Int. J. Elect. Enging. Educ., Vol. 36, pp. 139153. Manchester U.P., 1999.

Printed in Great Britain

TEACHING GENETIC ALGORITHM USING MATLAB

Y. J. CAO1 and Q. H. W U2
1Intelligent Computer Systems Centre, University of the West of England,
Bristol, UK
2Department of Electrical Engineering and Electronics, University of L iverpool,
L iverpool, UK

ABSTRACT In this paper, an attractive approach for teaching genetic algorithm (GA) is presented. This approach is based primarily on using MATLAB in implementing the genetic operators:
crossover, mutation and selection. A detailed illustrative example is presented to demonstrate that
GA is capable of finding global or near-global optimum solutions of multi-modal functions. An
application of GA in designing a robust controller for uncertain control systems is also given to
show its potential in designing engineering intelligent systems.

KEYWORDS

control systems; genetic algorithm; MATLAB

1 INTRODUCTION
Genetic Algorithm (GA) is a major topic in a neural and evolutionary computing under-graduate (advanced level)/postgraduate course. The course plays an
important role in designing intelligent control systems, which is now very
attractive and stimulating to students in Electrical Engineering Departments.
The usual way to teach genetic algorithm is by the use of PASCAL, C and
C++ languages. In this paper, we present an attractive and easy way for
teaching such a topic using MATLAB1. It has been found that this approach
is quite acceptable to the students. Not only can they access the software
package conveniently, but also MATLAB provides many toolboxes to support
an interactive environment for modelling and simulating a wide variety of
dynamic systems including linear, nonlinear discrete-time, continuous-time and
hybrid systems. Together with SIMULINK, it also provides a graphical user
interface (GUI) for building models as block diagrams, using click-and-drag
mouse operations.
The rest of the paper is organised as follows. Section 2 introduces a simple
version of genetic algorithm, canonical genetic algorithm (CGA), which was
originally developed by Holland2. Section 3 describes the implementation
details of the genetic operators: crossover, mutation and selection. Section 4
demonstrates the performance of the CGA with a multi-modal function.
Section 5 provides an application of the developed genetic operators in designing a robust controller for uncertain plants, and section 6 gives a conclusion.
139

140

2 CANONICAL GENETIC ALGORITHM

Genetic algorithms are adaptive algorithms for finding the global optimum solution for an optimization problem. The canonical genetic algorithm developed by
Holland is characterised by binary representation of individual solutions, simple
problem-independent crossover and mutation operators, and a proportional selection rule. To understand these concepts, consider the standard procedure of the
CGA outlined in Fig. 1. The population members are strings or chromosomes,
which as originally conceived are binary representations of solution vectors. CGA
undertakes to select subsets (usually pairs) of solutions from a population, called
parents, to combine them to produce new solutions called children (or offspring).
Rules of combination to yield children are based on the genetic notion of crossover,
which consists of interchanging solution values of particular variables, together
with occasional operations such as random value changes (called mutations).
Children produced by the mating of parents, and that pass a survivability test,
are then available to be chosen as parents of the next generation. The choice of
parents to be matched in each generation is based on a biased random sampling
scheme, which in some (nonstandard) cases is carried out in parallel over separate
subpopulations whose best members are periodically exchanged or shared. The
key parts in Fig. 1 are described in detail as follows.

2.1 Initialisation
In the initialisation, the first thing to do is to decide the coding structure. Coding
for a solution, termed a chromosome in GA literature, is usually described as a
string of symbols from {0, 1}. These components of the chromosome are then
labeled as genes. The number of bits that must be used to describe the parameters
is problem dependent. Let each solution in the population of m such solutions x ,
i
i=1, 2, ..., m, be a string of symbols {0, 1} of length l. Typically, the initial population of m solutions is selected completely at random, with each bit of each
solution having a 50% chance of taking the value 0.

FIG. 1

Standard procedure of a canonical genetic algorithm.

141

2.2 Selection
CGA uses proportional selection, the population of the next generation is
determined by n independent random experiments; the probability that individual x is selected from the tuple (x , x , ..., x ) to be a member of the next
i
1 2
m
generation at each experiment is given by
f (x )
i >0.
P{x is selected}=
(1)
i
m
f (x )
j
j=1
This process is also called roulette wheel parent selection and may be viewed
as a roulette wheel where each member of the population is represented by a
slice that is directly proportional to the members fitness. A selection step is
then a spin of the wheel, which in the long run tends to eliminate the least fit
population members.
2.3 Crossover
Crossover is an important random operator in CGA and the function of the
crossover operator is to generate new or child chromosomes from two parent
chromosomes by combining the information extracted from the parents. The
method of crossover used in CGA is the one-point crossover as shown in
Fig. 2(a). By this method, for a chromosome of a length l, a random number
c between 1 and l is first generated. The first child chromosome is formed by
appending the last lc elements of the first parent chromosome to the first c
elements of the second parent chromosome. The second child chromosome is
formed by appending the last lc elements of the second parent chromosome
to the first c elements of the first parent chromosome. Typically, the probability
for crossover ranges from 0.6 to 0.95.
2.4 Mutation
Mutation is another important component in CGA, though it is usually conceived as a background operator. It operates independently on each individual

FIG. 2 One-point crossover and mutation operators.

142

by probabilistically perturbing each bit string. A usual way to mutate used in

CGA is to generate a random number v between 1 and l and then make a
random change in the vth element of the string with probability p (0, 1),
m
which is shown in Fig. 2( b). Typically, the probability for bit mutation ranges
from 0.001 to 0.01.
Hollands early designs for CGA were simple, but were reported to be effective
in solving a number of problems considered to be difficult at the time. The field
of genetic algorithms has since evolved, chiefly as a result of innovations in the
1980s, to incorporate more elaborate designs aimed at solving problems in a wide
range of practical settings. Characteristically, GAs are distinguished in the following features: (1) they work with a coding of the parameter sets instead of the
parameters themselves; (2) they search with a population of points, not a single
point; (3) they use the objective function information directly, rather than the
derivatives or other auxiliary knowledge, to find maxima; (4) they process information using probabilistic transition rules, rather than deterministic rules. These
features make GAs robust to computation, readily implemented with parallel
processing and powerful for global optimisation.
3 IMPLEMENTATION OF CGA USING MATLAB
This section describes the procedure of implementing the canonical genetic
algorithm using MATLAB. For simplicity, we assume that the task is to achieve
the maximum of a one-variable multi-modal function, f (x)>0, for x belonging
to the domain [a, b]. The implementation steps are as follows.
3.1 Initialisation
For CGA, a binary representation is needed to describe each individual in the
population of interest. Each individual is made up of a sequence of binary bits
(0 and 1). Let stringlength and popsize denote the length of the binary sequence
and the number of individuals involved in the population. Each individual uses
a string codification of the form shown in Fig. 3. Using MATLAB, the whole
data structure of the population is implemented by a matrix of size
popsize(stringlength+2):

pop=

binary string 1

binary string 2

2
e

f (x )
1
f (x )
2
e

x
popsize

f (x
)
popsize

binary string popsize

FIG. 3 Binary string representation for the optimisation of a one-variable

function.

143

The first stringlength column contains the bits which characterise the binary
codification of the real variable x. The strings are randomly generated, but a
test must be made to ensure that the corresponding values belong to the
function domain. The crossover and mutation operators will be applied on this
stringlength-bit sub-string. The (stringlength+1)-th and (stringlength+2)-th
columns contain the real x value, used as auxiliary information in order to
check the algorithms evolution, and the corresponding f (x), which is assumed
to be the fitness value of the string in this case. Then the initialisation process
can be completed using the following code:
function [pop]=initialise(popsize, stringlength, fun);
pop=round(rand(popsize, stringlength+2));
pop(:, stringlength+1)=sum(2.9(size(pop(:,1:stringlength),2)1:1:0).
*pop(:,1:stringlength))*(ba)/(2.9stringlength1)+a;
pop(;, stringlength+2)= fun(pop(;, stringlength+1));
end

In the above routine, we first generate the binary bits randomly, and then
replace the (stringlength+1)-th and (stringlength+2)-th columns with real x
values and objective function values, where fun is the objective function, usually
denoted by a .m file.
3.2 Crossover
Crossover takes two individuals parent1, parent2, and produces two new individuals child1, child2. Let pc be the probability of crossover, then the crossover
operator can be implemented as follows:
function [child1, child2]=crossover(parent1, parent2, pc);
if (rand<pc)
cpoint=round(rand*(stringlength2))+1;
child1=[ parent1(:,1:cpoint) parent2(:,cpoint1+1:stringlength)];
child2=[ parent2(:,1:cpoint) parent1(:,cpoint1+1:stringlength)];
child1(:, stringlength+1)=sum(2.9(size(child1(:,1:stringlength),2)1:1:0).
*child1(:,:stringlength))*(ba)/(2.9stringlength1)+a;
child2(:, stringlength+1)=sum(2.9(size(child2(:,1:stringlength),2)1:1:0).
*child2(:,1:stringlength))*(ba)/(2.9stringlength1)+a;
child1(:, stringlength+2)= fun(child1(:, stringlength+1));
child2(:, stringlength+2)= fun(child2(:, stringlength+1));
else
child1=parent1;
child2=parent2;
end
end

At the top of the crossover routine, we determine whether we are gong to

perform crossover on the current pair of parent chromosomes. Specifically, we
generate a random number and compare it with the probability parameter pc.
If the random number is less than pc, a crossover operation is performed;
otherwise, no crossover is performed and the parent individuals are returned.
If a crossover operation is called for, a crossing point cpoint is selected between
1 and stringlength. The crossing point cpoint is selected in the function round,

144

which returns a pseudorandom integer between specified lower and upper

limits ( between 1 and stringlength 1). Finally, the partial exchange of crossover is carried out and the real values and fitness of the new individuals child1,
child2, are computed.
3.3 Mutation
Mutation alters one individual, parent, to produce a single new individual,
child. Let pm be the probability of mutation, then as in the crossover routine,
we first determine whether we are going to perform mutation on the current
pair of parent chromosomes. If a mutation operation is called for, we select a
mutating point mpoint, and then change a true to a false (1 to 0) or vice versa.
The real value and fitness of the new individual child are then computed:
function [child]=mutation(parent, pm):
if (rand<pm)
mpoint=round(rand*(stringlength1))+1;
child=parent;
child[mpoint]=abs(parent[mpoint]1);
child(:, stringlength+1)=sum(2.9 (size(child(:,1:stringlength),2)1:1:0).
*child(:,1:stringlength))*(ba)/ (2.9stringlength1)+a;
child(:, stringlength+2)= fun(child(:, stringlength+1));
else
child=parent;
end
end

3.4 Selection
The selection operator determines which of the individuals will survive and
continue in the next generation. The selection operator implemented here is
roulette wheel selection and this is perhaps the simplest way to implement
selection. We first calculate the probabilities of each individual being selected,
based on equation (1). Then the partial sum of the probabilities is accumulated
in the vector prob. We also generate a vector rns containing normalised random
numbers, by comparing the elements of the two vectors rns and prob, we decide
the individuals which will take part in the new population:
function [newpop]=roulette(oldpop);
totalfit=sum(oldpop(:,stringlength+2));
prob=oldpop(:,stringlength+2) / totalfit;
prob=cumsum(prob);
rns=sort(rand(popsize,1));
fitin=1; newin=1;
while newin<=popsize
if (rns(newin)<prob(fitin))
newpop(newin,:)=oldpop(fitin,:);
newin=newin+1;
else
fitin= fitin+1;
end
end

145

3.5 Accessing on-line information

We have shown how the three main pieces of CGA may be easily coded and
easily understood. After implementing the key genetic operators, implementation of the main program is straightforward. At the moment, the World Wide
Web and the Internet provide some quality MATLAB source codes for different
kinds of genetic algorithms. Chris Houck, Jeff Joines, and Mike Kay at North
Carolina State University have developed two versions of the genetic algorithm
for function optimization using both MATLAB 4 and MATLAB 5. The WWW
site is: http://www.ie.ncsu.edu/mirage. Andrew F. Potvin has also implemented
a simple version of genetic algorithm which is available on the WWW site:
http://unix.hensa.ac.uk/ftp/mirrors/matlab/contrib/v4/optim/genetic/.
It should be pointed out that although the above Toolboxes are very useful
in solving many engineering problems, they are sometimes too comprehensive
for the students. In many situations, the students need to modify the genetic
operators to meet their own problem-dependent requirements. However, they
found it very difficult to change some parts of the above GA Toolboxes, because
the whole package is heavily interconnected. With the source codes provided
in this section, the students found it quite easy to modify the genetic operators
and invent some new genetic operators. This is helpful in promoting the
students creativity, which is necessary and important in the new teaching
process.
4 AN ILLUSTRATIVE EXAMPLE
This section demonstrates that the CGA is able to find the global or nearglobal maximum of multi-modal functions. As an example, the following onevariable function is considered:
f (x)=

sin2(10x)
1+x

(2)

for x belonging to the domain [0, 1]. The characteristic of the function is
plotted in Fig. 4(a). As shown in the figure, the function has three maxima,
with a global maximum achieved at x=0.1527.
Several trials have been performed changing the fundamental parameter
values. Below, a relevant example, with popsize=30, stringlength=20, pc=
0.95 and pm=0.05 is reported. Fig. 4(a) shows the initial population for the
optimisation of a one-variable optimisation: the 1 represent the individuals of
the population, randomly generated in the function domain. Fig. 4( b) shows
the population after 50 generations of the CGA, and Fig. 4(c) gives the population after 100 generations of the CGA.
As can be observed from Fig. 4(a), at the first step, the population (represented by 1 in the figures) is randomly distributed in the whole function
domain. During the evolution of the algorithm, Fig. 4( b), the individuals of the
population tend to concentrate at the peaks of the function and at convergence,
Fig. 4(c), most of the individuals are clustered near the desired maximum value
of the function.

146

FIG. 4

Performance of the CGA with a multi-modal function.

5 AN APPLICATION EXAMPLE
In this section, we provide an application example of the developed GA in
designing robust controller for uncertain control systems. Robust control under
parametric uncertainty is an important research area with many practical
applications. When the system has a general nonlinear uncertainty structure,
the usual approach is to overbound it by an interval dynamical system. The

147

FIG. 4 (Continued.)
robust controller design can be transformed into solving an eigenvalue optimisation problem, which involves optimising the maximum real part of eigenvalues
of the characteristic polynomial for the uncertain plant. This kind of problem
is a typical non-smooth and nonconvex optimisation problem, and may have
several local optima. We demonstrate that GA is capable of solving this kind
of problem. A practical flywheelshaftflywheel system, including uncertainty
in the length of the shaft and shaft damping, is employed to carry out the
numerical simulation.
5.1 Uncertain control systems
Let us consider the following single-input single-output (SISO) linear systems:
dny
dn1y
a (q)
+a (q)
++a (q)y
n
n1
0
dtn
dtn1
dm1u
dmu
+b
(q)
++b (q)u
=b (q)
m1
0
m
dtm
dtm1

(3)

where n is the order of the model, y is the output variable, u is the input
variable, and q={q , q , ..., q } is the design parameter vector. Furthermore,
1 2
k
we acknowledge uncertainties in the design parameters by noting the design
parameter vector q lies in a box Q=[q , q+][q , q+][q , q+]. The
1 1
2 2
k k
transfer function of (3) is defined by
G(s, q)=

Y (s, q) W m b (s, q)si N(s, q)

= i=0 i
.
=
U(s, q) W n a (s, q)sj D(s, q)
j=0 j

(4)

148

For convenience, we define the entire family of uncertain plants as

G(s)={G(s, q) | qQ}.

(5)

We also consider a feedback controller, F(s, p)=F (s, p)/F (s, p), where the
1
2
numerator and denominator of F are polynomials in s, and p={p , p , ..., p }
1 2
l
is the parameter vector lying in a box
P=[ p , p+][ p , p+][ p , p+].
1 1
2 2
l l
The task of robust controller design considered in this section is to determine
the optimal value of p so that the feedback control system is stable for all
parameters qQ.
For the feedback system shown in Fig. 5, the characteristic polynomial is:
H(s)=F (s, p)D(s, q)+F (s, p)N(s, q)
2
1
and this polynomial varies over the corresponding uncertainty set

(6)

H(s, q, p)={F (s, p)D(s, q)+F (s, p)N(s, q) | qQ, pP}

2
1
Using Fig. 5, we cite the transfer functions pertinent to control

(7)

T y(s, q, p)=

Y (s)
F(s, p)G(s, q)
=
.
R(s) 1+F(s, p)G(s, q)

(8)

The robust controller design can be considered as determining the parameter

value pP so that the feedback control system is stable for all qQ. However,
the presence of the k-dimensional uncertainty vector, q, complicates the analysis
of the uncertain transfer functions such as (4) and equations such as (6).
Thorough robust analysis requires that the behaviour of every point in the
uncertainty space, i.e. all qQ, must be checked. As this is impossible, a
common technique involves creating a grid of points for each q [q, q+],
i
i i
i=1, ..., k, and analysing behaviour over this k-dimensional grid. This approach
is computationally intensive and inexact one can never be certain that a
finer grid might not change a robust stability prediction. In the next subsection,
we discuss solving this problem using GA optimisation approach.
5.2 Robust controller design via GA
As mentioned above, the main task of robust controller design is to determine
the optimal value of p so that the feedback control system is stable for all
parameters qQ. This can be transformed into solving eigenvalue optimisa-

FIG. 5 Uncertain feedback control plant.

149

tion problems. To be more precise, the optimisation problem may be stated

mathematically as
min g(p)
and
g(p)= max Re(l (H(s, q, p))).
(9)
i
pP
i,qQ
This is a typical minmax optimisation problem, and the following result is
straightforward4.
Proposition If ( p*, q*) is the global optimum of the problem (9), and the
corresponding objective value is less than 0, then the uncertain system (3) with
feedback controller F( p*, s) is stable for all qQ.
This proposition implies that, to design the robust controller for uncertain
system (3), an alternative approach is to develop a reliable approach to find
the global optima of problem (9). Recently, GA has been considered as viable
and promising techniques for global optimisation of complex functions. When
applying GA to the minmax problem (9), the fitness scaling technique in
Ref. [3] and the nesting technique in Ref. [5] are always used. The former
technique is to keep the fitness positive and avoid large deviation of fitness.
The latter technique involves an inner loop to maximise the Re(l (H(s, q, p)))
i
for fixed parameters p, and an outer loop to minimise g(p) over the set P.
Let us consider the practical shaftflywheel system illustrated in Fig. 6. The
input to the system is a torque applied to the first flywheel, and the output is
the angular velocity of the second flywheel. We model the shaftflywheel system
as two torsional inertias separated by a torsional spring and a torsional dashpot.
The transfer function for this system is
K
C
D s+
J J
J J
1 2
1 2
(10)
1
1
1
1
s3+C
s2+K
s
+
+
D J
J
J
J
1
2
1
2
where J and J are the inertia of the left flywheel and one-half of the shaft
1
2
inertia and the inertia of the right flywheel and one-half of the shaft inertia,
respectively. K is the torsional stiffness of the shaft, and C is the torsional
D
N(s, q)
G(s, q)=
=
D(s, q)

FIG. 6 Flywheelshaftflywheel plant.

150

damping of the shaft. We use standard relations to compute the torsional

inertias and stiffness, i.e.
J=

rpld4
32

(11)

Gpd4
32l

(12)

K=

where r is material density, l is the length of a circular shaft, d is the diameter,

and G is the shear modulus. Equation (11) also applies to the flywheels. We
assume proportional damping for the shaft, where C =C K.
D
0
We consider the case in which the uncertain design parameter vector, q,
consists of C , the proportional damping coefficient, the l , the shaft length.
0
s
This vector lies in the box
Q={q | C [0.0008, 0.0012],
l [0.9, 1.1]}
0
s
and the other parameters have the values: d =0.20, l =0.02, d =0.01,
w1
w1
s
d =0.15, and l =0.02.
w2
w2
We assume the controller to be designed having the following structure:
(s+p )(s2+p s+p )
3
4
5
(13)
s(s2+p s+p )
1
2
The developed GA is employed to solve the optimisation problem (9), involving
the transfer function G(s, q) in (10) and feedback controller F(s, p) in (13). The
population size and chromosome length in GA are chosen as 120 and 100,
respectively. The probabilities of crossover and mutation in GA are chosen as
F(s, p)=0.234

FIG. 7

Optimisation procedure for a robust controller with GA.

151

FIG. 8 Performance of an uncertain system with a robust controller.

0.95 and 0.05. Using the fitness scaling technique in Ref. [3] and the nesting
technique in Ref. [5], the optimal solution obtained using GA is: p =212.60,
1
p =11024.80, p =15.12, p =79.86, p =11026.84. The optimisation procedure
2
3
4
5
is shown in Fig. 7, where the best fitness in each population with respect to
generations is plotted. Fig. 8 and Fig. 9 show the Bode plots and the step

152

FIG. 9 Performance of an uncertain system with a robust controller.

responses of the feedback control system, when the uncertain parameters take
C =0.0008, l =1.0 and C =0.0012, l =1.1, respectively. Furthermore, the
0
s
0
s
uncertain plant (10) with the feedback controller (13) can be guaranteed stable
by Kharitonovs theorem.

153

6 CONCLUSION
In this paper, an attractive approach for teaching genetic algorithm has been
presented. This approach is based primarily on using MATLAB in
implementing the genetic operators: crossover, mutation, and selection. An
advantage of using such an approach is that the student becomes familiar with
some advanced features of MATLAB, and furthermore, with the availability of
other MATLAB Toolboxes such as The Control Systems Toolbox, Neural
Network Toolbox and Fuzzy Logic Toolbox, it is possible for the student to
develop genetic algorithm-based approaches to designing intelligent systems,
which could lead to his/her final year or MSc project.
REFERENCES
[1] MAT L AB 5, The Math Works Inc., Natick, MA (1997)
[2] Holland, J., Adaptation in Natural and Artificial Systems, University of Michigan Press, Ann
Arbor (1975)
[3] Goldberg, D. E., Genetic Algorithms in Search, Optimization and Machine L earning, AddisonWesley (1989)
[4] Borie, J. A., Modern Control Systems, Prentice-Hall International (1986)
[5] Sebald, A. V. and Schlenzig, J., Minimax design of neural net controller for highly uncertain
plants, IEEE T rans. Neural Networks, 5, pp. 7382 (1994)

ABSTRACTS FRENCH, GERMAN, SPANISH

Enseignement dalgorithme genetique par MATLAB
Dans cet article une approche attractive pour enseigner lalgorithme genetique (GA) est presentee.
Cette approche est basee sur lutilisation de MATLAB pour implementer les operateurs genetiques:
mutation, selection et croisement. Un exemple illustratif detaille est presente pour demontrer que
GA est capable de trouver un optimum global ou quasi-global pour des fonctions multi-modales.
Une application de GA dans la conception dun controleur robuste pour des syste`mes a` controle
incertain est egalement donnee pour montrer son potentiel dans la conception de syste`mes
intelligents en ingenierie.
Genetische Algorithmen mit MATLAB lehren
In diesem Bericht wird ein attraktives Vorgehen zum Lehren genetischer Algorithmen (GA)
vorgestellt. Dieses Vorgehen beruht in erster Linie auf der Verwendung von MATLAB in der
berkreuzung, Mutation und Selektion. Ein ausfuhrDurchfuhrung der genetischen Operatoren: U
liches illustratives Beispiel wird vorgestellt, um zu beweisen, da GA fahig ist, globale oder fast
globale optimale Losungen multimodaler Funktionen zu finden. Eine Anwendung von GA bei der
Auslegung eines robusten Reglers fur ungewisse Steuersysteme wird ebenfalls angegeben, um
dessen Potential in der Entwicklungstechnik intelligenter Systeme zu zeigen.
Ensenanza de algoritmos geneticos empleando MATLAB
En este artculo se presenta una aproximacion atractiva para la ensenanza de algoritmos geneticos
(AG). Esta aproximacion se basa, en primer lugar, en emplear MATLAB para la implementacion
de los operadores geneticos: cruzamiento, mutacion y seleccion. Se presenta un ejemplo ilustrativo
detallado para demostrar que el AG es capaz de buscar soluciones globales o casi globales optimas
de funciones multimodales. Se muestra as mismo una aplicacion de AG en el diseno de
controladores robustos para sistemas de control con incertidumbres que muestra su potencial en
el diseno de sistemas de ingeniera inteligentes.