of trials.
Evaluation of trials is the most demanding part of a
GA in offline learning. Consider the following case. Simulated control and evaluation of trials were carried out by
a numerical integration algorithm, using the fourth-order
Runge-Kutta method. Each trial knowledge-base was
allowed to control the model for a period equivalent to
approximately five open-loop time constants of a process,
and for five cycles from different initial conditions. Furthermore, in order to obtain more accurate simulation
using the Runge-Kutta method, it is usually desirable to
perform several integrations within a sampling interval.
I E E Proc.-Control Theory Appl., Vol. 142, No. 3,M a y 1995

There are other factors which could improve the performance of a GA, especially for fuzzy controller design.
These include: the need to speed up the evaluation using
parallel processing, the need to reduce the size of the
rule-bases, and improvement of the learning process with
knowledge-based genetic operators. The time taken in
evaluation of genetic structures imposes restriction on the
size of the population and also the number of generations
required to run the GA to a final solution. To alleviate
the problem, parallel processing can be used to reduce
generational execution times. There are two common
approaches to parallel processing in GAS. The first
approach is a one-population GA which only uses the
multiple processors to divide the evaluation task. Each
processor evaluates a given number of individuals and
returns their objective assessment to the population. The
second alternative is to have separate subpopulations on
each processor which develop individual solutions. In
such cases, usually, there is occasional migration of
members between the subpopulations. This model of
GAS is commonly known as a parallel or distributed GA
(PGA) [41, 421. Although studies have shown this mode
of GA to yield better solutions than the one-population
GA, it does not solve the problem that is of concern to
this study: reducing execution times. Thus. the first
method of parallelisation is preferred, because the
reduction times takes precedence over any possible
achievement of an optimal solution with a PGA.
Human reasoning and decision making are guided by
knowledge. Furthermore, fuzzy logic systems are
designed to imitate this reasoning mechanism and are
characterised by rules that are expressive knowledgebased models of the task domain. In contrast, C A in their
basic form are dependent only on coding, function evaluations and exchanges of the codes between individuals.
This is both an advantage and a disadvantage. Although
the ability to work without problem specific information
allows GAS to be applicable to different problems
without specific tailoring, failure to make use of available
domain knowledge would definitely put them at a disadvantage against other methods that could use such
information. One way of utilising available knowledge
would be to use it to seed the initial populations. The
advantage of this is that it can result in a reduction in the
number of generations required to reach a satisfactory
solution. Another way of utilising available knowledge is
to use knowledge-based genetic operators that exploit
any known regularities in the domain space. The evaluation of a genetic structure representing a fuzzy rule is
carried out at a knowledge level since the structures are
first interpreted into fuzzy rules, and then applied to the
169

task domain. It would, therefore, be desirable that the


processing and modification of these genetic structures
should also be conducted at the knowledge level. This
study proposed simple knowledge-based operators for
the GA-fuzzy learning system. Knowledge-based crossover is operated at the conditional level, and exchanges
complete segments representing one or more parameters
of a fuzzy set. Similarly, mutation when applied to a linguistic code alters it to another valid linguistic code,
which in this research was restricted to a linguistic distance of two, to avoid possible large deteriorations in
performance. These are illustrated in Figs. 7 and 8.

variable anaesthetic model is given by

[;I=[

(1

1.&-(1 + 10.64s)
+ 3.08sX1 + 4.81s)(l + 34.42s)

(1

+ 2.0s)

crossover point
I

c L!d] e

[ A

Fig. 7

a I b

IC I

D ] e I f

parents

offspring

1 ID]

A l B I C I d ] E I F

F ]

Knowledge-based crossover
F are code segments represenling linguistic names and fuzzy sets of rules

a .. . f and A . . .

Fig. 8
Knowledge-bused mufation
Mutation changes the linguistic fuzzy set to another. a distance of one or two
away

Learning and simulation results

This paper investigates the development of a suitable C A


technique for fuzzy design. A C A may be derived in
several ways depending on how its parameters are
chosen. Population size, for example, was a constraint in
view of evaluation times for large populations. Hence, it
was reasonable to decide on a suitable population size as
a starting point. The population size was fixed at 50,
although similar results have been obtained with different
sizes of population. Likewise, the genetic operators used
comprise the multiple-point crossover, and mutation
operators discussed previously. To determine the most
suitable choice of the other parameters of a genetic plan,
the reproduction and fitness techniques, also discussed
above, were then compared in a series of controlled
domain experiments. The best genetic plan is the one
which achieves the stated objective: to achieve the best
solution using the smallest number of evaluations.
The GA is used to derive rules for a multivariable
fuzzy controller. The fuzzy controller is then, applied to
the control of multivariable anaesthesia, comprising the
simultaneous regulation of paralysis and the depth of
unconscious in patients undergoing surgical operations.
This involves administration of two types of drugs, in this
case atracurium for muscle relaxation and isoflurane for
the depth of unconsciousness. Identification studies for
each of these drugs have been carried out by Weatherley
et al. [43] and Millard et al. 1441 for atracurium and
isoflurane, respectively. Identification of the interacting
dynamics has been carried out by Asbury, and is
described in Linkens et al. [45]. The resulting multiI70

where U , is the drug atracurium rate, (I2 is the drug


isoflurane rate, Yl is muscle relaxation (MR) response
(expressed as a % o f a l paralysis), Y2 is mean arterial
pressure change, A M A P .
In addition, the pharmacodynamic effect of atracurium,
which is described as the relationship between the drug
concentration in the body plasma and the dynamic effect
of the drug, is usually modelled by a Hill equation [46]

where E,, is the maximum effect of the drug possible,


X,,, is the concentration necessary to cause 50% effect
and a is a constant. The average values of a and X,, , are
given as 2.98 and 0.404, respectively [45].
4.1 Learning results

Results of comparisons between different genetic adaptive


plans are shown as profiles of the learning process in
each C A . The characteristics of the learning processes are
given by the mean fitness of the population, the fitness of
the best individual in the population and the all-time-best
fitness in each generation. In elitist techniques, the best
individual is also the all-time-best, since the best individuals are always retained. The fitness of an individual is
indicated as the value returned by the error criterion
which is minimised, that is the integral of absolute
control rate and error (IACRE). Each reproduction
method was evaluated with each of the fitness methods.
The results show the mean performance profiles of 10
runs of the GAS, each of which was initialised with a different random population. The parameters of the genetic
adaptive plans (population size, reproduction, crossover,
mutation and fitness scaling) were the same in all experiments.
Fig. 9 shows the fitness profiles for the generational
replacement (GR) method using fitness rescaling by the
three methods, functional normalisation (NORM),
ranking (RANK) and windowing (WIND). It is observed
that this method does not significantly improve the population fitness, after the initial few generations. In the
I E E Proc.-Control Theory Appl., Vol. 142, No. 3 , M a y 1995

usual application of 'generational replacement' the alltime-best individual may not exist in all generations since
entire populations are replaced in each sample step. This
is evidenced by the fact that the best individual in the
latter generations is worse than the all-time-best.
Without any means of deterministically selecting or

120

01

The generational gap (GG) reproduction technique


replaces a proportion of the population with new offspring. When the gap is small the method behaves similarly to the steady-state technique, especially if the
individuals that are replaced are the least fit in the population. When the gap is large it is similar to generational

'_r(

1201
I

1 @@OO
number of evaluations

200@0

Fig. 9
Learning profile of G R method
0 NORM
RANK
WIND
~1 mean fitness
h best fitness

01

retaining the best individuals in the population they are


likely to be lost without being exploited to pass on their
characteristics to new offspring.The task of improving the
population then relies on the fitness technique to create
competition between the fitter individuals. Fitness com;
petition, due to sampling errors, does not guarantee that
the best individuals are selected for recombination to
produce better trials.
The steady-state (SS) method of reproduction retains
the best individuals in the population and, in many cases,
introduces only a pair of new individuals into the next
generation of population. Competition for survival with
this technique, thus, is less stringent since individuals not
selected for recombination are likely to survive anyway,
except for the least fit. This, as is shown in Fig. 10, results
in a progressive improvement in the population fitness. It
is also shown that the method is less sensitive to the
fitness method used, and hence achieves good results
with all three fitness methods.
I E E Proc.-Control Theory Appl., Vol. 142, No. 3, M a y 1995

1@00@
number of evaluations

?@Cl 30

Fig. 10

Learning profile of SS method

0 NORM
A RANK
+ WIND
a mean fitness
h best fitness
Best values far b : 0 27.31; A 27.45;

+ 27.80

replacement. The generational gap in this study was set


at 20%. It is shown in Fig. 11 that the behaviour of the
C A has the characteristics of the generational replacement method, although a significant improvement in
fitness is obtained.
The selective breeding (SB) method was introduced in
this research as a technique which combines rapid generation of trials (obtained with generational replacement
methods) and the caution of the steady-state method. It is
designed to maximise both exploitation of fit individuals
and exploration of the search space by creating entire
populations of offspring while at the same time retaining
the fittest. As shown in Fig. 12 the characteristics of this
method are very similar to the steady-state method, and
this method achieves a marginally better final solution.
171

It can be concluded that the best CA reproduction


methods are the steady-state and selective breeding techniques, which achieved very similar results with all types
of fitness techniques. Comparison of the fitness techniques, however, does not offer conclusive merits of one

120

4 2 Performance results
The results of the learning process are the rule bases that
are used to control the multivariable anaesthesia model.
The rules obtained with the steady-state and the selective
breeding reproduction methods together with the ranking
I

120

I
1

01

20000

120

10000
number of evaluations

20000

Fig. 11

Fig. 12

10000 0
number of evaluations
b

Learning profile of S B method

Learning profile of GG method


0 NORM
A RANK
+ WIND
a mean fitness
b best fitness
&st values forb: 0 47.91; A 73.12; + 31.20

A RANK
+ WIND
a mean fitness
b best fitness
Best values for b : 0 27.25; A 26.84;

technique over the other. Although all fitness techniques


performed well with the steady-state and selective breeding reproduction methods, the generational methods
were more sensitive to the type of fitness technique used.
Ranking achieved the best result for generational replacement and selective breeding, but was the worst with the
generational gap method. Functional normalisation
achieved the best result with the steady-state method,
while windowing produced the best results with the generational gap method. It is worth noting that studies
which have investigated factors affecting performances of
C A have tended to favour either of the combinations.
Grefenstette [27] used generational gap and windowing
while Schaffer er al. [29] used ranking and the steady
state techniques.

and functional normalisation fitness methods were used.


This is because results obtained using the generational
replacement and generation gap methods were, generally,
unsatisfactory as demonstrated by the fitness profiles
shown in Figs. 9 to 12.
A typical linguistic rule-base is shown in Table 1.
Three such rule-bases were derived: one for control of
MAP, the second for direct control of MR and the other
for compensation of the interactions between the two
processes. Examination of the table suggests some of the
generated rules to be intuitively inappropriate. A more
correct picture, however, is obtained when the linguistic
labels are taken together with their actual membership
functions. This is because in the learning process, individual rules develop their own membership functions. So,

172

0 NORM

+ 26.14

I E E Proc-Control Theory Appl., Vol. 142, No. 3, M a y 1995

seemingly different linguistic values may actually have


similar membership functions, and vice versa. On the
other hand, it can be noted that because the inferred
control actions are incremental, rather than absolute, the
sign of the increment is more significant than its actual
magnitude. This is because even small increments
(decrements) can soon result in a large (small) control
action.

0.6

Lo[

Table 1 : Rules for control of M R


Change-in-error

Error

PB
PM
PS
PO
NO
NS
NM
NB

Another way of looking at the rules is a control decisions' table that shows inferred actions for different
process states (error and change-in-error).An example of

NB

NM

NS

NO

PO

PB
PO
PO
PS
NS
NS

PO
PS
PS
PM
NS
NB
NO
NM

PB
PO
PS
PO
NS
NM
NB
NB

PO
PB
PS
PM
NS
NS
NS
NS

PS PB
PO PO
PS PB
PO PS
NS NB
NS NS
N M NS
NO NO

NS
NO

PS

PM

PB

PO
PB
PS
PM
NB
NS
NO
NS

PO
PB
PM
PB
NO
NB
NO
NB

~_______

h
:
50 100 150 200
0

0
'

50

minutes

100 150
minu,tei

200

Fig. 14
Multivariable response, from selective breeding and fitness
normalisation methods
MAP response
band b MR responses. showing interactions from MAP
n and a'

10

0.8

0.6

d 40

I?
'h

6 0.4

20

E
0.2

'0

Fig. 15
methods

OCE

-E

100

150

200

Z 0 L L L - L 0
50
100

minutes
b

50

Fig. 13 Rules surface, showing location of dominant rules, for M R


(top) and M A P

150

200

minutes
b'

Multivariable response, from selective breeding and ranking

o and a'

MAP response
band b' MR responm, showing interactions from MAP

Table 2: Control table for M R

CE

E
-10

-8

-6

9.8
8.3
5.7
5.1
-1.6
-0.7
-2.0
-4.8
-3.9
-3.5
-3.5

9.8
8.6
6.6
5.1
-1.6
-0.8
-2.5
-4.9
-3.9
-3.5
-3.5

9.4
8.6
7.2
4.4
-1.4
-0.7
-1.9
-5.0
-4.3
-3.0
-3.0

-4

-2

7.9
7.0
5.2
3.5
1.9
-0.0
-0.0
-2.5
-3.6
-3.3
-3.3

5.1
5.2
4.9
2.0
1.1
0.1
1.1
-1.3
-3.5
-4.0
-3.9

5.5
5.2
5.1
1.0
1.2
0.7
0.8
-3.0
-2.9
-2.8
-3.5

4.5
4.5
4.0
1.9
1.5
-1.2
-2.9
-0.5
0.9

10

~~

10
8
6
4
2
0
-2
-4
-6
-8
-10

I E E Proc.-Control Theory Appl., Vol. 142, No. 3, M a y 1995

6.1
6.1
4.6
-1.0
-0.6
-0.4
-1.3
-3.9
-3.5
-3.1
-3.1

6.1
6.2
4.6
3.2
1.2
-0.6
-2.9
-4.2
-4.7
-3.4
-3.1

-0.8
-3.0

2.9
1.9
2.9
2.2
2.7
2.5
1.3
1.3
1.5
1.4
0.2
2.6
0.7
2.7
1.8
1.7
1.8
1.8
-2.7 -2.7
-4.7 -4.7

173

control decision tables derived from the linguistic rulebases is given in Table 2. This table is obtained from the
rule-base by scanning the input space and inferring an
input for each point. The table is similar to 'look-up
tables' sometimes used for direct fuzzy control [47,48].

Finally, results of the rules obtained in the learning


process were evaluated on the domain task, being the
simulated control of the multivariable anaesthesia model.
On the one hand, it can be seen that the selective breeding method (Figs. 14 and 15) achieved better results than

+-a'4 0

L
E

a'

100

60
2

rr' LO
I

20

minutes

minutes
b

b'

normnl.

Fig,
response,from steady
isation methods
(1 and a' MAP response
b and b MR responses,showing interactions from MAP

=t

"

'

"

'

rrinuies

b
b'
Fig. 18 Multivariable response, from a generational gap GA
a and a' MAP response
band b MR responses, showing interactions from MAP
For b : ITAE = 6525.92; IAECE = 90.60

the steady state method (Figs. 16 and 17), as indicated by


the ITAE and IACRE figures of merits. On the other
hand, use of the ranking fitness method (Figs. 15 and 17)
is observed to achieved better results than the functional
normalisation method. The combination of the selective
breeding and ranking techniques achieved the best result
(Fig. 15). As expected, control rules obtained from the
generational replacement schemes performed significantly
worse, as demonstrated in Fig. 18.

40

01

minutes

a'

Conclusion

The objective of this paper was to investigate the use of


genetic algorithms as a tool for the design of fuzzy controllers. To do this it was first necessary to derive a GA
scheme that produces an acceptable set of control rules,
in an acceptable time period. This is because learning
with GAS requires many iterations and consequently, a
ULU
large amount of processing time. For a problem of the
'0
50 100 150 200
'0
50
100 150 200
complexity considered in this research this could mean
minutes
minutes
b
b'
several hundred processor hours. In this regard the learning experiments in this research were carried out on a
Fig. 17 Multivariable response, from steady state and ranking
methods
parallel processing transputer platform.
a and 0' MAP response
Common reproduction and fitness techniques were
band h MR responses, showing interactions from MAP
studied in this work, and in addition new reproduction
For h : ITAE = 3377.73; IACRE = 16.67
and fitness techniques were investigated, both aimed at
improving on the usual methods. Experimental results
A further presentation of the results is a graphical
showed that only the steady-state reproduction method,
display showing the highest 'firing level' of all points in
among the usual methods, achieved good results in terms
the input space to the rules. This representation indicates
of the performance of the final rules. It was concluded
the number and location of the rules, since the highest
that the best genetic plans for this type of problem were
degree of firing is found at the position of the peak of the
the steady state and the selective breeding techniques,
rule fuzzy sets. This is portrayed in Fig. 13 for both MAP
and the best fitness methods were ranking and functional
and M R .
normalisation. However, it ought to be stated that this

'

174

I E E Proc.-Control Theory Appl., Vol. 142. No. 3, M a y I995

does not demonstrate that any reproduction method or


fitness technique is better than the other, except under
the specific circumstances investigated.
This research has investigated two different
approaches for applying CA in fuzzy controller design.
The second approach is discussed in part I1 of this paper.
This is an online approach in which the fuzzy control
rules are acquired in real time. The two approaches
represent two extremes of learning with CA. In the offline
method presented in this paper, a model of the process is
required, but more robust rules are derived. In the online
case, long periods of experimentation may be required
when no a priori knowledge is available. The method
may also not be applicable to all types of processes.
However, it is an effective method of rule acquisition,
when a certain amount of a priori knowledge is available
or can be assumed. The combination of these two
approaches presents an attractive technique for design of
adaptive fuzzy controllers, the offline method being used
to generate an initial rule-base which can then be modified online.
6

References

I ASSILIAN, S., and MAMDANI, E.H.: An experiment in linguistic


synthesis with a fuzzy log~ccontroller, lnt. J. Man Machine Studies,
1977,7, pp. 1-13
2 KING, P.J., and MAMDANI, E.H.: The application of fuzzy
control systems to industrial processes, Autornatica, 1977, 13, pp.
235-242
3 MAMDANI, E.H.: Appkations of fuzzy algorithms for control of
a simple dynamic process, Proc. I E E , 1974, 121, pp. 1585-1588
4 DAVIS, L.: A handbook of genetic algorithms (Van Nostrand
Reinhold, New York, 1990)
5 GOLDBERG, D.E.: Genetic algorithms in search, optimization
and machine learning (Addison-Wesley, 1989)
6 KARR, C.L., FREEMAN, L.M., and MEREDITH, D.L.: Improved
fuzzy process control of spacecraft autonomous rendezvous using a
genetic algorithm, SPIE lntelligenr Control & Adaptive Systems,
1989,11%, pp. 274-288
7 KARR, C.L.: Genetic algorithm based fuzzy control of spacecraft
autonomous rendezvous, in BELEW, R., and BOOKER, L. (Eds.):
Proceedings of the fourth international conference on Genetic algorithms (Morgan Kauffman Publishers, Cambridge, MA, 1991). pp.
43-51
8 COX, E:.: Fuzzy fundamentals, l E E E Spectrum, 1992, 29, (lo), pp.
58-61
9 DALEY, S., and GILL, K.S.: A design study of a self organizing
fuzzy logic controller, Proc. Inst. Mech. Eng., 1986, 200, (CI), pp.
59.~69
I O LINKENS, D.A., and ABBOD, M.F.: Self-organizing fuzzy logic
control for real-time processes. Proceedings of the IEE international conference Control, Edinburgh, UK, 1991
1 I PROCYK, T.J., and MAMDANI, E.H.: A linguistic self organizing
process controller. Automatica, 1979, 15, pp. 15-30
12 ABBOD, M.F.: Supervisory intelligent control for industrial and
medical systems. PhD dissertation, Department Automatic Control
& Systems. Engineering, University of Shefield, UK, 1992
13 PETERS, I., BECK, K.. and CAMPASANO, R.: Fuzzy logic with
dynamic rule-set. Proceedings of the IEEE symposium on Intel.
control, Glasgow, UK, 1992, pp. 216-219
14 MIZUMUTO, M.: Fuzzy controls under various reasoning
methods. Inform. Sci.. 1988. 45, pp. 129-151
15 LINKENS, D.A., and ABBOD, M.F.: Self organizing fuzzy logic
control and the selection of its scaling factors, Trans. Inst. Meas. &
Control, 1992, 14, (3). pp. 114-125
16 CELLIER. E.F., and ALVAREZ, F.M.: Systematic design of fuzzy
controllers using inductive reasoning. Proceedings 1992 IEEE international symposium on intelligence control, Glasgow, 1992, pp.
198-203
17 LINKENS, D.A., and SHIEH, J.: Self-organizing fuzzy modelling
for nonlinear system control. IEEE international symposium on
intelligence control, Glasgow, UK, 1992, pp. 210-215
18 MICHALSKI, R.S.: A theory and methodology of inductive
reasoning, Artificial Intelligence, 1983, 20, pp. I1 1-161
19 MITCHELL, T.M., MAHADEVAN. S., and STEINBERG, L.:
LEAP: a learning apprentice for VLSI design. Proceedings of the
IEE Prm-Control Theory Appl., Vol. 142, No. 3, M a y 1995

9th international joint conference on Artificial intelligence (MorganKauffman Publishers, 1985). pp. 573-580
20 KARR, C.L.: Design of an adaptive fuzzy logic controller using a
genetic algorithm, in SCHAFFER, I. (FA.): Proceedings of the third
international conference on Genetic algorithms (Morgan Kauffman
Publishers, Cambridge, MA, 1989), pp. 450-457
21 THRIFT, P.: Fuzzy logic synthesis with genetic algorithms, in
BELEW, R., and BOOKER, L. (Eds.): Proceedings of the fourth
international conference on Genetic algorithms (Morgan Kauffman
Publishers, Cambridge, MA, 19911, pp. 509-513
22 HOLLAND, J.H.: &Geneticalgorithms and the optimal allocation of
trials, SIAM J. Compur., 1973,2, (2), pp. 89-104
23 HOLLAND, J.H.: Adaptation in natural and articifical systems
(Addison-Wesley, 1975)
24 ANTONISSE, H.J., and KELLER, K.S.: Genetic operators for high
level knowledge representation, in GREFENSTETTE. J. (Ed.): Proceedings of the second international conference on Genetic algorithms (Lawrence Erlbaum Publishers, Hillsdale, NI. 1987). pp.
69-76
25 CARUANA, R.A., and SCHAFFER, J.D.: Representation and
hidden bias: Gray vs. Binary coding for genetic algorithms, in
LAIRD, J. (Ed.): Proceedings of the fifth international conference on
Machine learning (Morgan-Kauffman Publishers, Camhridge, MA.
1988), pp. 153-161
26 SCHRAUDOLPH, N.N., and BELEW, R.K.: Dynamic parameters
encoding for genetic algorithms, Machine Learning, 9 (Kluwer Academic, 1992), pp. 9-21
27 GREFENSTETTE, J.J.: Optimization of control parameters for
genetic algorithms, IEEE Trans. Syst. Man & Cyhern., 1986,
SMC-16, pp. 122-128
28 DAVIS, L.: Adapting operator probabilities in genetic algorithms,
in SCHAFFER, J. (Ed.): Proceedings of the third international conference on Generic algorithms (Morgan Kauffman Publishers, Cambridge, MA, 1989). pp. 61-69
29 SCHAFFER, J.D., CARUANA, R.A., ESHELMAN, L.J., and DAS,
R.: A study of control parameters affecting online performance of a
genetic algorithm for function optimization, in SCHAFFER, J.
(Ed.): Proceedings of the third international conference on Genetic
algorithms (Morgan Kauffman Publishers, Cambridge. MA, 1989),
pp. 51-60
30 DE JONG. K.A.: Are genetic algorithms function optunizers?, in
MANNER, R., and MANDERICK, B. (Eds.): Proceedings of the
second workshop on Parallel problem soloing Jrom nature (SpringerVerlag, Berlin, 1992). pp. 2-13
31 GOLDBERG, D.E.: Sizing populations for serial and parallel
genetic algorithms, in SCHAFFER, J. (Ed.): Proceedings of the
third international conference on Genetic algorithms (Morgan KauITman Publishers, Cambridge, MA, 1989),pp. 71-79
32 DE JONG. K.A., and SPEARS, W.M.: An analysis of the interacting roles of population size and crossover in genetic algorithms,
in SCHWEFEL, H.-P., and MANNER, R. (Eds.): Proceedings of
the first workshop on Parallel problem solving Jrom nature,
Dortmund, West Germany (Springer-Verlag, Berlin. 1990), pp,
39-47
33 ODETAYO, M.: Optimal population sue for genetic algorithms
an investigation. IEE digest of the colloquium on genetic algorithms in control systems engineering (London, May 1993)
34 ROBERTSON, G.G.: Population size in classifier systems, in
LAIRD, J. (Ed.): Proceedings of the fifth international conference on
Machine learning (Morgan-Kauffman Publishers, Cambridge, MA,
1988), pp. 142-152
35 SYSWERDA, G.: Uniform crossover in genetic algorithms, in
SCHAFFER, I. (Ed.): Proceedings of the third international conference on Genetic algorithms (Morgan Kauffman Publishers, Cambridge, MA, 1989). pp. 2-9
36 SYSWERDA, G.: A study of reproduction in generational and
steady state genetic algorithms, in RAWLINS, G.J.E. (Ed.): Proceedings of a workshop on foundations of Generic algorithm and
classifier systems (Lawrence Erlbaum Publishers, Hillsdale, NJ,
1991), pp. 94-101
37 SCHAFFER, J.D., and MORISHIMA, A.: An adaptive crossover
distribution mechanism for genetic algorithms, in GREFENSTETTE, 1. (Ed.): Proceedings of the second international conference on Genetic algorithms (Lawrence Erlbaum Publishers,
Hillsdale, NJ, 1987), pp. 36-40
38 MUHLENBEIN, H.: How genetic algorithms really work. I: Mutation and hill-climbing, in GREFENSTETTE, 1. (Ed.): Proceedings
of the second international conference on Genetic algorithm
(Lawrence Erlhaum Publishers, Hillsdale, NI, 1992). pp. 15-25
39 BACK, T.: Optimization of mutation rates in genetic search, in
FORREST, S. (Ed.): Proceedings of the fifth international conference on Genetic algorithms (Morgan Kauffman Publishers, C a n bridge, MA, 1993).pp. 2-8
175

40 BAKER, J.: Adaptive selection methods for genetic algorithms.

44 MILLARD, R.K., MONK, C.R., WOODCOCK, T.C., and PRYS-

Proceedings of an international conference on Genetic algorithms


(Lawrence Erlbaum Publishers, Hillsdale, N.J., 1985), pp. 100-111
41 MANDERICK, B., and SPIESSENS, P.: Fine grained parallel
genetic algorithms, in SCHAFFER, J. (Ed.): Proceedings of the
third international conference on Genetic algorithms (Morgan Kaullman Publishers, Cambridge, MA, 1989), pp. 428-433
42 TANESSE, R.: Distributed genetic algorithms, in SCHAFFER, J.
(ed.): Proceedings of the third international conference on Genetic
algorithms (Morgan KaulTman Publishers, Cambridge, MA, 1989),
pp. 434-439
43 WEATHERLEY, B.C., WILLIAMS, S.G., and NEILL, E.A.M.:
Pharmacokinetics. pharmacodynamics and dose-response relationships of atracuriurn administered i.v., Brit. J. Anae., 1983, 55, pp.
39s-45s

ROBERTS, C.: Controlled hypotension during ENT surgery using


self-tuners, Biomed. Meas. Infor. Contr., 1988, 2, pp. 59-72
LINKENS, D.A., MAHFOUF, M., and ASBURY, A.J.: Multivariable generalized predictive control for anaesthesia. Proceedings
of the European control conference, Grenoble, France, 1991
GODFREY, K.R.: Pharmacokinetics, Trans. Imtiture Meas. &
Control, 1982,4, pp. 213-223
CHUNYU, H.T., TOGUCHI, K., SHENOI, S., and FAN, L.T.: A
technique for design and implementation of fuzzy logic controllers.
Proceedings of the 1989 American control conference, pp. 27542755
KWOK, D.P., et al.: Linguistic PID controller. Proceedings of
IFAC 11th congress, Tallin, USSR; pp. 192-197

176

45

46
47

48

IEE Proc.-Control Theory Appl., Vol. 142, No. 3 , M a y 1995

You might also like