See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220742263

Self-adaptive simulated binary crossover for real-parameter optimization

Conference Paper · January 2007

DOI: 10.1145/1276958.1277190 · Source: DBLP

CITATIONS READS
82 2,834

3 authors:

Kalyan Deb Karthik Sindhya

Michigan State University University of Jyväskylä
421 PUBLICATIONS   76,611 CITATIONS    44 PUBLICATIONS   1,580 CITATIONS   

SEE PROFILE SEE PROFILE

Tatsuya Okabe
Honda R&D Innovation Lab Tokyo
22 PUBLICATIONS   1,248 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

SUMSOR project View project

MultiForest - Management for multifunctionality in European forests in the era of bioeconomy View project

All content following this page was uploaded by Karthik Sindhya on 26 May 2014.

The user has requested enhancement of the downloaded file.

 Self-Adaptive Simulated Binary Crossover
for Real-Parameter Optimization

Kalyanmoy Deb S. Karthik Tatsuya Okabe

Dept. of Mechanical Dept. of Mechanical Honda Research Institute
Engineering Engineering Japan
Indian Inst. of Tech. Kanpur Indian Inst. of Tech. Kanpur 8-1 Honcho, Wako-shi
Kanpur, PIN 208016, India Kanpur, PIN 208016, India Saitama, 351-0188, Japan
[email protected] [email protected] [email protected]

ABSTRACT must be blended to create two or more offspring vectors of

Simulated binary crossover (SBX) is a real-parameter re- real numbers [11, 2, 9, 7]. These recombination operators
combination operator which is commonly used in the evo- employ a non-uniform probability distribution around the
lutionary algorithm (EA) literature. The operator involves parent solutions to create an offspring solution. A theo-
a parameter which dictates the spread of offspring solutions retical study [1] attempted to find similarities among these
vis-a-vis that of the parent solutions. In all applications of operators.
SBX so far, researchers have kept a fixed value throughout In this paper, we concentrate on a particular recombina-
a simulation run. In this paper, we suggest a self-adaptive tion operator – the simulated binary crossover (SBX) op-
procedure of updating the parameter so as to allow a smooth erator [2]. The SBX operator uses two parent vectors and
navigation over the function landscape with iteration. Some apply the blending operator variable by variable to create
basic principles of classical optimization literature are uti- two offspring solutions. The operator involves a parameter,
lized for this purpose. The resulting EAs are found to called the distribution index (ηc ), which is kept fixed to a
produce remarkable and much better results compared to non-negative value throughout a simulation run. If a large
the original operator having a fixed value of the parameter. value of ηc is chosen, the resulting offspring solutions are
Studies on both single and multiple objective optimization close to the parent solutions. On the other hand, for a small
problems are made with success. value of ηc , solutions away from parents are likely to be cre-
ated. Thus, this parameter has a direct effect in controlling
the spread of offspring solutions. Since a search process of
Categories and Subject Descriptors finding the minimum solution of a function landscape largely
I.2.8 [Computing Methodologies]: Problem Solving, Con- depends on controlling the spread (or diversity) of offspring
trol Methods, and Search solutions vis-a-vis the selection pressure introduced by the
chosen selection operation, fixing the ηc parameter to an
General Terms appropriate value is an important task.
Here, we suggest a self-adaptive procedure of updating the
Algorithms
ηc parameter by using the extension-contraction concept in
a classical optimization algorithm. If the created child so-
Keywords lution is better than the participating parent solutions, the
Self-adaptation, simulated binary crossover, real-parameter child solution is extended further in the hope of creating even
optimization, recombination operator. a better solution. On the other hand, if a worse solution is
created, a contraction is performed. Either task will result
1. INTRODUCTION in a update of ηc , so that the newly-created extended or
contracted offspring solution has an identical probability of
Most real-world optimization problems involve decision
creation with an updated ηc . This modification procedure
variables which are real-valued. Despite the dedicated real-
has been applied to three single-objective and three two-
parameter EAs, such as evolution strategy, differential evo-
objective optimization problems and compared with corre-
lution etc., real-parameter GAs have gained adequate pop-
sponding GAs with a fixed ηc value. In all cases, better
ularity in the recent past. The main challenge in developing
performance of the suggested self-adaptive procedure is ob-
an efficient real-parameter GA lies in devising a recombina-
served.
tion operator in which two or more real-parameter vectors
In the remainder of the paper, we briefly describe the SBX
operator. Thereafter, we suggest the self-adaptive ηc update
procedure and show simulation results on single-objective
Permission to make digital or hard copies of all or part of this work for optimization problems. In each case, a parametric study
personal or classroom use is granted without fee provided that copies are with a parameter α is performed to find a suitable working
not made or distributed for profit or commercial advantage and that copies range of this parameter. Then, a scale-up study by vary-
bear this notice and the full citation on the first page. To copy otherwise, to ing the number of decision variables is performed to demon-
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee. strate the polynomial complexity of the suggested algorithm.
GECCO’07, July 7–11, 2007, London, England, United Kingdom. Finally, the self-adaptive update of ηc is extended to multi-
Copyright 2007 ACM 978-1-59593-697-4/07/0007 ...$5.00.

1187
objective optimization and results are discussed. Finally, tion, the offspring are calculated as follows:
conclusions from the study are made. h i
(1,t+1) (1,t) (2,t)
xi = 0.5 (1 + βq i )xi + (1 − βq i )xi , (3)
h i
(2,t+1) (1,t) (2,t)
xi = 0.5 (1 − βq i )xi + (1 + βq i )xi . (4)
2. SIMULATED BINARY CROSSOVER (SBX)
As the name suggests, the SBX operator [2, 6] simulates The SBX operator biases solutions near each parent more
the working principle of the single-point crossover opera- favorably than solutions away from the parents. Essentially,
tor on binary strings. In the above-mentioned studies, it the SBX operator has two properties:
was shown that this crossover operator respects the inter-
1. The difference between the offspring is in proportion
val schemata processing [7], in the sense that common in-
to the parent solutions.
terval schemata between the parents are preserved in the
(1,t+1)
offspring. The procedure of computing the offspring xi 2. Near-parent solutions are monotonically more likely
(2,t+1) (1,t) (2,t)
and xi from the parent solutions xi and xi is de- to be chosen as offspring than solutions distant from
scribed as follows. A spread factor βi is defined as the ratio parents.
of the absolute difference in offspring values to that of the
parents: An interesting aspect of this crossover operator is that for
˛ ˛ a fixed ηc the offspring have a spread which is proportional
˛ x(2,t+1) − x(1,t+1) ˛ to that of the parent solutions
˛ i i ˛
βi = ˛ ˛. (1) “ ” “ ”
˛ x(2,t) − x(1,t) ˛ (2,t+1) (1,t+1) (2,t) (1,t)
i i xi − xi = βq i xi − xi . (5)
First, a random number ui between 0 and 1 is created.
Thereafter, from a specified probability distribution func- In initial populations, where the solutions are randomly
(2,t) (1,t)
tion, the ordinate βq i is found so that the area under the placed, making the difference in parents ((xi − xi ))
probability curve from zero to βq i is equal to the chosen ran- large, this allows to create a child solution which is also far
dom number ui . The probability distribution used to create away from the parents. However, when the solutions tend
an offspring is derived to have a similar search power to that to converge due to the action of genetic operators (thereby
in a single-point crossover in binary-coded GAs and is given making the parent difference small), distant solutions are
as follows [2]: not likely to occur, thereby focusing the search to a nar-
row region. As we discuss in the following section that this
( aspect of self-adaptive nature of SBX operator is not ade-
0.5(ηc + 1)βiηc , if βi ≤ 1;
P(βi ) = (2) quate alone in reaching near the optimum in large-sized and
0.5(ηc + 1) ηc1+2 , otherwise.
βi complex functions.

Figure 1 shows the above probability distribution with ηc = 3. SELF-ADAPTIVE SBX

2 and 5 for creating offspring from two parent solutions The SBX operator discussed above involves a parameter:
(1,t) (2,t)
(xi = 2.0 and xi = 5.0). In the above expressions, the distribution index ηc . In most applications of SBX, a
the distribution index ηc is any non-negative real number. fixed value of ηc = 2 is used for single-objective optimization
A large value of ηc gives a higher probability for creating [2]. For a fixed value of ηc , the difference between the created
‘near-parent’ solutions (thereby allowing a focussed search) child solution and the closest parent solution depends on
and a small value of ηc allows distant solutions to be selected the net difference between the two parent solutions, thereby
as offspring (thereby allowing to make diverse search). causing a self-adaptive property of such a operator [5]. It
has always been a research issue whether such a self-adaptive
1
property is adequate in solving difficult optimization prob-
Probability density per offspring

ηc = 2
lems. Past studies of real-coded genetic algorithms with the
0.9
ηc = 5 SBX operator was not found to be suitable for multi-modal
0.8
problems, such as Rastrigin’s function [4]. Here, we sug-
0.7
gest a procedure for updating the ηc parameter adaptively
0.6 to solve such problems.
0.5 To illustrate the modified procedure, let us consider Fig-
0.4 ure 2, in which a typical child solution c is shown to be
0.3 created from two parent solutions p1 and p2 by using SBX
0.2 with a distribution index of ηc . Say, the random number
0.1
used for this particular crossover operation is u (∈ (0, 1)).
o o
Then, if β (> 1) corresponds to the spread factor for the
0
−1 0 1 2 3 4 5 6 7 8 specific case of crossover, we obtain the following from the
Offspring solution definition of spread factor:
2(c − p2 )
β =1+ . (6)
Figure 1: The probability density function for cre- p2 − p1
ating offspring under an SBX-ηc operator. By noting that
Z β
0.5(ηc + 1)
dβ = (u − 0.5),
After obtaining βq i from the above probability distribu- 1 β ηc +2

1188
 000000000
111111111
111111111
000000000 is obtained by the above equation, we set it to be zero and
ηc 000000000
111111111
000000000
111111111
ηc
000000000
111111111
000000000
111111111 if ηc > 50 is used, we set it to be 50. This equation gives
000000000
111111111
000000000
111111111 us an update procedure of ηc for providing a distribution
000000000
111111111
000000000
111111111
000000000
111111111 index which is able to create children solutions in the right
000000000
111111111
000000000
111111111
000000000
111111111 direction away from the parents. It is interesting to note that
000000000
111111111
u u
000000000
111111111
000000000
111111111 if a child c was created to the left of p1 , a similar update
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111 η’c relationship will also be achieved.
000000000
111111111
000000000
111111111 If the child solution c is worse than both parent solutions,
000000000
111111111
p1 p2 c c’ we would like to move the modified child c to get closer to
the parents and we may use 1/α instead of α in equation 6:
(ηc + 1) log β
Figure 2: A schematic diagram showing the ηc up- ηc = −1 + . (10)
date procedure. log (1 + (β − 1)/α)
However, if a child solution is created inside the region
bounded by both parents, a different update relationship will
we obtain the following relationship:
„ « be obtained, since the SBX probability distribution function
log 2(1 − u) is different in this case. By following the above arguments,
ηc = − 1 + . (7) we obtain β  = β α and the relationship for an improved child
log β
solution between ηc and ηc values is obtained as follows:
We now get into four scenarios (resulting in equations 9 to
1 + ηc
12), as discussed below. ηc = − 1. (11)
If this child solution c is better than both parent solu- α
tions, we assume that the region in which the child solution Once again, we restrict its value within [0, 50] to avoid any
is created is better than the region in which the parents so- computational error and to make ηc meaningful.
lutions are and intend to extend the child solution further For a scenario of creating a worse child, the above update
away from the closest parent solution (p2 , in the case shown relationship changes to
in the figure) by a factor α (> 1). This idea is similar to
that followed in Nelder and Meade’s simplex search method ηc = α(1 + ηc ) − 1. (12)
[10], in which an intermediate solution is either expanded If the child solution c has a function value which is within
or contracted or kept the same depending on the function the function value of the two parent solutions, we set ηc =
value of the solution compared to that of the previously com- ηc .
puted solutions. Rechenberg’s 1/5th rule also keeps track of To implement the update concept, we assign a ηc value
proportion of successful mutations over a certain number of within [ηcL , ηcU ] in the initial population. In all simulations
trials [13]. If the success happens too often, the mutation here, we use ηcL = ηcU = 2. For a child solution, one random
strength is increased to create solutions away from the par- number is created and a corresponding β is computed. This
ent, else the mutation strength is reduced. By using α > 1, β is used for all n variables and a child is created by variable-
the difference between the new child solution c with closest wise application of the SBX operator. This results in a line-
parent p2 is made α times than that between original child SBX operator which produces a child along the line joining
solution c and p2 , such that the two parent vectors. Thereafter, depending on the event
(c − p2 ) = α(c − p2 ). of whether a better child (than both parents) or a worse
child (than both parents) is created, ηc is updated and is
We then modify the distribution index to a value ηc such assigned to the corresponding child solution. We assign a
that with this value and having the identical random number new ηc value to each child solution separately.
u, the modified child solution c gets created from parent It is interesting to note that if α = 1 is used, all the
solutions p1 and p2 , thereby yielding above update of ηc procedure result in ηc = ηc and the
„ « above procedure is identical to the fixed ηc (original SBX)
log 2(1 − u)
ηc = − 1 + . (8) procedure.
log β 
Here, β  is the corresponding distribution index, given by 3.1 Effect of mutation operator
 In addition to the above modified SBX operator, a mu-
2(c − p2 )
β = 1+ , tation operator can be used to perturb the created child c .
(p2 − p1 ) In such an event, there is an extra solution evaluation per
= 1 + α(β − 1). child creation. Solution c gets evaluated during the crossover
Using the above relationship between β  and β and using operator and the mutated version of the modified solution
equations 7 and 8, we obtain the following update relation- c is evaluated. In our implementation, we add such extra
ship for a child solution being found to be better than the function evaluations in the computation of performance of
nearest parent solution and the child lies outside the region the modified procedure. However, if the mutation probabil-
bounded by parents: ity is so low that no variable gets mutated by the mutation
operator, there is no extra evaluation recorded.
(ηc + 1) log β
ηc = −1 + . (9)
log (1 + α(β − 1)) 4. SIMULATION RESULTS
To make the ηc
value meaningful and free from numerical We now apply the modified approach to three different un-
underflow error, we restrict it within [0, 50], that is if ηc < 0 constrained functions which are popularly used in the GA

1189
literature. Many studies in the literature on the chosen test out of 11 runs. It is clear that compared to the fixed-ηc
problems use a population which is initialized symmetrically schemes, the self-adaptive scheme is able to steer the search
around the global minimum of the functions. When an algo- towards the true minimum solution quickly and converge
rithm with such a population uses recombination and muta- close to the minimum. It is clear that even with 300,000
tion operators, which have a tendency to create solutions in function evaluations, the fixed ηc scheme is not able to find
the central region of the search space bounded by the pop- a near-optimum solution, whereas with a median of 184,050
ulation members, it finds an easier time converging to the function evaluations, a solution with three decimal places
global minimum. To avoid any such undue advantage from accuracy from the true optimum is found repeatedly in 11
the operators, in all simulations here, the initial population runs.
is bounded in the range xi ∈ [10, 15] for all i, such that the
global minimum is not bracketed in the initial population. 4.1.1 Parametric Study
This provides a stiff test to an algorithm to first get out of Figure 4 does a parametric study with α in the range
the region bounded by the initial population and then keep [1.05, 2.00]. Recall that the parameter α signifies the extent
moving in the correct direction so as to reach the global of change in the offspring solution performed to recompute
minimum. the ηc parameter for an identical probability event of creat-
ing the modified solution. Once again, 11 runs are performed
4.1 Sphere Function for every case and the best, median and worst number of
The sphere function is the simplest of the three functions function evaluations are shown in the figure. It can be said
used in this study: that the effect of α is not significant. Although there is a
X
n degradation of the median performance with increasing α,
f (x) = x2i . (13) the performance is best in the range α ∈ [1.05, 1.50]. Thus,
i=1 despite the introduction of a new parameter (α) for updat-
First, we employ the real-coded GA with the original fixed- ing an existing parameter (ηc ), the effect of the parameter
ηc based SBX operator on the 30-variable sphere function α is not significant.
with following parameter settings: population size = 150,
pc = 0.9, ηc = 2, pm = 1/30, and ηm = 50. A run is
4.1.2 Scale-up Study
terminated when a solution having a function value equal Finally, we perform a scalability study by varying the
to 0.001 is found. Eleven runs are made from random ini- number of variables (n) from 20 to 200. In this study, we
tial populations. Figure 3 shows a typical variation of best have used the following update of parameter due to the in-
population function value with the number of function eval- crease in number of variables: population size = 5n. All
uations. To investigate the effect of recombination operator other parameters are kept the same as before and we use
alone, we make another set of runs with pm = 0 and a typ- α = 1.50. Figure 5 shows that (i) the real-coded GA with
ical performance is also shown in the figure. Both results self-adaptive recombination operator is able to find a solu-
indicate that the fixed-ηc based GA is unable to find the tion close to the true minimum (within a tolerance of 0.001
true optimum in any reasonable number of evaluations for in the function value) and (ii) the increase in number of func-
a 30-variable sphere function. However, it is worthwhile to tion evaluations is polynomial to the increase in number of
mention here that after a variable-wise creation of two off- variables (O(n2.21 )). The GA with the self-adaptive SBX
spring vectors, if the variables values are swapped randomly operator suggested here does not exploit the variable sepa-
between the two offsprings (similar to a uniform crossover rability of the objective functions. This is the reason why
operator or the crossover operator in differential evolution the proposed GA takes more number of evaluations than
[12]), a much quicker convergence with a fixed ηc = 2 can other approaches which favor the variable separability and
be obtained for variable-wise separable functions, like the unimodality of the sphere function [5, 4].
sphere function [5]. This is because variables can indepen-
dently reach near the true optimum in different population 4.2 Rosenbrock Function
members and a uniform swapping of such solutions can cre- Next, we consider the Rosenbrock’s function:
ate a complete solution having near optimal variables values.
However, this uniform-like crossover operation may not work X
n−1
f (x) = 100(x2i − xi+1 ) + (xi − 1)2 . (14)
well in rotated and more complex problems. In this paper,
i=1
we only concentrate on the blending part of the operation
and investigate whether the fixed nature of ηc or the self- This function has the global minimum at xi = 1 for all i with
adaptive nature of ηc is more effective without having any a function value equal to zero. Near the minimum region,
crossover-like swapping effect. this function has a very small slope towards the minimum.
Next, we employ our self-adaptive update procedure of ηc . This property of the landscape causes an algorithm to have
All initial population members are initialized with ηc = 2 a slow convergence to the minimum.
and then allowed to change based on our update procedure Once again, we first employ the original SBX operator
described above. All other parameters are the same as above with a fixed ηc = 2 and initialize xi ∈ [10, 15] for all i.
and to investigate the effect of the self-adaptive recombina- Figure 6 shows a typical variation of population-best func-
tion operator alone, we use pm = 0 and reduce the crossover tion value with the number of function evaluations for the
probability to pc = 0.7. For the ηc update we use α = 1.50. 30-variable Rosenbrock function with following parameter
A typical variation of the best population function value settings: population size = 150, pc = 0.9, pm = 1/30, and
is shown in Figure 3 and is found to have good converging ηm = 50. The algorithm is terminated when a solution with
property. Table 1 shows the best, median and worst number a minimum function value of 0.001 is found. The procedure
of function evaluations to reach a solution with f = 0.001 is not able to come close to the true minimum. When we use

1190
 10000
1e+08
Orig. (p_c=0.9,p_m=0.033) 450000
1000
400000

Function Evaluations
100 1e+07

Function Evaluations
Function Value

Orig. (p_c=0.9,p_m=0)
350000
10
300000
1 1e+06

250000 Slope = 2.21

0.1 Self−adaptive

200000 100000
0.01

0.001 150000
0 50000 100000 150000 200000 250000 300000
Function Evaluations 10000
100000 20 30 50 80 100 200
1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
Number of Variables
α

Figure 3: Variation of population-

best function value with number Figure 4: Parametric study of α for Figure 5: Scale-up study with
of function evaluations for the 30- the 30-variable sphere function. number of variables for the sphere
variable sphere function. function.

Table 1: Performance of real-coded GAs with fixed and self-adaptive ηc update on 30-variable sphere function.

Method Optimized function value (func. eval.)

Original (pc = 0.9, pm = 0.033) 1.24e03 (300,000) 4.16e03 (300,000) 4.30e03 (300,000)
Original (pc = 0.9, pm = 0) 1.21e03 (300,000) 4.13e03 (300,000) 4.24e03 (300,000)
Self-adp.(pc = 0.7, pm = 0) 10−3 (151,800) 10−3 (184,050) 10−3 (213,450)

pm = 0, a somewhat better performance is observed, but the local optima and one global minimum at xi = 0 for i =
procedure is unable to reach near to the global minimum. 1, 2, . . . , n:
Now, we apply the self-adaptive recombination operator
starting with ηc = 2 to all initial population members. We X
n
f (x) = x2i + 10 (1 − cos(2πxi )) . (15)
use α = 1.5 for the update procedure. Figure 6 shows a
i=1
typical variation in the population-best function value. The
self-adaptive property of the recombination operator is able In an earlier study [4], this function was difficult to solve
to adjust the ηc adequately to navigate through the fitness for global optimality using the real-coded GA with a parent
landscape to reach near the global minimum. Table 2 shows centric recombination operator, particularly when the ini-
the best, median and worst performance out of 11 runs of tial population did not bracket the global optimum. In this
original and self-adaptive GAs. For the self-adaptive case, study, we initialize a population with each xi created ran-
we have shown results with α = 1.4, which produces the best domly in [10, 15], away from the global minimum solution.
result. With as many as 10 million function evaluations, the In a single dimension, an algorithm has to overcome at least
fixed ηc scheme is not able to find a near-optimum solution 10 different local optima to reach to the global minimum.
(in fact, the best solution has a function value of 6.85(106 )), With a larger variable size, exponentially more local optima
whereas with a median of about 6.9 million function evalu- must be overcome to reach the global minimum.
ations, a solution with three decimal places accuracy from First, we apply the real-coded GA with the original SBX
the true optimum is found repeatedly. operator on the 20-variable Rastrigin function with a stan-
dard parameter setting: population size = 100, pc = 0.9,
4.2.1 Parametric Study ηc = 2, pm = 1/20, ηm = 50. GAs are run from 11 different
A parametric study on α is shown in Figure 7. Although initial populations till a maximum of 40,000 generations or
there is a upward trend in number of function evaluations till a solution having a function value of 10−4 is obtained.
with increasing α, with α ∈ [1.05, 1.50] the performance is Table 3 shows the best function values obtained by the pro-
better. Interestingly, in this problem also we find that the cedure. In this case, the best obtained function value (in
effect of α in a good range of values is not significant. 11 runs) with four million function evaluations is 319.523,
whereas the globally best solution has a function value equal
4.2.2 Scale-up Study to zero. It is clear that no run is able to find a solution close
A scale-up study is made next by varying the number of to the global minimum. With pc = 0.7 and pm = 0.01, we
variables from 20 to 200. Figure 8 shows that the number obtain slightly better performance, but even now no solution
of function evaluations needed to reach up to three decimal close to the global minimum is found.
places of accuracy varies as O(n3.496 (log n)−9.147 ) by the Next, we apply the real-coded GA with our proposed self-
proposed self-adaptive procedure. adaptive SBX operator with pm = 0.7 and pm = 0.01. All
initial population members are initialized with ηc = 2 and
4.3 Rastrigin’s Function then allowed to change using the proposed ηc update proce-
Next, we consider the Rastrigin’s function which has many dure described earlier. We use α = 1.5 here. Table 3 shows

1191
 1e+08 1e+09
1.2e+07
Orig. (p_c=0.9, p_m=0.033)
1.1e+07
1e+06
1e+07

Function Evaluations
Function Evaluations
Orig. (p_c=0.9,p_m=0)
Function Value

10000 9e+06
1e+08
8e+06
100 7e+06
6e+06
1 Self−adaptive 1e+07
5e+06
4e+06
0.01
3e+06
0 2e+06 4e+06 6e+06 8e+06 1e+07
2e+06 1e+06
Function Evaluations 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 20 30 50 80 100 200
α Number of Variables

Figure 6: Variation of population-

best function value with number Figure 7: Parametric study of α for Figure 8: Scale-up study with
of function evaluations for the 30- the 30-variable Rosenbrock’s func- number of variables for the Rosen-
variable Rosenbrock’s function. tion. brock’s function.

Table 2: Performance of real-coded GAs with fixed and self-adaptive ηc update on 30-variable Rosenbrock’s
function.
Method Optimized function value (func. eval.)
Original (pc = 0.9, pm = 0.033) 6.85e06 (10M) 7.44e06 (10M) 8.30e07 (10M)
Original (pc = 0.9, pm = 0) 9.94e06 (10M) 4.94e07 (10M) 5.19e07 (10M)
Self-adp.(pc = 0.7, pm = 0) 10−3 (2,200,650) 10−3 (6,832,950) 10−3 (7,836,300)

that in all cases the self-adaptive update of ηc is able to solutions in terms of their function values and show its varia-
find a solution with desired accuracy in a fraction of total tion with number of function evaluations in Figure 9. The y-
evaluations used in the case of fixed-ηc procedures. Despite axis of this figure is made logarithmic. Although, the above
being started far away from the global minimum, the pro- ηc value seems to end at 10−5 in the figure, the actual value
cedure is able to converge to the correct globally minimum recorded by the procedure is ηc = 0. An interesting aspect is
solution. Figure 9 shows the decrease in best function value that the ηc value seems to take a value zero at various stages
with number of function evaluations for a typical simulation of the simulation. The line corresponding to the best func-
run with a fixed ηc = 2 procedure (pc = 0.7, pm = 0.01) and tion value (marked as ‘Adaptive SBX’) indicates that there
with the self-adaptive procedure. The figure clearly shows are a number of function values, especially within f = 10
that the fixed ηc run is poor in its performance, whereas to f = 1, in which the algorithm seems to get stuck for
the self-adaptive procedure steadily finds better and better a large number of evaluations before finding a better solu-
solutions with function evaluations. tion. Within this range of function values, this function has
one local minimum at every integer value of the function,
10000
thereby having a possibility to get stuck 10 times. Interest-
1000 Original SBX ingly, every time the best population member gets stuck at
100 a local minimum, the ηc of the best 15 percentile solution
gets updated to zero, thereby increasing the spread of cre-
f(x) or eta_c

10 Adaptive SBX
ated solutions by the SBX operator. Since an ηc = 0 will be
eta_c
1
the smallest possible ηc which provides the maximum spread
0.1 in created solutions, the algorithm finds that the best way
0.01 to counteract a local statis is to increase the diversity of
created solutions to the extent possible. However, as soon
0.001
as a better solution is found, the ηc is immediately updated
1e−04 to a value close to one, thereby providing a more focussed
1e−05 search around population members. With an assigned fixed
0 1e05 2e05 3e05 4e05 5e05 6e05
Function Evaluations value of ηc over the entire simulation run, such a variation
in spread in solutions in an offspring population is not possi-
ble. The proposed procedure seems to employ this principle
Figure 9: Decrease in best population function value adaptively and multiply in as many times as the algorithms
with number of function evaluations for the 20- get stuck to a locally optimal solution and improve from such
variable Rastrigin’s function. a situation. In the absence of such an adaptive update of ηc
with the original SBX operator, the corresponding GA was
To understand the effect of the proposed self-adaptive up- not able to improve its performance efficiently every time it
date of ηc , we record the ηc value of the top 15 percentile gets stuck to a locally optimal solution.

1192
Table 3: Performance of real-coded GAs with fixed and self-adaptive ηc update on 20-variable Rastrigin’s
function.
Method Optimized function value (func. eval.)
Original (pc = 0.9, pm = 0.05) 319.523 (4M) 338.093 (4M) 342.363 (4M)
Original (pc = 0.7, pm = 0.01) 124.368 (4M) 210.929 (4M) 335.380 (4M)
Self-adp.(pc = 0.7, pm = 0.01) 10−4 (287,822) 10−4 (429,511) 10−4 (569,597)

4.3.1 Parametric Study uations increased polynomially (O(n1.807 )) with n over the
A parametric study of α to investigate its effect on the entire range of number of variables used in this study. It is
performance of the proposed procedure is made next. For
the 20-variable Rastrigin’s function and with above param-
eter settings, 11 different runs are made. Figure 10 shows 1e+08

Number of Function Evaluations

number of function evaluations needed to find a solution
with a function value equal to or smaller than 10−4 . It
is interesting to note that for a large range of values of α
(∈ [1.05, 8.00]), the performance of the proposed procedure 1e+07
remains fairly independent of α. However, the best perfor-
mance seems to happen for α = 3 with best, median and
worst number of function evaluations of 221,082, 342,741, Slope=1.807
and 718,680, respectively. 1e+06

1e+06

900000
1e+05
Function Evaluations

800000 20 30 50 100 200

Number of Variables
700000

600000

500000
Figure 11: A polynomial increase in function eval-
uations with number of decision variables for the
400000 Rastrigin’s function using proposed algorithm.
300000

200000
1 2 3 4 5 6 7 8 9 10 noteworthy that the proposed procedure is able to overcome
alpha exponentially many local minima to converge to the globally
minimum solution with polynomially increasing number of
function evaluations to as complex as a 200-variable Rastri-
Figure 10: Effect of parameter α on number of func- gin’s function.
tion evaluations to obtain the global minimum with
four decimal places of accuracy for the 20-variable
Rastrigin’s function.
5. SELF-ADAPTIVE SBX FOR
4.3.2 Scale-up Study MULTI-OBJECTIVE OPTIMIZATION
Motivated by the success of the self-adaptive SBX pro- Like the way we made the SBX operator self-adaptive
cedure on the 20-variable Rastrigin’s function, we now try for single-objective optimization, we extend the idea here
to solve larger size Rastrigin’s function with n varying in for multi-objective optimization. The main difficulty arises
[20, 200]. With an increase in n, the number of local minima in deciding when a child solution is better than a parent.
in a particular range of the variable values increase exponen- Here, we simply use the idea of non-domination to decide
tially and the resulting problem is likely to provide more dif- on this matter. Let us consider Figure 12. For the two
ficulty to an optimization algorithm. Since a function with parent objective vectors shown in the figure, if a child lies on
a larger number of variables should ideally require a larger the non-dominated shaded region (marked with ‘A’), we call
population size for a GA initialized with a randomly created that the child to be better than the parents and we use the
population [8], we use the following parametric update for ηc update equations described earlier. On the other hand,
different n: population size=5n, pm = 0.2/n, and pc = 0.7. if the child lies on the region marked as ‘B’ in the figure,
Here also, we initialize population with xi ∈ [10, 15], so as we call that the child is worse than the parents and we use
to not bracket the global minimum in the initial population. the appropriate equation (described earlier) to update ηc .
We also use α = 1.5 for all n. We run till a solution with If, however, the child lies on the region marked as ‘C’ in
a function value equal to or smaller than 10−4 is obtained. the figure, we do not update ηc . The remaining part of the
Figure 11 shows that the required number of function eval- NSGA-II algorithm [3] is used as usual.

1193
 f2 111111111111111111
000000000000000000
000000000000000000
111111111111111111
000000000000000000
111111111111111111
7. CONCLUSIONS
000000000000000000
111111111111111111
000000000000000000
111111111111111111
000000000000000000
111111111111111111
000000000000000000
111111111111111111 In this paper, we have suggested a self-adaptive procedure
000000000000000000
111111111111111111
000000000000000000
111111111111111111 for updating the distribution index ηc used in the simulated
000000000000000000
111111111111111111
000000000000000000
111111111111111111
000000000000000000
111111111111111111
B
000000000000000000
111111111111111111
000000000000000000
111111111111111111 binary crossover or SBX operator which is a commonly-used
000000000000000000
111111111111111111
000000000000000000
111111111111111111 real-parameter recombination operator. The update pro-
000000000000000000
111111111111111111
000000000000000000
111111111111111111
000000000000000000
111111111111111111
000000000000000000
111111111111111111
000000000000000000
111111111111111111 cedure mimics the extension-contraction concept in Nelder
000000000000000000
111111111111111111
000000000000000000
111111111111111111 and Meade’s simplex search procedure and also follows, in
000000000000000000
111111111111111111
000000000000000000
111111111111111111
000000000000000000
111111111111111111
000000000000000000
111111111111111111
000000000000000000
111111111111111111
parents principle, Rechenberg’s 1/5-th update rule. On three single-
000000000000000000
111111111111111111
000000000000000000
111111111111111111 objective optimization problems and on three two-objective
000000000000000000
111111111111111111
000000000000000000
111111111111111111
000000000000000000
111111111111111111
0000000
1111111
0000000
1111111 optimization problems, the suggested procedure is found to
0000000
1111111 C
0000000
1111111 be perform much better than the original SBX procedure.
0000000
1111111
0000000
1111111
0000000
1111111
A
0000000
1111111 Further investigations are now needed for solving problems
0000000
1111111
0000000
1111111 f1
0000000
1111111
0000000
1111111
0000000
1111111
having a linkage among variables and problems having more
than two objectives. A similar self-adaptive idea can also be
used with other real-parameter recombination operators.
Figure 12: A sketch showing different regions in
which the child may lie for an appropriate update 8. ACKNOWLEDGMENT
of ηc .
This study is supported by a research grant from Honda
R&D, Japan.
6. SIMULATION RESULTS
Here, we show the working of NSGA-II with the self- 9. REFERENCES
adaptive SBX on three test problems: 30-variable ZDT1 [1] H.-G. Beyer and K. Deb. On the desired behavior of
and ZDT2 and 12-variable DTLZ2 problems. self-adaptive evolutionary algorithms. In Parallel Problem
Solving from Nature VI (PPSN-VI), pages 59–68, 2000.
6.1 ZDT1, ZDT2 and DTLZ2 Problems [2] K. Deb and R. B. Agrawal. Simulated binary crossover for
continuous search space. Complex Systems, 9(2):115–148,
We use the hyper-volume measure to indicate the perfor- 1995.
mance of a procedure. We compare the performance of the [3] K. Deb, S. Agrawal, A. Pratap, and T. Meyarivan. A fast
self-adaptive procedure with NSGA-II having the fixed-ηc and elitist multi-objective genetic algorithm: NSGA-II.
based SBX procedure with standard setting: ηc = 15 and IEEE Transactions on Evolutionary Computation,
ηm = 20. We use the following parameter setting for self- 6(2):182–197, 2002.
adaptive EA: population size = 100, pc = 0.9, pm = 1/30, [4] K. Deb, A. Anand, and D. Joshi. A computationally
efficient evolutionary algorithm for real-parameter
pm = 0 and maximum number of generations = 500. We optimization. Evolutionary Computation Journal,
initialize all population members with ηc = 2, as before. Ta- 10(4):371–395, 2002.
ble 4 shows the hyper-volume measure for various α values [5] K. Deb and H.-G. Beyer. Self-adaptation in real-parameter
with the self-adaptive procedure. genetic algorithms with simulated binary crossover. In
We observe that with Proceedings of the Genetic and Evolutionary Computation
α values near 1.70, Conference (GECCO-99), pages 172–179, 1999.
the performance of [6] K. Deb and A. Kumar. Real-coded genetic algorithms with
Table 4: Performance
simulated binary crossover: Studies on multi-modal and
the self-adaptive NSGA- (hyper-volume) comparison multi-objective problems. Complex Systems, 9(6):431–454,
II is better than of self-adaptive NSGA-II 1995.
the fixed-ηc based with fixed-ηc based SBX on [7] L. J. Eshelman and J. D. Schaffer. Real-coded genetic
NSGA-II. Table 4 ZDT1 and ZDT2. algorithms and interval-schemata. In Foundations of
also shows the per- Genetic Algorithms 2 (FOGA-2), pages 187–202, 1993.
formance for 30-variable [8] D. E. Goldberg, K. Deb, and J. H. Clark. Genetic
α Best Median Worst algorithms, noise, and the sizing of populations. Complex
ZDT2 problem. Here,
ZDT1 Systems, 6(4):333–362, 1992.
we observe that for Original, fixed ηc = 15 [9] T. Higuchi, S. Tsutsui, and M. Yamamura. Theoretical
α values larger than 0.72745 0.72133 0.72075 analysis of simplex crossover for real-coded genetic
1.20 the performance Self-Adaptive SBX algorithms. In Parallel Problem Solving from Nature
of the self-adaptive 1.05 0.72768 0.68531 0.62985 (PPSN-VI), pages 365–374, 2000.
NSGA-II is better. 1.20 0.73949 0.72203 0.56250 [10] J. A. Nelder and R. Mead. A simplex method for function
For the three-objective 1.50 0.72908 0.72222 0.61360 minimization. Computer Journal, 7:308–313, 1965.
1.70 0.73984 0.72228 0.72210 [11] I. Ono and S. Kobayashi. A real-coded genetic algorithm
DTLZ2 problem, self-
2.00 0.72282 0.72238 0.72196 for function optimization using unimodal normal
adaptive EA obtained ZDT2 distribution crossover. In Proceedings of the Seventh
better hyper-volume Original, fixed ηc = 15 International Conference on Genetic Algorithms
values of 0.57461, 0.38883 0.38846 0.38800 (ICGA-7), pages 246–253, 1997.
0.56975 and 0.56597, Self-Adaptive SBX [12] K. V. Price, R. Storn, and J. Lampinen. Differential
compared to 0.55993, 1.05 0.38959 0.37661 0.15350 Evolution: A Practical Approach to Global Optimization.
0.55645 and 0.55539 1.20 0.38940 0.38912 0.38874 Springer-Verlag, Berlin, 2005.
as best, median and 1.50 0.38957 0.38920 0.38891 [13] I. Rechenberg. Evolutionsstrategie: Optimierung
1.70 0.39040 0.38930 0.38890 Technischer Systeme nach Prinzipien der Biologischen
worst values obtained 2.00 0.38942 0.38922 0.38880
using original fixed- Evolution. Stuttgart: Frommann-Holzboog Verlag, 1973.
ηc EA.

1194

View publication stats