a study on the convergence of genetic algorithms
TRANSCRIPT
Pergamon Computers ind. Engng Vol. 33, Nos 3--4, pp. 581-588, 1997
© 1997 Elsevier Science Ltd Printed in Great Britain. All rights reserved
0360-8352/97 $17.00 + 0.00 PII: S0360-8352(97)00198-8
A Study on the Convergence o f G e n e t i c A l g o r i t h m s
B.M. Kim Y.B. Kim C.H. Oh
Dept. o f Industr ia l The Research ins t i t u te o f Dept of Industr ia l
Engineering, Industr ial Technology Engineering,
Universi ty. o f UI-san Myon9 Ji Univ. The College of Su-won
Abstract
This paper extends genetic algorithms to achieve fast solut ions to d i f f i c u l t problem.
To accomplish this, we present empirical resul ts on the terminated condi t ion by bias
and the funct ionized model of mutation rate in genetic algerithms. The terminated
condi t ion by bias enable to reducing computation tima(CPU time) according to l imitted
and pre-astimated number o f generations. The funct ionized model o f mutation operator
reducing computation time and improving solut ion should be accomplished by applying
qui te low mutation rate on the continuing generation with remaining 95 percentage o f
b i a s . © 1997 E l s e v i e r S c i e n c e L td
Keywords : Bias, Generation, Mutation, Population
I. In t roduct ion
A successful e f f ec t o f system under the
dynamic environments usual ly demands the
adaptive solut ions. I t is so d i f f i c u l t to get
a optimal so lu t ion using noise, mul t i fu l
dimensions, a various react ion which are
uncertain in the various re lated appl icat ion
f ie lds . On these f ie lds , a successful resul t
can be obtained by using the appl icat ion o f
genetic algorithms for the global optimum. But
th is appl icat ion o f genetic algorithms to have
a 91obal optimum may be d i f f e ren t i a t ed by
f i tness function, i n i t i a l population, mutation
rate, crossover rate. The resul ts are also
changed depending on the number o f generation
o f the evolut ion or degree of convergence.
Such above parameters have been used as
various value under the given environments in
the genetic algorithms. I t would be not
guaranteed that these parameters are always
optimal. Many problems such as CPU time and
space o f storage would be influenced by the
parameters. Eventhough the adaptive parameters
were setted up to the given circumstances, a
proper combination on the parameters is
required to get more improved solut ion.
For instance Goldber9[5] researched the
technique o f computation for the re la t ionship
between populat ion size and computation time
used to convergence o f populat ion as applying
the expected number o f schemeta. I~ck[2]
su9gested mutation rate o f 1/l accelarated
search of so lu t ion in case o f pseudo-boolean
funct ion which the evaluat ion funct ion is
mul t imodal.
Meanwhile, genetic algorithma comes to an end
when i t a r r ives at the predetermined
generation upon convergence stage o f so lu t ion
or i f the increasing range o f f i tness funct ion
is less than s<:~te I imi ts(~:) [3] [7]. This
terminations techinque have the some problem
spending massive computation time and memory
according to the number of generation.
Moreover, the operat ion is continuous while
the value of f i tness funct ion increases fas t l y
on the i n i t i a l stage, slow or immaterial
increase on the some level. Such a rate o f
increase for f i tness funct ion becomes too
small on the f a l l i n g below the some level.
Hence, the total so lu t ion time may be reduced
and the qua l i t y of so lu t ion may be also
preserved well i f we terminate the algori thm
on the proper time which the increasing time
rate o f the f i tness funct ion value is too slow
or unnecessary. In th is aspect termination o f
condit ion for computation re f l ec t i ng the
character is t ics o f parameter is demanded. This
study suggests the terminated condi t ion o f
algori thm reducing the computation time which
appl ies terminated condi t ion by using bias in
the solving the problem as a genetic
581
582 Proceedings of 1996 ICC&IC
algorithms and pa r t i cu l a r l y in the case o f / <~
n, methods e levat ing the qua l i t y for so lu t ion
as improving the f i tness funct ion using the
character is t i cs o f murat ion operator.
2. The c h a r a c t e r i s t i c s o f parmmters us ing
by bias
This chapter 9ives p r i o r i t y to grasp a
charac te r is t i c point o f genetic system using
the l imit ted generation su i t ing the subject o f
problem. As grasping a character is t ics for the
various parameters o f such a system, the
analysis reducement to computation time and
pre-estimate the genetic algorithms for real
problem is enable to reduce total cost which
the number o f generation regards as a cost.
Bias d is t r i bu tes var ious ly between 50/, and
100/. as the s t r ing measures a rate of gene
wi th in population. Since such a bias is 100"/.
i f a l l s t r ings in population a r e al l the some
level on each loci. a so lut ion is not improved
any more as a l l s t r ings are coincident.
Let 's consider loci is b i t a on each
chromosome that the population size is 20, i f
15 uni ts have the value o f b i t i and the odd 5
uni ts have the value o f b i t 0 on chromosome
which loci is 9, th is bias becomes 75Y., and
th is bias appl ies the larger one out o f the
rate of 0 or 1. The s t r ing bias to use to
terminated condit ion of so lu t ion in th is study
is as fo l lowing ; i f each s t r ing bias is the
mean value o f I un i ts o f b i t biases. From such
a b i t bias, using s t r ing bias to grasp the
convergence state o f the whole population, and
we make i t the termination condit ion.
Applying the convergence rate o f genetic
a I 9or i thm, Operator i s cont i nuous and
converges to 100/. bias while get t ing connected
from the present generation to the next
generation, increasenent rate increases
gradual ly and on some level gets gradual ly
slow or immaterial, Hence, each chromosomes of
populat ion which has the bias above 95~ are
very s imi lar and converged to the speci f ied
str ing.
Af ter computing the measurement o f convergence
rate l i ke as bias, we should determine whether
i t is the termination o f algori thm or not. And
i t continuously comes into operat ion by stages
i f algori thm is not terminated.
The terminated condit ions applied to genetic
algorithms thus far is that algorithm is
continuously executed unt i l i t ' s ar r ived at
the determined maximum generation or the
increasement size o f f i tness funct ion is less
then some l imits(~). However, th is terminated
condithion have an important e f f ec t upon
population s ize and mutation operator.
To search the so lu t ion wi thin I imi ted
genration might be ef fected since the small
popu lat ion i s comparat ivel y converged upon
smaller generation then the large population.
Naximum generation arranged by population size
or the terminated condit ion by the increasing
size o f f i tness function gives r i se to trouble
to CPU time and storage space because o f the
res t r i c t i on by c ~ u t a t i o n time. Meanwhile, i f
the rate o f mutation operator is large, the
qua l i t y o f so lu t ion on i n i t i a l stage is
improved but the more number o f generation is
increased, the more i t impedes in convergence
state o f population because the mutation
operator at the convergence stage of so lu t ion
in ter feres with being a gene o f str ing. And
therefore, we are enable to ear ly confirm the
approached state to optimal so lu t ion according
to apply the terminated condit ion by bias.
Now we apply the method by bias as fol lows :
Aoplying Stage
Stage 1. Using the applying stage for ex is t ing
genetic algorithms,
Stage 2. In case that the bias is 95"/. and
over, terminating the algori thm i f
not increasing one percent or more
then the present bias, even a f te r
operat i ing from n time to 2n times.
At these above the method for termination,
operating times for algori thm could be
properly reduced according to the subject o f
problem.
3. Practical experiment
The purpose of th is experiment is to grasp the
problem point for the reduction o f computation
time, to make a pre-est imat ion and to improve
qua l i t y of so lu t ion using the terrniated
condit ion. The using computer is IB4 PC586
Pentium 100 N-Iz, the language is Nathematica
Vet 2.2.3. The subject o f experiment is the
Proceedings of 1996 ICC&IC 583
unconstraint condi t ion and the problem o f
quadrat ic fucnt ion, d iscont inu i ty ,
unimClel i ty, nonconvexity as fo l lowing :
f (x l , xz) = Integer(x, z + Xz z)
-5.12~x~, xz~5.12
This stage o f experiment suggests, select ing
character i st i cs o f parameter us i n8 B i as
applied from chapter 2, the technique that
keep the qua l i t y o f so lu t ion on time point
above rate for 95 percent a f te r reviewing and
comparing each individual genes for each
generation and solve a problera cause to
reduc i n9 c~pu ta t i on time.
3. I Influences by population size
One of the common problems in performing
genetic algorithms is how many influences
algori thm is given to populat ion size. In case
o f small population size, i t is very easy to
be prematurely converged cause occurrence o f
chromosome is too deminent or recessive in
search the so lu t ion space. And in case o f
large population size, some troubles are
ex is ted to perform the algori thm cause to a l l
search chromosome on every generation in
searching the so lu t ion space. I f the number o f
generation to be used l imit ted, some problems
are accurred to the convergent course o f
solut ion, consequently the population size to
be establ ished is adapt ively required to
perform genetic algorithms e f f ec t i ve l y to
improve the computation time for computer and
the qua l i t y o f solut ion. Fortunately, most o f
the experiment resu l ts are presented that 30
uni ts populat ion size su i t the subject of
problem, but th is case is influenced by
establ ished i n i t i a l population. Goldberg
recent ly proceeds an experiment using smaller
populat ion suggests an establishment through
the simulat ion about the size of adaptive
populat ion to e f f e c t i v e l y operate algorithm,
not to a f fec t the r e l i a b i l i t y o f solut ion.
On Figure 1, i f we consider the number of
generation of I00 percents convergence rate by
simulat ion in applying the population size
n=50 and mutation rate PaFO. OI, th is is about
14=35. Most of loss for gene in a whole
populat ion means convergence state o f
population. In such case, the i n c r e ~ t o f
populat ion size has an e f f ec t to considerably
loose the convergence state o f geno for
funct ion f, th is is understood to recluse the
poss i b i I i ty for premature convergence. Hence,
the increesement o f populat ion to avoid
premature convergence makes the convergence
rate o f so lut ion slow.
To confirm such an above assumption, Figure 2
appreared mutation rate P=--O, 01, population
size 30, 50, 100 appl ied algorithm performance
for funct ion f. And we can f ind out, in case
o f n=30 as c~mpared with population size n=50,
n=lO00 convergence rate o f 9ene wi th in
l imi t ted rect ion is remarkably reduced. That
is, the problem for premature convergene of
9erie could be e f f e c t i v e l y reduced as
increasing the population size.
In the s i tua t ion al lowed, the various
population size to maximum value f" and mean
calue j o f f i tness funct ion to measure the
f i tness o f funct ion f was applied. Figure 3
shows the maximum f i tness for funct ion f based
on population size. Convergence rate o f
algori thm is able to be quickly got from n=30
as compared with population size n=50, n=iO0
but the resu l t o f premature convergence which
is not optimal state is appeared. This
presents that a bet ter algori thm even though
convergene rate o f so I u t i on for Iarge
population is slower than small population.
Meenwhile, Figure 4 shows average f i tness for
funct ion f based upon population size.
3.2 The effect iveness for mutation rate
Mutation operator is genetic operator to be
decisive in applying genetic algorithms and
also one o f the genetic operator to apply
random walk method by the space o f s t r ing [ I ]
[11]. Mutation operator in previous research
about genetic algorithms appl ied mutation to
one b i t for every one thousand b i t s to get
bet ter e f fec t , however, in recent researches
and e f fec ts of experiment about genetic
algorithms, appl icat ion o f one mutation
operator for every one thousand b i t s has been
useless to improve the qua l i t y o f so lut ion in
case o f smaller Population size than length o f
s t r ing [5 ] [7 ] . Par t icu lar ly . in the e f fec t to
apply rate of mutation operator, the
584 Proceedings of 1996 ICC&IC
discontinuous case for the subject o f problem
has been proved to be more r e l a t i ve l y
e f f i c i e n t than the continuous case in
improving qua l i t y o f so lu t ion [4].
Losed gone can be restored as increasing the
rate o f mutation operator since the degree of
convergence is reduced according as rate for
mutation operator is increased.
But, when the rate o f mutation operator is
increased, apparently suggested research for
being influence to the performance, degree o f
algori thm is acarcely existed, the reduced
gene by premature convergence can be only
improved as increasing rate o f mutation
operator. And an increasement o f the rate for
mutation operator increases the number o f gene
which samplin9 can be taken, th is exer ts
no-9ood influence upon algori thm performance
degree.
For instance, Figure 5 shows the e f fec t to
apply populat ion size n=50, mutation rate Pe
=0.001, 0.01, 0.03, 0.1. And Fi9ure 5 shows
average convergence rate o f gene based on rate
o f the various mutation operator. Since the
more rate o f mutation operator is increased.
the more convergence rate o f 9ene get low,
performance degree for algori thm is given some
troublers. In case that premature convergence
for 9ene is accurred, i t could be solved
problems as increasing the rate o f mutation
operator but convergence ef fect iveness based
on performance degree o f algorithm should be
surely considered.
Figure 6 is showing the ef fect iveness for
maximum performance degree of genetic
algorithms when the rate o f mutation operator
is increased.
In th is Figure 6, as B~ck presented, i t can' t
be v e r i f i e d maximum f i tness defree o f
algorithm is the best one when the rate o f
multat ion operat ion is 1/[. Increasement o f
rate for mutation operator have an
ef fect iveness to improve the f i tness o f
f i tness funct ion upon i n i t i a l stage when
algori thm is performed, but th is is considered
to be due to rather cause troubles for
improvement o f so lu t ion a f te r that the
generation is proceeded on a cer ta in degree.
In case that mutation operator o f low rate was
applied, f i tness o f f i tness funct ion cause by
premature convergence o f gene is reduced
according as generation is proceeded.
Figure 7 shows by a diagram for convergence
ef fect iveness of so lu t ion for average f i tness
degree based on increasement o f rate for
mutation operator when genetic algorithms is
appl led to funct ion f. And also the
increasement o f random rate for genetic
operator l i ke mutation reduce a performeance
degree o f genetic algorithms could be found
out.
Table 1 presented the resu l t to solve the
subject o f problem to f ind the computation
time and precision of so lut ion for terminated
condit ion by bias. By experiment data(30 uni ts
for population size 10, 50, 100), computation
time o f average 9W. was saved as compared with
1000 generations o f the typical genetic
algorithms which is the present, terminated
condit ion o f algorithm, and i t was maintained
so lu t ion qua l i t y of 92"/, from the problem
knowing optimal solut ion.
4. E f fec t i ve convergence of so lut ion
4. I Improvement of so lu t ion based on change
of b ias
One o f the problem points which could be
occurred in searchin9 so lu t ion o f the subject
o f problem based on genetic algorithms is that
the p o s s i b i l i t y o f premature convergence is
appeared on performin9 algorithm. This
premature convergence is var iously occurred
depending on value o f p robab i l i t y or number o f
generation. Especial ly, qua l i t y so lu t ion is
decreased is the case of />n, that is, most
solut ions would be prematurely converged in
the event that the size o f population is
smaller than s t r in9 length. However, there is
no technique to confirm in advance the
premature convergence state as the method o f
the present genetic algorithm. Improved
so lu t ion can be get, no enter ing a premature
convergence state as making populat ion size
large, but the expected ef fect iveness is not
much cause by r e s t r i c t i o n of computation time
and problem of memory. And the change o f
crossover rate can be considered, however, new
chromosomes consist ing o f next generation are
very herd to be expected due to the occurrance
of s imi lar chromosomes based on increasemant
Proceedings of 1996 ICC&IC 585
fo r the number of generation in genetic
algorithms.
The change of mutation rate play a decisive
ro le for performance o f algori thm as the
generation is proceeded. Since the
predetermined mutation rate under the
performance o f algori thm is applied constant
value unt i l the termination for algorithm, i t
is o f no use in confirming premature
convergence. Hence a new technique to improve
the qua l i t y of solut ion, not entering
premature convergence state on performing
algorithm. This section presents the qua l i t y
o f so lu t ion based on increasing the mutation
rate on stage eighty percent [examPle;
populat ion size 10, ra ise mutation rate up to
0.3] that the p o s s i b i l i t y for new gene
occurrence is decreased.
To ident i fy above case, from Figure 8 to
Figure 10 showed the technique to improve
p o s s i b i l i t y for premature convergence as
ra is ing up the rate o f mutation operation on
stage o f bias 80"/. in the population size n=10
which has high p o s s i b i l i t y o f premature
convergence. From the experiment results, a
case of the small population size n=10,
pa r t i cu la r l y , on the discontinuous subject of
problem, the correctness degree for tota l
so lu t ion is appeared to about 95% of optimal
solut ion, and computation time is reduced to
91% as compared with the typical genetic
algor i thins techniques.
4.2 An appl ication of functionized mutation
operator for iBproving solution
Parameter exer t ing a d i rec t influence to
algori thm performance is a population size and
the rate of mutation operator. Especial ly the
rate o f operator act as a d i rec t factor to
improve so lu t ion in f ina l stage when we
perform the algorithm. Accordingly, th is
sect ion suggests that the large mutation rate
is appl ied to diminish the stage o f premature
convergence in the i n i t i a l stage o f algorithm
to improve the qua l i t y of solut ion, and that
the small mutation rate is applied to improve
so lu t ion in the convergent stage for optimal
so lu t ion on generation process.
Let us simply express the rate o f mutation
operator o f funct ionize form to improve the
quat i ty o f solut ion.
P= = a exp(-~tN), a < 0.1, /t ~ 0.1
From this, a and ~ are not exceeded 0.1,
which is mutation rate l i ke as tentave search
method in the i n i t i a l stage for algorithm, and
we apply the rate in so small that i t couldn' t
be interrupted to improve solut ion at a point
o f time maintained as bias above 95% on
9enerat ion process.
Since the crossover operation in the bias
s tate above 95% acts a rate of mutation
operator, and from a point of time maintained
as bias above 95%, i t may be e f f i c i e n t to
maintain the rate o f mutation operat ion
exert ing a influence to qua l i t y o f so lu t ion in
so small. An applying the case o f n=lO which
p o s s i b i l i t y of premature convergence is high,
n=50 which the improvement degree of algori thm
is also high and crossover rate 0.6, and the
experiment using terminated condit ion that is
previously suggested was presented by draw on
Figure ] l and Figure 12. The resu l t o f
experiment to apply a mutation rate for the
function form, the qua l i t y o f so lu t ion doesn't
have some change as compared with the present
techniques but computation time is come down
to 94%. This appears an ef fect iveness to save
the computation time.
5. Conclusion
In th is paper, a terminated condit ion o f
genetic a 19or i tl'lns reducing computaion
time(CPU t i m ) by using bias to solve the
given problem has been proposed. And also,
funct ionized mutation operator was u t i l i z e d to
improve f i tness function, as a resul t , the
method for ira Droving premature has been
studied.
The resul ts for various numerical
experimentations to apply terminated condit ion
by bias to the discontinuous funct ion have
been proved to 9et an optimal so lu t ion in
tota l experimentation even though there is a
some d i f ference according as populat ion rate,
mutation rate and crossover operator rate.
Accordingly, th is study set up the computation
time with in number o f generation enabling to
perd ic t ion and charac ter is t i c for parameter to
improve the qua l i t y of so lut ion as using the
s im i l a r i t y re la t ionship a~ong genes. I t was
586 Proceedings of 1996 ICC&IC
confirmed through the experimentation that
th is technique, as compared with the typical
genetic algorithms, d idn ' t come down the
qua l i t y of so lu t ion and enabled to make
computation time reduced above 91~.
Addi t ional ly , we obtian that the funct ionized
mutation rate should be decreased during
convergence. The assumptions made for th is
explanation regarding the frequency o f
absorption end the losses due to mutation rate
can in p r inc ip le be v e r i f i e d experimentally.
This approach to determine the funct ionized
mutation rate seems fundamental and an
experimental ve r i f i ca t i on of the assumptions
used here is planned.
References
lO. Stephanie,
1. Yong-Beom Kim, "A Study on the
convergence o f Genetic Algorithms', PH.D
thesis, Univers i ty of Myeong-Ji,1996.
2. B&ck, T., "The Interact ion of
Mutation Rate, Selection, and
Self-Adaptat ion wi th in a Genetic
Algorithms', In Reinhard, M. and Bernard,
M., pp 85 - 94,1992.
3. De Jon9 K. A., "An Analysis o f the
Behaviour o f a Class o f Genetic Adaptive
Systems", P h . D . thesis, Univers i ty o f
Michgan, 1975.
4. Goldber9, D. E.,"Genetic Algorithms" in
Search, Optimization and Machine
Learning, Addison-Wesiey, Reading, MA, 1989.
5. Hesser, J., l~rmner, R., "Towards an
Optimal Mutation Probabi l i ty for Genetic
Algorithms', In Schwefel, H. and M~nner,
R., pp 23 - 32,1990.
6. Michalewicz, Z., "Genetic Algorithms
+Data Structures=EvolutionPrograms ", Spring
- Verlag, 1993.
7. Nicholas, J. R., Fe l i c i t y , A. W., "A
Study in Set Recombination', In Stephanie,
F., pp 23 -30,1993.
8. Reinhard, M., Bernard, M. , 'Para l le l
Problem Solving from Nature',Proceedings
o f the Second Conference on Paral le l
Problem Solving , North-Holland, 1992.
9. Scheefel, H., Minner, R. , "Para l le l
Problem Solving from Nature', 1st Workshop,
PPSN1, Springer-Verlag, 1990.
F. , 'Genet ic AI9orithms',
Proceedings of the f i f t h internat ional
conference on Genetic Algor i thm, Morgan
Kaufmann Publishers, Sen Mateo, CA, 1993.
i J= l
Figure 1. Bias o f f i tness funct ion f for
n=5O, P,=O, O. 001, O. 01, O. 05, O. 1
L ~
| =;
, o y Figure 2. Bias of f i tness funct ion ] for
n = 30,50,100
~o
t s
i ' °
20
Figure 3. Evaluate o f maximum f i tness
funct ion ] for n=30 ,50 ,100
o~ | i J
-LLo. ' ~ ' ° °
Figure 4. Evaluate of average f i tness
funct ion ] for n=30 ,50 ,100
Proceedings of 1996 ICC&IC 587
Ii I I '~ l ,~/ t J
SO ~ ISO ~
Figure 5. Bias of fitness function J for
P.=O.O01,O.Ol,O.03,0.1
M
I"
¢ 1 1 u l
~s
=o
Figure 6. Evaluate of maximum f i tness
function ./ for P,=O.O01,O.Ol,O.03.0.1
Figure 7. Evaluate of average f i tness
function J for P.=0.001,0.01,0.03,0.1
if/ . [
Figure 8. Bias of J by mutation rate at the
state of bias 80"/. for n - - lO
s a m I m l
,! J o l
= = l
Figure 9.
j u
Evaluate of maxirnum fitness
function J by mutation rate at the state of
bias 80/, for n = ] O
* =
I" Jo
= l
= *
Figure 10. Evaluate of average f i tness
function j by mutation rate at the state of
bias 80"/. for n = l O
tD sB
l"
==
Figure 11. Evaluate of maximum fitness
function ] by functional mutation rate for
n=lO
,=
i"
Figure 12. Evaluate of maximum fitness
function ] by functional mutation rate for
n = 50
588 Proceedings of 19961CC&lC
Table 1. Comparative results between t radi t ional G.A, method of bias, method of Bias
80~ by mutation vari ty, and functionized mutation model
Terminate
method
Pop_size
Terminae
generation
Quality of
solution(%)
Computing
time(sec)
~ i t i o n ~ G.A
10 ~ 1~
1 ~ 1000 1 ~
~ .0 2 .8 97.4
5~.1 ~ . 8 9101
Method of Bias
10 50 100
21 60 108
82.0 96.8 97.7
15.0 234.6 932.6
Method of Bias 80%
by mutation varity
10 50 100
25 68 123
~ .7 97.0 97.5
18.9 267.7 1067.1
Functional
mutation Model
10 56 100
17 49 96
~ .0 ~ .3 ~.3
12.2 156.6 759.0