Com plex Systems 9 (1995 ) 431-454

Real-coded Genetic Algorithms with Simulated

Binary Crossover: Studies on Multimodal and
Multiobjective Problems

Kalyanmoy D eb*
Amarendra Kumar
Departm en t of Mechanical Engineering ,
Indian Inst itute of Technology,
Kanp ur, UP 208 016, Indi a

A bstract . Real-coded genet ic algorit hms (GAs) do not use any cod-
ing of the problem variab les, instead they work dir ectly wit h t he vari-
ab les. The main differen ce in the implementation of real-coded GA s
and binary-coded GA s is in t heir recombination op erators. Alt ho ugh a
number of real-cod ed crossover implement ati on s wer e suggested, most
of them were developed wit h intuition and wit hout much analysis. Re-
cen tly, a real-cod ed crossover op er at or has been developed based on
the search characteristics of t he single-point crossover operator used in
binary-coded GAs. T his simulated binary cros sover (SBX) ope rator
has been fou nd to work well in many test problems having continuo us
search space when compared to exis t ing real-coded crossover imple-
mentation s. In this paper the performan ce of t he real-cod ed GA with
SBX in solving mu lt imodal and mu lti ob jective problems is further
investigat ed . Sharing function ap proach and nond omi na ted sort ing
implement ati ons are incl ud ed in the real-coded GA with SBX to solve
mult imodal and mult iobjective problems, resp ecti vely. It is observ ed
t hat t he real-coded GAs pe rform equally well or bet t er t han binary-
coded GAs in solving a nu mbe r of test pro blems . One advant age of the
SBX op erator is t hat it can restri ct childr en solut ions to any arb it rary
closeness to t he parent solutions , t hereby not requi rin g any separate
mating restrict ion sche me for bet t er p erformance. F inally, rea l-coded
GA s with SBX have b een successfully used to find mu lt iple P ar et o-
optimal solut ions in solving a weld ed beam design pr obl em . These
simulation results ar e encour aging a nd suggest the applica t ion of real-
cod ed GAs with SBX op er ator to rea l-world optimization problems at
large.

*Electronic mail address: debClliitk. ernet. in.

432 K alyanmoy Deb an d Amarendra Kumar

1. Introd uction

W it h the success of binar y-coded genet ic algorithms (GAs) in var ious opti-
mization problems, real-coded GAs are finding som e at te ntion pr imarily in
solving cont inuous search space pr obl ems. Real-coded GAs differ from the
binar y-cod ed GAs in the coding of the problem variab les. Since the problem
vari abl es are used dir ectl y in real-coded GAs, there lies a need for developing
new, yet efficient , crossover and mu t at ion ope rators . Alt hough t here exist
a number of st udies of real-cod ed GAs wit h different cross over and mu t a-
tion opera tors [1, 2], recent ly a crossover and a mutation ope ra tor have been
developed by simulat ing the working of their bina ry count erparts [3] . In
t hat st udy, the main motivation was to develop real-cod ed geneti c ope rators
havin g similar search power as t hat in the binar y genet ic op erators. T he sim-
ulated bin ar y crossover (SBX) operator used in [3] had sea rch power similar
t o t hat of a single-point binary-coded crossover operator. The sea rch power
was defined as t he ability to create any arbit rary child solution from two par-
ent solutions . Based on a derived probabi lity distribut ion of creating a child
solution in the single-p oint crossover ope rator, a similar probability distribu-
t ion was used dir ect ly to choose a child solution in SBX. In t hat st udy, t he
perform an ce of t he proposed real-coded GAs with SBX was compared wit h
some of the earlier real-coded GAs and t he bin ary-coded GA. Motivated by
the success of t he proposed real-coded GAs in [3], we extend it s ap plication
t o mult imodal and multiobj ecti ve fun ction optimization pr obl ems.
In this pap er we br iefly describ e the SBX ope rat or and exte nd the concept
of sharing in the real-coded GAs to solve a number of mult imodal functions
for mult iple soluti ons simultaneously. Since children solut ions arb itrar ily
close to t he par ent solutions can be create d using t he SBX operat or , t here is
no need to use a separate mating restrict ion scheme . Thereaft er , we extend
the principle of nondominat ed sorting in real-cod ed GAs to solve a number of
multiobj ecti ve optimization problems for P ar et o-opt imal solutions . To show
the efficacy of the proposed techniques, we also show simulation resu lts of the
pr esent ed sharing and nondominat ed sort ing te chniques to optimize the engi-
neering design of a welded beam. The successful working of these techniques
suggest s t hat the real-cod ed GA using the SBX operator perfor ms similar to
the binar y-cod ed GA and can be used efficiently in solving cont inuous sea rch
space pr oblems.

2. Simulated b inary crossover

When problem variables ar e directl y used in a GA, binary-cod ed crossover

op erators can no longer be applied . A number of real-coded crossover op-
erato rs have b een develop ed that create two children solut ions from two
par ent solut ions. In most cases, a probability distributi on centering t he par-
ent solutions is ass ume d and two children solutions are created based on t hat
probability distribution. Creat ing children solut ions using a fixed probability
distribution , which do es not depend on t he location of the parent solut ions,
St udies on Mu ltimodal and M ultiobj ective Problems 433

makes the search adapt ive. For examp le, if the par ent solutions are close
t o each other , the children solut ions are expecte d to lie in the neighborhood
of t he parent solutio ns . On t he other hand, if t he parent solutions are far
away from each ot her , childre n solutions far away from t he par ent solut ions
are expected. In early generations of a GA simulat ion, parent solutions are
expected t o be away from each ot her and almost any solution can be cre-
ated as a child solu tion. But when the search converges towards a solut ion,
par ent solut ions becom e similar and children solutions also becom e closer t o
t he par ent solut ions . T his adaptiveness in t he sea rch power of t he real-cod ed
cross over ope rator is simila r in principle to the search power of t he binar y-
coded single-point crossover ope rator. However , one m ain differen ce is t hat
for t he binar y-cod ed crossove r ope rator , no explicit pro bability distribution
is used to creat e a child solution . But t here is an implicit prob ability dis-
t ribution t hat dep ends on t he st ring lengt h used t o cod e t he variable. In a
real-cod ed crossover operator , a pro bability distribution is explicitly used t o
create a child solution . Sinc e an explicit probability dist ributi on is used , t he
performan ce of real-cod ed GAs dep end on that dist rib ution . In earlier real-
coded GA implement ati ons, the cho ice of the probability dist ribut ion was
somew hat arbitrary and based on int uition . Recent ly in [3], a SBX op era-
tor has b een sugges te d with a probability distribut ion simila r t o t hat in the
single-point crossove r op erator used in binary-cod ed GAs . In t he following ,
we briefly describe t hat crossove r ope rator.
It has been observed t hat in the binary sing le-po int crossove r ope rator,
both children solutions lie either inside (cont racting crossove r ) or outside (ex-
panding crossover) t he region bounded by the par ent solut ions . Mor eover , t he
dist ance of one child solution from one par ent is exac t ly the sa me as t hat of
the other child from the other par ent solut ion . Further , there is no apparent
bias for eit her cont racting or expanding crossover, on expec tat ion . T hus , on
an average, the overall prob abilit ies of contracting an d expanding crossove rs
are the sa me. T he SBX op erat or for real-coded GAs has b een developed
wit h the above pr op erties. In ord er to imp lement this cross over ope rat or for
any two parent solution s PI and P2, a nondimension alized spread factor 13 has
been defined as the rat io of t he spread of creat ed childre n solut ions CI and
C2 to that of the parent solutions as follows:

f3= I~I
PI - P2
(1)

It can be shown that 13 ::::: 1 correspo nds to a cont rac ting crossover and 13 ~ 1
corresponds to an expanding cross over. Writing t he decod ed values of two
arb itrary par ent st rings in terms of t heir allele values (0,1) and writi ng t he
children st rings created from the parent st rings as a function of t he cross-sit e
k along the st ring length, the spread fact or 13 has b een writ ten in t erm s of k .
From t his dist ribut ion , it has also been p ossib le t o calculate t he probab ility
of creat ing a pair of children solut ions having a certain 13. That probability
distribution has been approximated by a polynomial pro babilit y distrib ut ion
434 Kalyanmoy Deb and Amarendra Kum ar

l'
~ II ~

n
).1,
c:
o
5
."<;;
.0
I"' Ii
is I, : !
~ I
I' ,\'\
:c
co I ,\
.0

,,
o I : \
<t \
\
I
/

Figure 1: Probability dist ribut ions used in the simulated binary

crossover operator are shown for different values of th e distribution
index n.

as follows:

0.5(n + l) ,6n, if,6 :::; 1;

,6 = {
P () (2)
0.5(n + 1) {3}+2' otherw ise.

In equat ions (1) and (2), t he distribution index n is any nonnegati ve real
number. A large value of n gives a high er probability for creating solutions
near to the parent and a sm all value of n allows distant po ints to be selecte d as
children solut ions. F igure 1 shows t he probabi lity dist rib ution as a fun ction
of ,6 for different values of the dist ribution index n . T he advantage of using
a pr ob ab ility dist rib ution as a fun ct ion of ,6 is t hat t he create d childre n
solut ions are relative to the par ent solut ions. If the parent solut ions are
distant , children solut ions far away from the parent solution s can be create d .
On the other hand , if t he par ent solutions are close to each other , t he children
solu tions, in general, cannot be far away from t he parent solut ions .
In order to create two children solu t ions CI and Cz from the parent solu-
tions PI and pz using the ab ove probab ility dist rib ution , t he following proce-
dure is used .

Cr eat e a random number u between 0 and 1.

Find a ,6' for which t he cumulative probability

fo{3' P (,6)d,6 = u. (3)

Know ing t he value of ,6', t he children po ints are calculated as

CI 0.5 [(PI + pz) - ,6' lp2 - PII] ,

C2 0.5 [(PI + P2) + ,6' lp2 - PI ll
St udies on M ult imo dal and Multiobjective Problems 435

T he pr eceding SBX op erator is allowed t o create any solut ion in the

entire real space [-00,00], however , it can also be modified for var iables wit h
kn own lower and up p er bounds (x~L) :::; X i :::; x~U) ) . This can be achieved by
modifying the prob abi lit y distribution (equation (2)) so that t he pr obabi lit y
for a solut ion outs ide t he above bo und s is zero. A simple way to achieve this
is to first calculat e the cumulat ive pr obab ilities
f3{L)
p 'I 10 P ((3 )d(3 , (4)

p'2 Jo
r: P ((3 )d(3. (5)

In equations (4) and (5) , t he parameters (3(L) and (3(U) ar e t he spread fact ors
for t he lower and upper bou nd s of the problem var iable, resp ect ively:

T hereaft er , two spread facto rs (3~ and (3~ are calculat ed using mod ified prob-
ability distributions P ((3) /P~ and P ((3) /P~ , resp ect ively, (by using equa-
t ion (3)) t o create two children solutions as follows:

CI 0.5 [(PI + P2) - (3~ lp2 - PI I] ,

C2 0.5 [(PI + P2) + (3~lp2 - P I ll .

T hus, t he modified prob abi lity distribut ions do not create any solution out-
side t he given bounds, inst ead they scale up t he probabi lity for solutions
inside t he bo un ds , as shown by the solid line in F igur e 2. In t he figur e,
t he par ent solutions are PI and P2. The probability distributions for t he un-
bounded case (x E [- 00, 00]) and t he bo unded case (x E (x(L), x(U))) are
shown in dashed and solid lines , resp ect ively. It can be observed in the fig-

- - - - Unbounded
- Bounded
c:
o
5
.D
C
0;
is
~
:0
"'o
.D

a:
I
/

x (L) p,
x

F igur e 2: P rob ability distribu ti on s are shown for bounded and un-
bounded cases .
436 Kalyanmoy Deb and Amarendra Kum ar

ur e that a solu tion outside the lower and upper bounds has zero probability
to become a child solution, t hereby restricting t he search wit hin t he given
bounds.
The focus of this pap er is to st ud y the performance of the real-cod ed
GA with SBX on multimodal and mu ltiobj ect ive problems . The applica t ion
of binar y-cod ed GA s for the solut ion of these problems is finding increasing
attention. This is because the solution of such problem s requ ires mu ltiple
solutions to be found simultaneously and GAs ar e particula rly suitable for
these problems b ecau se of their inherent population ap proach . In the fol-
lowing , we first investi gat e the applicat ion of real-cod ed GAs to mu lt imodal
problems and then st udy t he multiobj ect ive problems.

3. Multimodal problems using sharing genetic algorithms

A mu lt imodal problem cont ains mu lt ipl e opt imal solutions in it s search space.
The objective of a mu ltimodal fun ction opt imiza t ion procedure is to find
mu ltiple optimal solut ions having eit her equal or un equal objective fun ct ion
values. The kn owledg e of mu ltiple op t imal solut ions is particularl y useful to
design engineers for choosing an alternative optimal solu tion, as and when
requ ired . In solving for multiple optimal solut ions, tradition al optimization
algorit hms need to be applied as many t imes as the number of optimal so-
luti ons. This is becau se most of those algorithms are point-by-point search
methods and can on ly find on e op timal soluti on at a time. Sin ce GAs create
a population of solutions inst ead of one solut ion in each it erati on , it may
b e possible to capt ure mu lt iple optimal solut ions in t he population, thereby
allowing GAs t o be applied only once t o find mu ltiple solut ions . After the
pioneering work in [4], resear chers have also develop ed different GA s to find
and main t ain a stable subpop ulat ion of optim al solut ions in the population
(e.g. , [5, 6]) .
In solving multimod al pr obl ems using GA s, the st udy in [4] was moti-
vat ed by a "gedanken" expe riment based on a modified two-armed bandit
problem. It was observed tha t if the reproduction phase is performed with a
modifi ed fitn ess fun ction obtained by degrading the original fitn ess value of
a solution by a cumulat ive measure of the proximity between that solut ion
and the rest of t he population, a stable subp opulati on can be maintained
in t he populati on . They defined a proximit y measure (di j ) between any two
solutions i and j eit her phenotyp ically (with t he problem vari ab les dir ectl y)
or genotyp ically (with corres po nding st rings) . A sharing fun ct ion Sh(di j )
is used to define the sharing effect of on e solut ion t o another based on the
proximity measure of two solu tions as follows:

_ { 1 - d~i , if di j 2: a ;
Sh (di j ) - (6)
0, otherwise.

The par am et er a is the maximum dist an ce between two solutions t hat can
be long to on e optimal solut ion . Thereafter , the cumula t ive proximity (m;) of
Studies on M ult imodal and M ult iobjective Prob lems 437

the ith solut ion in t he ent ire population (of size N) is calculated by summ ing
t he indi vidu al sharing function values as follows:
N
m~ = ~ Sh(di j ). (7)
j=l

T he repro ducti on is t hen performed wit h the shared fitn ess ii = h l mi in-
stead of the original fitn ess I; T his sha ring t echn ique has been successfully
applied to many mu lti modal problems [5, 6J .
We implement t he st a ted sharing functi on concept in real-coded GAs. It
is important to note that th e implement a tion of sharing is only in the repro-
du ction ph as e; th e crossover and mutation opera tors need not b e changed .
The primary differences in t he implement ati on of binary-cod ed and real-
coded GAs are in t he coding and in t he recombination operat ors. T hus, the
concept of sharing can also be used in real-coded GAs . Since, in real-coded
GAs th e problem variab les are used dir ectly, phenotypic sharing is a natural
choice. To calculate the proximity measure di j and the paramete r (J , the
following guidelines were suggested in [5J and ar e used here:

di j t
k= l
( x k(i ) - x k(j)r , (8)

(J
1
- -
2{!q
t( x k(U ) - x k(L) r , (9)
k= l

where p is t he num ber of variables and q is t he desired number of optimal

solutions .

3 .1 Simulation r esults
T he real-coded GA wit h SBX operator and t he shared reprodu cti on scheme
have been implement ed to solve a number of t est pro blems obtained from
the lit erat ure and also to a number of random bimod al functions. T he per-
form anc e of real-cod ed GAs has b een compa red wit h binary-cod ed GAs on
all t est fun ct ions.
In solving mult imodal problems using GAs , t he pr imary objective has
been not only t o find all optima l solut ions but also to dist ribute th e po pu-
lation memb ers well amo ng mu ltiple optimal solut ions [5]. To compare t he
performance of different GA implement ati ons, a deviat ion meas ure similar
to chi-squar e was suggeste d in [5]. T his meas ure calculat es th e deviation of
th e distribution of the population from an ideal distribution, which can be
obtained using the modified two-ar med bandit gedanke n expe riment from
[4]. The chi-square deviati on measure is given as

(10)
438 Kalyanmoy Deb and Amarendra K umar

where Tk and Tk ar e th e expec ted valu e and standard deviation of the number
of ideal solut ions in the kth optimal solution in the population, and rk is the
actual number of th e solutions in the kth optimal solution in the populati on .
In all our simulations, all solutions havin g a fitn ess greater t han or equal to
70 percent of t he globally b est fitness are counte d as the number of solutions
(rk) near th e corres ponding optimum. However, the suggest ed valu es of the
par am et ers Tk and r k of th e ideal dist ribution for q optimal solut ions having
fun cti on valu es f k' k = 1,2, . .. , q are as follows [5J :

For nonpeak solutions, Tq+l = 0 and r q+l = V~k=l T~ . Equation (10) sug-
gest s that smaller deviat ion measures W yield better distribut ion of t he pop-
ulati on on t he optimal solut ions.

3.1.1 Five test function s

Fiv e t est functions , which were used in [5], ar e solved using real-cod ed GAs.
The fun ct ions, th eir dom ain , and the description of the optimal solutions are
given as follows.

MMl: sin 6 (5?rx) 0 :::; x <1

Fi ve equispaced maxima wit h equa l fun ct ion
values .
MM2: exp (-2In2 (X~.~.lr) sin 6(5?rx) 0:::; x:::; 1
Fiv e equispaced maxima with unequ al fun c-
ti on values.
MM3: sin 6 (5?r(x0 75- 0.05)) 0 :::; x :::; 1
Five unequispaced maxim a with equal fun c-
tion values.
MM4: 6(
exp ( - 2I n 2 ( X~.~.lr) sin 5?r(x0 75- 0.05)) 0 :::; x:::; 1
Five unequispaced maxim a with unequal func-
t ion values.
MM5: [1 - ((xi + X2 - 11)2 + ( Xl + x~ - 7)2) /2186J - 6 :::; Xl , X 2 :::; 6
Four maxima with equal fun ction values.

The simulat ion results are compared with binary-cod ed GAs having a
single-point crossover operator. The following GA par am et ers are used in all
simulations.
Studies on M ultimodal and Mul tiobj ective Problems 439

Populat ion size N: 100

St ring length R.: 30 (for binary-coded GAs )
Crossover pro bability: 0.9
Mutation prob abili ty: o
Distribution index n : 0-500 (for real-coded GAs ).

F igur e 3 shows the deviat ion measure ijJ versu s the distribution index n
on the function MMl. In order to invest igat e the effect of n on t he per-
formance of t he real-cod ed GA s, we have varied n from 0 to 500. Average
deviation measures for generati on s 101 to 200 ar e plot t ed to demonstrat e
that the algorit hms no t only distribut e t heir p opulati ons well among all the
optimal solutions , but also maintain the distribution for a lar ge number of
generations . Fi gure 3 reveals t hat for small values of n , the SEX operator is
unable to maintain stable subpopulations on all five optim al solu tions . The
horizont al das hed lin e is t he deviation measure obtained for t he binar y-cod ed
GAs with single-point cross over on t his function. Fi gur e 3 also shows that
real- coded GAs perform better than binary-coded GAs when n is some what
lar ger than 30. T he distributi on of p opulati on members is further illustr at ed
in Fi gure 4, whe re the deviation measure is plotted with generat ion numbe r.
Figure 4 shows t hat the real-coded GAs with n = 35 and binary-cod ed GAs
perform mor e or less the same and t hat t he real-cod ed GAs wit h n = 200 pe r-
form mu ch better than t he binary-coded GAs . T he reason a comparatively
larger value of n is required to adequat ely solve the problem can b e explained
as follows. T he deviati on measure described in equation (10) becomes b etter
if the pop ula t ion is well dist ributed among t he optimal solut ions and the
numbe r of lethal individuals (solut ions not belonging to any opt imal solu-
ti on) is low. A little thought will revea l that if two in divid uals belonging t o
two different optimal solutio ns are allowed to cross over , lethal solutions may

10
~ 9
::J
~ 8 Binary GA
'"c
::; 7
RealGA
Binary GA with Mating Re striction
.1
tii
":;: 5
o'"
'"a>co 3
Q;
>
<I:

o L.:.::::'=~==::::::';:'==~::;::::;::3:'::':~~~
o 50 100 150 2 00 250 300 350 400 4 50 500
Distribution Index, n

F igure 3: Average deviation measure (for generations 101 to 200) is

plotted with distributi on in dex n for function MMl.
440 K alyanm oy Deb and Amaren dra Kumar

13 "
.--~-~---~~-~------,
12
11 Binary GA
10 Real GA (n=35)
Ql
~ 9 - Real GA (n-200)
'"
~ 8
::;; 7
c;
o 6
~ 5
~ 4
3
2
1
o L::::iS~~~~~~S~~~~::::::::J
o 20 40 60 80 100 1 20 1 40 160 180 200
Generat ion Number

Figure 4: Deviation measure is plotted wit h generation number for

function MMl.

be created . Und er SBX , t his po ssibility can be minimized by suitably choo s-

ing a dist ribution index, n . In function MM 1, the two ex treme peaks ar e a t
x = 0.1 and x = 0.9 , resp ect ively. Since solutions having a function value
greater than 0.7 are considered to belong to t he niche of an optimal solution ,
the difference between a child solut ion from t he closest parent solution must
be at most equa l to 0.022. T hus, for contracting crossovers, the minimu m
spread factor must b e equa l to f3 = 0.945. If we want t o succeed in 99 per-
cent of crossovers , the correspo nding n can be approximately calculate d by
equa t ing the cumulat ive probab ility of success to 0.99 , as follows:
1 - f3n+1 = 0.99.
For the two ext reme opt imal solutions, the ab ove equat ion demands n R> 80.
For t he two near est optimal soluti ons (x = 0.1 and x = 0.3) t he requ ired
dist ribution index is n R> 18. T he value of n = 35 observed in Figur e 3 is a
compromise between t hese two values of n.
In ord er to improve t he performance of binary-coded GAs wit h sha ring ,
oftentimes, a mating restricti on scheme [5] is used . In a mating rest ricti on
scheme, similar solut ions are only allowed to mate wit h each other, thereby
disallowing t he creation of lethal solutions . The SBX ope rator, on the other
hand , can restrict children solutions to lie in t he vicinity of the parent solu-
tions by using a large value of t he distribu tion ind ex n . Thus, t he mating
restrict ion scheme may not b e necessary with t he SBX operat or to improve
performan ce (eit her on-line p erformance as defined in [7] or our deviation
measure \If) of GAs . In F igur e 3, the dashed-dot line shows t he deviat ion
measur e of the binar y-cod ed GAs with sha ring and ma tin g rest rict ion . It
can be seen in t he figur e t hat t he real-coded GAs with n higher t ha n abo ut
150 can achieve similar or bet t er p erformance t ha n binary-coded GAs wit h
sharing and mat ing rest ricti on schemes t oget her .
Similar experiments are perform ed using the ot her four tes t functions and
simila r result s are obt ained. Fi gur es similar to Figur es 3 and 4 are pr esented
in the Appendix for the fun ct ions MM2 t hrough MM5 .
St udies on M ult imodal and Mu lt iob jective Prob lems 441

3.0

2.5 + I ~ Real GA
Binar y GA

'"
0
<::
+
+ + +
+

'E" 2.0
+ + + +
~ 1.5
+ +

'"
a.. +
'"
Cl + + +
'Qj>" 1.0
+ +
+ +
+
+ + +
0.5 + + + +
o 0 0 0
00 oo .oC?o90ooo0n066 apq)ooooOoOOoOOOOQOoOO~?~OoO '
0.0
o 10 15 20 25 30 35 40 45 50
Function Number

Figur e 5: Average deviat ion measure (from generations 101 to 200)

of real-coded GAs and binary-coded GAs are plott ed for 50 different
bimodal functions.

3.1.2 Random bimodal function

To fur ther compare t he p erforman ce of real-cod ed GA s with SBX and binary-
coded GA s, we have created 50 fun ctions having two peaks locat ed at any
tw o random po ints in the range (0,1). The fun ctional form of t he two-p eaked
fun ction (with opti mal solutions close t o X l and X 2 ) is as follows:

MM6: O:S X:Sl.

The paramet er b acts like the spread of the opt imal solu tions. We use MM 6
with 50 different random combinat ions of X l, X2 E (0, 1) , and b E (0,0 .05) t o
create 50 bimodal functions .
Binary -code d and real-cod ed GA s are used t o find both op timal solu-
ti ons in the sa me 50 random fun cti ons wit h GA parameters as menti on ed
earlier. The avera ge of t he deviation measure \II of the simulation runs from
generation s 101 to 200 is calcula te d in each case and shown plotted versus
t he fun ct ion number in Figur e 5. The figur e shows that t he p erforman ce of
real-cod ed GAs wit h SBX (n = 200) is consistent ly bet t er t han that of the
binar y-coded GAs. T his is expected becau se in real-co ded GAs t he search
power can be cont rolled using the distribution index, n , With a large value
of n , the real-c od ed GA wit h SBX beh aves like a bin ary-cod ed GA with
sharing and mating restricti on. With large values of n , t he sea rch is always
confined in the vicinity of t he parent solutions . T hus, the soluti ons improve
margin ally in each generation and t he combined effect of shared reproduc-
tio n and reco mbination ope rat ors guide t he search par allel y towards each
optimal solut ion . Since, in t he above simula t ion of bin ary-cod ed GAs t he
mating restrict ion scheme is not used , there is no restrict ion for the solu-
t ions from different op t ima t o mat e wit h each other. T hese cross overs some-
442 Kalyanm oy Deb and A marendra K umar

times create some nonoptimal solut ions, which deterior ate the pe rform ance
of binary-cod ed GAs.

3 .2 Fractional sh aring
One crit icism against t he implement ation of shari ng st rategy in GAs is t hat
for every individual in the population the sharing fun cti on needs to be calc-
ulat ed for every ot her indi vidual [4, 8]. T his requir es N 2 evaluat ions of
sha ring functions at each generation . However , if pr oper book-keeping is
maintained , this complexity can be reduced to half b ecau se t he sha ring func-
tion is commutative (i.e., Sh(d i j ) = Sh(d j i ) ) . Alt hough it is suggeste d in [4],
and later in [16], to use a smaller subpopulation (of O(N )) for calcu lat ing
t he cumulat ive proximity, no simulat ion result s hav e been reported to the
best of our kn owledge. In this sect ion, we investigate the effect of using a
fr act ional population for calculat ing the cumulat ive proximity measur e.
For each ind ividual in t he pop ulation , a small subpopulation P1) of size TJ
(sh ar e size) is chosen at random and sharing function values are calculat ed
for each indi vidu al in t his sub populat ion . Sharing st rat egy remain s the same
as b efore except equat ion (7) is now replaced by the following:
1)

m; = L Sh(d i j ). (11)
j =l
j E P'1

It is obvious th at if TJ is equal t o the population size, equa t ion (11) is identi cal
to equat ion (7).
In order to st udy the effect of TJ in the performance of sharing GAs, we
apply both binary- cod ed and real-coded GAs t o t est functi ons MM1 t hrough
MM5 wit h different values of TJ . T he dist ributi on index n is kept t he same
as t hat used in earlier simulat ions (i.e ., wit h TJ = N ). Fi gur e 6 shows t he
var iation of the average deviati on measur e W of GAs in generat ions 401 to 500
wit h TJ (which is also the percentage measur e of sha re size to t he p opulati on

22.5
CD MM1
3 20.0
MM2
<n
~ 17.5 ... MM3
::::;: MM4
6 15 .0 MM5
~ 12 .5
>
CD
o 10.0
CD
:ir 7.5
CD
>
-c

0.0 L......~~-'-'---'_~~~~~~---'~~~

o 10 20 30 40 50 60 70 80 90 100
Share Size

F igure 6: Average deviation measure (from generat ions 401 t o 500)

of rea l-cod ed GA s is plotted versu s share size TJ .
Studies on Multimo dal and Multiobjective Problems 443

size). F igure 6 reveals t hat beyond a crit ical value of n (much smaller than
the comp lete popula tion size) , t he pe rformance of G As are similar to t hat of
'f) = N . For example, t his crit ical value is only five percent of the po pulat ion
size in t he case of function MM 1, 15 percent in the case of fun ct ion MM2,
and t en pe rcent for MM3 , MM4 , an d MM5 .
Sim ilar performan ce is obse rved with bin ar y-cod ed GAs in [17]. These
expe rimental result s show that only a small fract ion of t he p opulation size is
sufficient to calculate t he cumulative proximity for prope rly dist rib u ting t he
population to all opt imal solutions .

4. Multiobj ective problems using nondominated sorting genetic

algorithms
In a multi objec tive opt imizat ion problem , t here is more than on e ob ject ive
fun ct ion , all of which are to be optimized simultaneously. Tr ad it ionally, the
pr act ice is to convert multiple ob ject ives int o one ob jective functi on (usually
a weighted average of t he ob ject ives is used) and t hen treat the pr oblem
as a single ob ject ive opt imizat ion problem. Unfort unately, this tech nique
is subjec tive to the user , with the opt imal solut ion being depe ndent on the
chosen weight vect or . In fact , t he solut ions of the mult iob jecti ve op ti mization
prob lem can be thought of as a collect ion of optimal solut ions obtained by
solving different single ob ject ive fun cti ons t hat are formed by using different
weight vectors . T hese solutions are kno wn as Pareto -optimal solut ions .
In or der to find a number of Paret o-op timal solutions, different exte nsions
of binar y-cod ed GAs have b een tried (e.g., [9, 10, 11]). Because of their
population ap proach , GAs are idle candidates to solve this problem . In one
implement at ion of binary-coded GA s , the concep t of non dominat ed sorti ng of
populat ion members is success fully used t o solve some t est problems [11] and
a number of t russ-structure optimization problems [12]. We briefly describ e
t hat method and show how the sa me conce pt can be used in real-coded GAs
wit h t he SBX operator.
GAs require only one fitness value for an individ ual solution in t he pop-
ulati on . T hus, an artificial fitness value mu st b e assigned t o each solutio n in
t he populat ion depe nding on the comparat ive values of each obj ective func-
tion. In order to assign a fitn ess measur e t o each solution , in [11] the idea
of non dom inat ion among pop ulation m emb ers from [13] is used . In a popu-
lation , t he nondominat ed solut ions are defined as t hose solut ions which are
bet t er in at least one object ive than any ot her solu tion in the population.
In order to implement such a non dominat ed sort ing conce pt, the following
pro cedure is adopte d .
T he population is sorte d t o find the nondominati on set of solut ions . All
individuals in t his subpo p ula t ion are ass igne d a large artificial fitness
value .
Since t he ob ject ive is to find a number of Pareto -opti mal solutions ,
shar ing is p erformed among t hese nondominat ed solut ions and a new
shared fitness is calculate d for each .
444 Kalyanmoy Deb and Amarendra Kumar

These solut ions are temporar ily counted out of t he populat ion and
the next nondomina ted set is found . These solutions are ass igned an
artificial fitn ess value mar gin ally smaller t han the least shared fitness
value in the pr evious nondominat ed set. This is don e to impose a high er
pr eference for solutions in t he pr evious set than for the cur rent set .

Sharing is perform ed aga in among the new nondominated set and this
proc ess continues until all population membe rs are ran ked in descend -
ing order of t he nondominated sets.

Thereafter , t he reproduction op eration is perform ed usin g these art ifi-

cial fitness values.

Crossove r an d mu t ati on ope rators are applied as usu al.

The preceding procedure is implemented with real-coded GA s and sharing is

p erformed phenotypi cally. The performance of real-cod ed GA s is compared
wit h binary-cod ed GAs in solving t hre e test functions used in an earlier
study [11].

4 .1 Simulat ion r e s ult s

All of the test fun ctions have two objective fun ctions to be minimized , al-
tho ugh t he pr eceding pr ocedure can be used for mo re t han two obj ecti ves.
Mor eover , t he nondominat ed sorting GA s do not restrict the ob jectives to be
of a minimizati on ty pe only. A combination of minimizati on and maximiza-
tion obj ecti ve fun cti ons can also be handled equally efficient ly. This feature
of nondominated sort ing GA s makes them attract ive in solving mult iobj ec-
t ive problems . In each of t he following three t est problems, t he ob jecti ve is
to find mult ipl e Pareto-optim al solutions.

MOl
{ mmmuze x 2
mmnrnze (x - 2)2
Par et o solution :
0 :::; x:::; 2.

-x if x:::; 1
- 2 + x if 1 < x :::; 3
minimize
M02 4- x if 3 < x :::; 4 Par et o solut ion :
1 :::; x :::; 2 and 4 < x :::; 5.
- 4 +x if x > 4
minim ize (x - 5)2

mimrru ze 2 + (X I 2)2 + (X2

M03
{ minimize 9 XI -
-

(X2 - 1)2
- 1)2
Par et o solution :
Xl = - 2.5.
Studies on Multimodal and Multiobj ective Problems 445

14 -
13
12 Binar y GA
11 Real GA (n:30)
(])
:5 10
~ 9
~ 8
c: 7
o
~ 6
's
(])
o

50 100 1 50 200 250 300 350 40 0 450 500

Generation Number

Figur e 7: The deviat ion measures of real-coded GAs and binary-coded

GAs are plotted versus generation numb er for mult iobjective prob lem
MOL
14 -
13
12 BlnaryGA
11 Real GA (n- 30)
(])
s 10
~ 9
~ 8
c: 7
o
~ 6
0:; 5
(])
o 4

1
~t~~MtJlM
O L.....~~~~~~~~~~~~~~~-"-----'

o 50 10 0 1 50 200 250 300 35 0 40 0 4 50 500

Generation Number

Figure 8: Th e deviati on measures of real-coded GAs and binary-coded

GAs are plotted versus generation numb er for th e multiobj ective prob-
lem M02 .

The sa me GA p ar ameters used in the multimodal study are chosen here.

The chi-squar e like deviation measure defined in equat ion (10) is used to
judge the working of the algorit hms. For the t est fun ctions MOl and M02 ,
the dev iation me asures are computed for ten equal int ervals (q = 10) in t he
P areto-optim al set . The p erforman ce of b oth real-coded GAs with SBX and
binary-coded GAs on problem s MOl and M02 are shown in Figures 7 and
8, resp ectively. From bo th figur es it ca n b e observed that the p erformance
of real-cod ed GAs is simi lar to that of binar y-coded GAs.
T he rea l-coded GA has also b een able to find a number of mult iple Pareto-
optimal solut ions in M03 . As shown in [11], the P ar eto-optim al set for this
problem is the line Xl = -2.5. Figure 9 shows that all p opulation members,
even in generation 500, are distributed along the Pareto-optimal line. Simi lar
resu lt s were found for binary-coded GA s in [11].
446 Kalyanm oy Deb and Amarendra Kumar

20

15

10
,.I:
N
'<
0 ......sP..';..
-5 t
-10 ~
~
-15

-20
-20 -15 -10 -5 10 15 20
x,
F igur e 9: T he pop ula tion of solut ions a t gene ration 500 obtaine d using
t he real-cod ed GA s is shown t o have found t he P aret o-opt imal line in
pro blem M03.

T hese simulation results suggest t hat , similar t o t he nondominated sort-

ing binar y-cod ed GAs, t he nondom inat ed sort ing principle can b e used in
real-cod ed GAs equa lly efficient ly. In order to fur ther test the efficacy of t he
above t echnique, the real-coded GAs with sha ring and nondomin at ed sort ing
are applied to the engineering design of a welded beam .

5. Welded b eam p roblem

A beam of rect an gul ar cross-section is welded to a memb er to carry a certain

load (F igur e 10). T he welded beam pr oblem is a popu lar engineering design
problem where t he object ive is t o find a set of four vari ab les (h, f.., t , and b)
such that t he cost of fabricat ion of t he welded beam is minimum , subject t o
sat isfying a numb er of const raints [14]. In t he origina l welded beam pr oblem
there are five cons traints .

1. T he bending stress anyw here in t he welded b eam is limi ted to the

allowable st rengt h of t he beam material.

2. The shear st ress in the weld is lim it ed to t he allowable st rengt h of the

weld.

3. Maximum buckling load that can b e carr ied by t he plat e is caused only
by the applied load .

4. T he deflect ion of the end of t he beam is limit ed t o a maximum permis-

sible deflect ion.

5. T he weld size must be smaller than t he thickness of the b eam .

St udies on Multim odal an d Multiobjective Problems 447

F
I
I
I t
~ . ~11+-_
I
Figure 10: T he welded beam problem.

T he sing le-ob ject ive, non linear optimization problem can be written as fol-
lows:

Minimize cost sub ject to : 91 (X) = Syt - O" (x ) 2: 0,

92(X) = S sy - T(X) 2: 0,
93(X) = Pc(x) - F 2: 0, (12)
94(X) = o max - o(x) 2: 0 ,
9s (X) = b - h 2: o.
For brevi ty, the expressions for the pr ecedin g constraint s are not given here.
T hey can be found elsewh ere [14, 15J.

5.1 Multiobj ective welded b eam de sign

We convert constraint 94(X) to an addit ional objective fun ct ion o(x). Thus ,
the welded b eam pr oblem now has t he two object ives of bo th mi nimi zing
cost and end-deflect ion of t he beam. T here are now four const raints (91, 92,
93, and 9s) . We use the nondominat ed sor ting GAs discussed in sec tion 4 to
solve t his problem .
At first , the variables are all ass umed to take cont inuous values. T hus,
we use nond ominat ed sorting real-cod ed GAs . A po p ulation size of 100 , a
crossover probability of 1.0, and a mu t ation pr obabi lity of 0.0 are used . A
stochastic rem ain der roulette-wheel select ion scheme is used. A distributi on
index n of 30 is used. The P ar eto-op ti mal solutions foun d aft er 500 gen-
erations are prese nted in Figur e 11. Alt ho ugh the Par eto-optim al solutions
are found wit hin t he first 50 generations , th e simulat ion is prolonged for 500
generat ions to det ermine whet her t he GA can main t ai n a stable set of solu-
ti ons long afte r t hey are ini ti ally discovered. Figur e 11 shows that the GA
has been able to find man y Par eto -optim al solut ions in one simula tion run.
T he solution set cont ains a design wit h a low cost of abo ut $3.90 , having
a deflection of abo ut 0.005 inches . The corr esponding design var iables are
h = 0.423, . = 2.457, t = 9.982 , and b = 0.433 inches . The solution set also
contains a cost ly design wit h a cost of about $40, having a low deflect ion
of only 0.0005 inches. T he correspond ing design var iab les are h = 0.426 ,
. = 2.466, t = 9.981, and b = 4.921 inches. The ot her solut ions in the solu-
t ion set reveal t hat the P areto-opt im al solutions vary only in t he value of the
varia ble b. However , neit her of t he two solut ions ment ioned above (or any
450 Kalyanm oy Deb and A marendra K umar

The simula ti on resu lt s show promise in the use of these t echniques to ot her
similar opt imization problems.
The successful applica ti on of real-coded GA s wit h t he SBX ope rator on
continuous variab les in this st udy and in [3] suggests the development of a
combined approach t o genet ic adap tive search using both binar y-cod ed GA s
and real-coded GA s. Oft en , optimization problems in engineering and sci-
ences involve mixed varia bles where some are discrete variables including
zero-one var iab les and some are cont inuous var iables. In t he combined GA
approach, a mixed coding representing discret e and continuous variables may
be used . The bin ary -coded GAs may be used to handle discret e variables and
real-coded GA s may be used to handle con tinuous variables. Althoug h any
standard repr oducti on scheme can be used , the recombinat ion op erators will
dep end on the underlyin g varia ble. T he binary single-p oint crossover or the
simulated binar y crossover may be used depending on whet her t he var iable
being crossed is discrete or cont inuous, resp ecti vely. A similar considerat ion
can b e m ad e wit h the mut ati on ope rator. Su ch a GA will allow only the
feasible values of the vari ables t o be searche d, thereby reducing t he search
effort required t o find the op ti mal solu tion . T his GA will const it ute a ro-
bust and flexible search technique that can be used t o solve mixed-variab le
optim ization p roblems effectively [3] .

Acknowle dgment
This work is funded by t he Depar tment of Scien ce and Techno logy, New Delhi
under grant SR /SY/E-06/93 . The aut hors acknowledge the programming
help of Ram Bhush an Agrawal.

Appendix
D e v ia t io n measure plots fo r test functions MM2 through MM5

Q)
Binary GA
~ 6 RealGA
'"
Q)
::; 5
Binary GA with Mating Restriction
c
.Q
OJ 4
"o
Q)
3
Q)
rn
~ 2 -
:> - - - - - -- - - - - - - - - - - - - - - - - - - - -- --
-c 1 -

o L-------~----~--~--'----'
o 50 1 00 150 200 2 50 300 350 400 450 500
Dist ribut ion Index, n

F igur e 13: Average deviation meas ure (for generations 101 to 200) is
plotted wit h distributi on index n for fun ct ion MM 2.
S t udies on Multim odal and Mu ltiobjective Prob lems 451

12 r-~~~~~~~r-~r-r-r-r-r-~~~~----,
11
10 Binary GA
Real GA (n:35 )
'" 9 Real GA (n=200 )
~ 8
m
::;;
7
c:
.Q
0;
.~ 4
o

20 40 60 80 100 12 0 140 160 180 200

Ge ne ra tio n Numb er

F igure 14: Devia tion measure is plot ted wit h generation numb er for
funct ion MM2 .

10
Binary GA
'"
:; RealGA
'"
'" Binary GA with Mating Restrict ion
'"
::;; 7
c
.Q
0;
's
o'"
'"
'"
'Q";
>
-c

o L.::=~=~~==::::=:=;:::::::::8
o 50 100 1 50 200 250 300 350 400 450 500
Distri but ion Inde x n

F igur e 15: Aver age dev ia tion measur e (for gener a tions 101 to 200) IS
plot t ed wit h distrib ut ion index n for function MM 3.

13
12
11 Binary GA
10 Real GA (n=60 )
'"
:; 9 Real GA (n 200)
2

''""
'c:"
::;;
8
7
.Q 6
0;
5
's
Cl>
0 4
3
2
1
0
0 20 40 60 80 100 120 140 160 180 20 0
Ge nera t ion Numb er

F igure 16: Deviati on measure is plot t ed wit h generation number for

fun ct ion MM 3. I
452 Kalyanmoy Deb and A marendra K umar

Binary GA
~ 7
Real GA
ii: Binary GA with Mating Restriction
'"
Q)
::;:
.2 5 -
;;;
.;; 4 -
Q)
o
Q) 3
Ol

'"i>ii 2

o L..,~~~~~~~~~~~~~~~.o.-J
o 50 100 150 200 250 300 350 400 450 500
Distribution Index n

Figure 17: Average deviation measure (for generations 101 to 200) is

plotted with distrib ution ind ex n for function MM 4.

12
11
10 Binary GA
9 Real GA (n=35)
Q)
Real GA (n=200)

"
''""
Q)
::;:
8
7
c; 6
0
~ 5
.S;
Q) 4
0
3
2
1
0
0 20 40 60 80 100 120 140 160 180 200
Ge ne ra tio n Num be r

Figure 18: Deviation measure is plotted wit h gen eration number for
function MM4 .

Q)

"'"'"
Q)
::;:
Binar y GA
Real GA
Binar y GA with Mating Restriction
c
.Q
~ 5
>
Q)
o
~ 4 -
'>"
iii
3

2 L..,~_~~~~~lLc.-~~.!..-.L~....L~~--.J
o 50 100 150 200 250 300 350 400 450 500
Dist ributi on Index n

F igure 19: Aver age deviation measure (for gen erations 101 to 200) is
plotted with distribution index n for fun ction MM5 .
Studies on Mu ltim odal and Mu ltiobjective Problem s 453

13 ,------.~ ~ ~~ ~~~ ~~~ ~~ ~ ~~~____,

12
11 Binary GA
'" 10 Real GA (n-35)
~ 9 Real GA (n-200)
:g 8
::;:
c: 7
o
:; 6
.~ 5
o

20 40 60 80 100 120 140 1 60 180 200

Gen eration Number

Figure 20: Deviati on measure is plot t ed wit h gene ration number for
fun ct ion MM5 .

R eferences
[1] Eshelm an , L. J. and Schaffer , J. D. , "Real-coded Genet ic Algorit hm s and
Interval Schemata," in Foun dations of Geneti c Algorithm s, II, edite d by
D . Whitl ey (Morgan K au fman n , San Mateo, CA , 1993) .

[2] Wr igh t , A., "Genet ic Algorithms for Real P aram eter Op timization ," in Foun-
dations of Gen etic A lgorithm s, edit ed by G. J. E. Rawlins (Morgan K aufmann ,
San Mateo, CA, 1991).

[3] Deb , K. and Agr awal, R , "Sim ulat ed Binary Crossover for Cont inu ous Search
Space," Complex Sys tems , 9 (1995) 115-148.

[4] Goldb erg , D . E ., and R ichardson , J ., "Genet ic Algorit hms with Sharing
for Mult imo dal Funct ion Optimization ," in Proceedings of the Second In-
ternational Conference on Gen etic A lgorithms, ed it ed by J . J . Gr efenstet te
(Law re nce Erl ba um Associat es, Hillsd ale, NJ , 1987) .

[5] Deb , K. , "Gene t ic Algori thms in Mul ti modal Fu nction Opt imizat ion," (Uni -
vers ity of Ala bama, Tu scaloosa, T CGA Rep ort Number 89002) .

[6] Goldberg, D . E. , Deb , K. , a nd Horn , J. , "Massive Mult imo dality , Deception ,

and Genetic Algor it hms," in Parallel Problem S olving from Nature II, edite d
by R Manner and B. Manderick (Nort h-Holland , Amsterdam, 1992) .

[7] De J ong, K. A., "An An alysis of t he Behavior of a Class of Genetic Adap-

tive Sys te ms ," (doctoral dissertation , Univers ity of Michigan , Ann Arbor) .
Dissertat ion A bstracts Intern ational, 36 (1975) 5140B (University Microfilms
Number 76-9381) .

[8] Sm it h , R E. , Forr est , S., and P erelson , A. S. , "Searching for Diverse, Co-
ope rative P opula tions wit h Gene t ic Algorithms ," (Uni ver sit y of Alabama,
Tu scaloosa , T CGA Rep or t N umb er 92002, 1992) .
454 Kalyanmoy Deb and Amarendra K umar

[9] Fonseca , C . M. and F leming P. J ., "Ge net ic Algorit hms for Mult iobject ive
Opti mi zati on : Formulat ion , Discussion , and Generalization ," in Proceedings
of the Fift h Intern ation al Confe rence on Genetic A lgorit hm s, edite d by S. For -
rest (Morgan K aufmann , San Mateo , CA, 1993).

[10] Horn, J . and Nafpliot is, N ., "Mult iob jec tive Opt imization using Niched P ar et o
Genet ic Algorit hms ," (Un iversit y of Illinois at Urban a-C hampaign, Illi GAL
Report Number 93005, 1993) .

[11] Srinivas, N . and Deb , K , "Mult iobjec t ive Funct ion Op timization using Non-
d omin at ed Sorting Gen etic Algorithms," Evo lution ary Comp uta tion, 2 (1995)
221-248.

[12] Sr inivas , N., Mult iobjective Optim izat ion using Nondomin ated S orting in
Gen etic Algorit hm s (mas te r 's t hesis, Indian Inst itute of Technology, Ka npur,
1994) .

[13] Goldberg, D . E. , Gen etic A lgorithms in S earch, Optimization, and Machin e

Learning (Addison-Wesley, Reading , 1989).

[14] Reklaitis, G. V ., Ravindra n , A ., and R agsdell, K M., En gin eering

Optimi zation-Meth ods and App lications (W iley, New York , 1983) .

[15] Deb , K , "Optimal Design of a Welded Beam Structure via Geneti c Algo-
ri thms," AIA A Journ al, 29 (1991) 2013-2015.

[16] Oei, C . K , Goldberg, D. E ., and Chang, S-J ., "Tourn ament Select ion , Nich-
ing, and the Preservat ion of Diversity," (Univers ity of Illinois at Urbana-
Champaign , IlliGAL Rep ort Num be r 91011, 1991) .

[17] K umar , A., "Mult imod al a nd Mult iob ject ive Optimiza tion using Real-cod ed
Genetic Algorit hms ," (mast er 's thes is, Indi an Instit ute of Techno logy, 1996 ).

[18] Deb , K , "GeneAS: A Rob ust Optimal Design Techn iqu e for Mechanical
Comp onent Design ," in Evolut ionary Algorithms in En gin eering App lications,
edited by D . D asgup t a and Z. Michalewicz, (Sp ringer-Verlag, New York , in
press) .