Constrained Multi-Objective Optimization Using Steady

State Genetic Algorithms

Deepti Chafekar, Jiang Xuan, Khaled Rasheed
Computer Science Department
University of Georgia
Athens, GA 30602 USA
{chafekar, xuan, khaled}@cs.uga.edu
Abstract. In this paper we propose two novel approaches for solving
constrained multi-objective optimization problems using steady state GAs.
These methods are intended for solving real-world application problems that
have many constraints and very small feasible regions. One method called
Objective Exchange Genetic Algorithm for Design Optimization (OEGADO)
runs several GAs concurrently with each GA optimizing one objective and
exchanging information about its objective with the others. The other method
called Objective Switching Genetic Algorithm for Design Optimization
(OSGADO) runs each objective sequentially with a common population for all
objectives. Empirical results in benchmark and engineering design domains are
presented. A comparison between our methods and Non-Dominated Sorting
Genetic Algorithm-II (NSGA-II) shows that our methods performed better than
NSGA-II for difficult problems and found Pareto-optimal solutions in fewer
objective evaluations. The results suggest that our methods are better applicable
for solving real-world application problems wherein the objective computation
time is large.
1 Introduction
This paper concerns the application of steady state Genetic Algorithms (GAs) in
realistic engineering design domains which usually involve simultaneous optimization
of multiple and conflicting objectives with many constraints. In these problems
instead of a single optimum there usually exists a set of trade-off solutions called the
non-dominated solutions or Pareto-optimal solutions. For such solutions no
improvement in any objective is possible without sacrificing at least one of the other
objectives. No other solutions in the search space are superior to these Pareto-optimal
solutions when all objectives are considered. The user is then responsible for choosing
a particular solution from the Pareto-optimal set later.
Some of the challenges faced in the application of GAs to engineering design domains
are:
Deepti Chafekar, Jiang Xuan, Khaled Rasheed
The search space can be very complex with many constraints and the feasible
(physically realizable) region in the search space can be very small.
Determining the quality (fitness) of each point may involve the use of a simulator
or an analysis code which takes a non-negligible amount of time. This simulation
time can range from a fraction of a second to several days in some cases. Therefore
it is impossible to be cavalier with the number of objective evaluations in an
optimization.
For such problems steady state GAs may perform better than generational GAs
because they better retain the feasible points found in their populations and may have
higher selection pressure which is desirable when evaluations are very expensive.
With good diversity maintenance, steady state GAs have done very well in several
realistic domains [1]. Significant research has yet to be done in the area of steady state
multi-objective GAs. We therefore decided to focus our research on this area.
The area of multi-objective optimization using Evolutionary Algorithms (EAs) has
been explored for a long time. The first multi-objective GA implementation called the
Vector Evaluated Genetic Algorithm (VEGA) was proposed by Schaffer in 1985 [9].
Since then, many Evolutionary algorithms for solving multi-objective optimization
problems have been developed. The most recent ones are the Non-Dominated Sorting
Genetic Algorithm-II (NSGA-II) [3], Strength Pareto Evolutionary Algorithm-II
(SPEA-II) [16], Pareto Envelope based selection-II (PESA-II) [17]. Most of these
approaches propose the use of a generational GA. Deb proposed an Elitist Steady
State Multi-objective Evolutionary Algorithm (MOEA) [18] which attempts to
maintain spread [15] while attempting to converge to the true Pareto-optimal front.
This algorithm requires sorting of the population for every new solution formed
thereby increasing its time complexity. Very high time complexity makes the Elitist
steady state MOEA impractical for some problems. To the best of our knowledge,
apart from Elitist Steady State MOEA, the area of steady state multi-objective GAs
has not been widely explored. Also constrained multi-objective optimization which is
very important for real-world application problems has not received the deserved
exposure. In this paper we propose two methods for solving constrained multi-
objective optimization using steady state GAs. These methods are relatively fast and
practical. It is also easy to transform a single-objective GA to a multi-objective GA by
using these methods.
In the first method called the Objective Exchange Genetic Algorithm for Design
Optimization (OEGADO) several single objective GAs run concurrently. Each GA
optimizes one of the objectives. At certain intervals these GAs exchange information
about their respective objectives with each other. In the second method called the
Objective Switching Genetic Algorithm for Design Optimization (OSGADO) a single
GA runs multiple objectives in a sequence switching at certain intervals between
objectives.
Our methods can be viewed as multi-objective transformations of GADO (Genetic
Algorithm for Design Optimization) [1, 2]. GADO is a GA that was designed with the
goal of being suitable for the use in engineering design. It uses new operators and
Constrained Multi-Objective Optimization Using Steady State Genetic Algorithms
search control strategies that target engineering domains. GADO has been applied in a
variety of optimization tasks which span many fields. It has demonstrated a great deal
of robustness and efficiency relative to competing methods.
In GADO, each individual in the GA population represents a parametric description
of an artifact. All parameters have continuous intervals. The fitness of each individual
is based on the sum of a proper measure of merit computed by a simulator or some
analysis code, and a penalty function if relevant. A steady state model is used, in
which several crossover and mutation operators including specific and innovative
operators like guided crossover are applied to two parents selected by linear rank
based selection. The replacement strategy used is a crowding technique, which takes
into consideration both the fitness and the proximity of the points in the GA
population. GADO monitors the degree of diversity of the GA population. If at any
stage it is discovered that the individuals in the population became very similar to one
another, the diversity maintenance module rebuilds the population using previously
evaluated points in a way that restores diversity. The diversity maintenance module in
GADO also rejects proposed points that are extremely similar to previously evaluated
points. The GA stops when either the maximum number of evaluations has been
exhausted or the population loses diversity and practically converges to a single point
in the search space. Floating point representation is used. GADO also uses some
search control strategies [2] such as a screening module which saves time by avoiding
the full evaluation of points that are unlikely to correspond to good designs.
We compared the results of our two methods with the state-of-the-art Elitist Non-
Dominated Sorting Algorithm-II (NSGA-II) [3]. NSGA-II is a non-dominated sorting
based multi-objective evolutionary algorithm with a computational complexity of
O(MN
2
) (where M is the number of objectives and N is the population size). NSGA-II
incorporates an elitist approach, a parameter-less niching approach and a simple
constraint handling strategy. Due to NSGA-IIs low computational requirements,
elitist features and constraint handling capacity, it has been successfully used in many
applications. It proved to be better than many other multi-objective optimization GAs
[3, 18].
In the remainder of the paper, we provide a brief description of our two proposed
methods. We then present results of the comparison of our methods with NSGA-II.
Finally, we conclude the paper with a discussion of the results and future work.
2 Methods for Multi-Objective Optimization Using Steady State
GAs
We propose two methods for solving constrained multi-objective optimization
problems using steady state GAs. One is the Objective Exchange Genetic Algorithm
for Design Optimization (OEGADO), and other is the Objective Switching Genetic
Algorithm for Design Optimization (OSGADO). It should be noted that for multi-
objective GAs, maintaining diversity is a key issue. However we did not need to take
Deepti Chafekar, Jiang Xuan, Khaled Rasheed
any extra measures for diversity maintenance as the diversity maintenance module
already present in GADO [1, 2] seemed to handle this issue effectively. We focused
on the case of two objectives in our experiments for simplicity of implementation and
readability of the results, but the methods are applicable for multi-objective
optimization problems with more than two objectives.
2.1 Objective Exchange Genetic Algorithm for Design Optimization
(OEGADO)
The main idea of OEGADO is to run several single objective GAs concurrently. Each
of the GAs optimizes one of the objectives. All the GAs share the same representation
and constraints, but have independent populations. They exchange information about
their respective objectives every certain number of iterations.
In our implementation, we have used the idea of informed operators (IOs) [4]. The
main idea of the IOs is to replace pure randomness in traditional GA operators with
decisions that are guided by reduced models formed using the methods presented in [5,
6, 7]. The reduced models are approximations of the fitness function, formed using
some approximation techniques, such as least squares approximation [5, 7, 8]. These
functional approximations are then used to make the GA operators such as crossover
and mutation more informed. These IOs generate multiple children [4], rank them
using the approximate fitness obtained from the reduced model and select the best.
Every single objective GA in OEGADO uses least squares to form a reduced model
of its own objective. Every GA exchanges its own reduced model with those of the
other GAs. In effect, every GA, instead of using its own reduced model, uses other
GAs reduced models to compute the approximate fitness of potential individuals.
Therefore each GA is informed about other GAs objectives. As a result each GA not
only focuses on its own objective, but also gets biased towards the objectives which
the other GAs are optimizing.
The OEGADO algorithm for two objectives looks as follows:
1. Both the GAs are run concurrently for the same number of iterations, each GA
optimizes one of the two objectives while also forming a reduced model of it.
2. At intervals equal to twice the population size, each GA exchanges its reduced
model with the other GA.
3. The conventional GA operators such as initialization (only applied in the
beginning), mutation and crossover are replaced by informed operators. The IOs
generate multiple children and use the reduced model to compute the approximate
fitness of these children. The best individual based on this approximate fitness is
selected to be the newborn. It should be noted that the approximate fitness function
used is of the other objective.
4. The true fitness function is then called to evaluate the actual fitness of the newborn
corresponding to the current objective.
5. The individual is then added to the population using the replacement strategy.
Constrained Multi-Objective Optimization Using Steady State Genetic Algorithms
6. Steps 2 through 5 are repeated till the maximum number of evaluations is
exhausted.
If all objectives have similar computational complexity, the concurrent GAs can be
synchronized, so that they exchange the current approximations at the right time. On
the other hand, when objectives vary considerably in their time complexity, the GAs
can be run asynchronously.
It should be noted that OEGADO is not really a multi-objective GA, but several
single objective GAs working concurrently to get the Pareto-optimal region. Each GA
finds its own feasible region, by evaluating its own objective. For the feasible points
found by a single GA, we need to run the simulator to evaluate the remaining
objectives. Thus for OEGADO with two objectives:
Total number of objective evaluations = Sum of objective evaluations of each GA +
Sum of the number of feasible points found by each GA
A potential advantage of this method is speed, as the concurrent GAs can run in
parallel. Therefore multiple objectives can be evaluated at the same time on different
CPUs. Also the asynchronous OEGADO works better for objectives having different
time complexities. If some objectives are fast, they are not slowed down by the slower
objectives. It should be noted that because of the exchange of reduced models, each
GA optimizes its own objective and also gives credit to the other objectives.
2.2 Objective Switching Genetic Algorithm for Design Optimization
(OSGADO)
The main idea of OSGADO is to use a single GA that optimizes multiple objectives in
a sequential order. Every objective is optimized for a certain number of evaluations,
then a switch occurs and the next objective is optimized. The population is not
changed when objectives are switched. This continues till the maximum number of
evaluations is complete.
We modified GADO [1, 2] to create multi-objective OSGADO. OSGADO is inspired
from the Vector Evaluated GA (VEGA) [9]. Schaffer (1985) proposed VEGA for
generational GAs. In VEGA the population is divided into m different parts for m diff
objectives; part i is filled with individuals that are chosen at random from current
population according to objective i. Afterwards the mating pool is shuffled and
crossover and mutation are performed as usual. Though VEGA gave encouraging
results, it suffered from bias towards the extreme regions of the Pareto-optimal curve.
The OSGADO algorithm looks as follows:
1. The GA is run initially with the first objective as the measure of merit for a certain
number of evaluations. The fitness of an individual is calculated based on its
measure of merit and the constraint violations. Selection, crossover and mutation
take place in the regular manner.
Deepti Chafekar, Jiang Xuan, Khaled Rasheed
2. After a certain numbers of evaluations, the GA is run for the next objective. When
the evaluations for the last objective are complete, the GA switches back to the
first objective.
3. Step 2 is repeated till the maximum number of evaluations is reached.
In order to fairly compare the methods, in the experiments we first ran OEGADO and
obtained the number of feasible points found by each of the two GAs. We then ran
OSGADO for the number of evaluations calculated as follows,
Total number of objective evaluations = Sum of evaluations of each objective in
OEGADO + Sum of the number of feasible points found by each objective in
OEGADO
OSGADO has certain advantages over VEGA. In VEGA every solution is evaluated
for only one of the objectives each time and therefore it can converge to individual
objective optima (the extremes of the Pareto-optimal curve) without adequately
sampling the middle section of the Pareto-optimal curve. However OSGADO
evaluates every solution using each of the objectives at different times. So OSGADO
is at less risk of converging at individual objective optima.
3 Experimental Results
In this section, we first describe the test problems used to compare the performance of
OEGADO, OSGADO and NSGA-II. We then briefly discuss the parameter settings
used. Finally, we discuss the results obtained for various test cases by these three
methods.
3.1 Test Problems
The test problems for evaluating the performance of our methods were chosen based
on significant past studies. We chose four problems from the benchmark domains
commonly used in past multi-objective GA research, and two problems from the
engineering domains. The degree of difficulty of these problems varies from fairly
simple to difficult.
The problems chosen from the benchmark domains are BNH used by Binh and Korn
[10], SRN used by Srinivas, Deb [11], TNK suggested by Tanaka [12] and OSY used
by Osyczka, Kundu [13]. The problems chosen from the engineering domains are
Two-Bar Truss Design used by Deb [14] and Welded Beam design used by Deb [14].
All these problems are constrained multi-objective problems. Table 1 shows the
variable bounds, objective functions and constraints for all these problems.
Constrained Multi-Objective Optimization Using Steady State Genetic Algorithms
Table 1: Test problems used in this study, all objective functions are to be minimized
Problem
Variable
bounds
Objectives functions f (x) and
Constraints C(x)
BNH
] 3 , 0 [
] 5 , 0 [
2
1

x
x
7 . 7 ) 3 ( ) 8 ( ) x (
25 ) 5 ( ) x (
) 5 ( ) 5 ( ) x (
4 4 ) x (
2
2
2
1 2
2
2
2
1 1
2
2
2
1 2
2
2
2
1 1
+ +
+
+ =
+ =
x x C
x x C
x x f
x x f
SRN
] 20 , 20 [
] 20 , 20 [
2
1


x
x
0 10 3 ) x (
225 ) x (
) 1 ( 9 ) x (
) 2 ( ) 2 ( 2 ) x (
2 1 2
2
2
2
1 1
2
2 1 2
2
2
2
1 1
+
+
=
+ + =
x x C
x x C
x x f
x x f
TNK
] , 0 [
] , 0 [
2
1

x
x
5 . 0 ) 5 . 0 ( ) 5 . 0 ( ) x (
0 ) arctan 16 cos( 1 . 0 1 ) x (
) x (
) x (
2
2
2
1 2
2
1 2
2
2
1 1
2 2
1 1
+
+
=
=
x x C
x
x
x x C
x f
x f
OSY
] 10 , 0 [
] 5 , 1 [
] 6 , 0 [
] 5 , 1 [
] 10 , 0 [
] 10 , 0 [
6
5
4
3
2
1

x
x
x
x
x
x
0 4 ) 3 ( ) x (
0 ) 3 ( 4 ) x (
0 3 2 ) x (
0 2 ) x (
0 6 ) x (
0 2 ) x (
) x (
] ) 1 ( ) 4 (
) 1 ( ) 2 ( ) 2 ( 25 [ ) x (
6
2
5 6
4
2
3 5
2 1 4
1 2 3
2 1 2
2 1 1
2
6
2
5
2
4
2
3
2
2
2
1 2
2
5
2
4
2
3
2
2
2
1 1
+

+
+

+
+ + + + + =
+ +
+ + =
x x C
x x C
x x C
x x C
x x C
x x C
x x x x x x f
x x
x x x f
Two-bar
Truss
Design ] 3 , 1 [
] 01 . 0 , 0 [
] 01 . 0 , 0 [
3
2
1

x
x
x
3 2
2
3 2
3 1
2
3 1
5
2 1 1
2 1 2
2
3 2
2
3 1 1
/ 1 80
/ 16 20
10 ) , max( ) x (
) , max( ) x (
1 16 ) x (
x x x
x x x
C
f
x x x x f
+ =
+ =

=
+ + + =

Deepti Chafekar, Jiang Xuan, Khaled Rasheed
Welded
Beam
Design
] 10 , 1 . 0 [
] 10 , 1 . 0 [
] 5 , 125 . 0 [
] 5 , 125 . 0 [

t
l
b
h
3
2
2 2
2 2
"
'
2 2 " ' 2 " 2 '
4
3
2
1
3
2
2
1
) 0282346 . 0 1 ( 022 . 64746
/ 504000
) ) ( 25 . 0 12 / ( 2 2
) ) ( ( 25 . 0 ) 5 . 0 14 ( 6000
2 / 6000
) ) ( ( 25 . 0 / ) ( ) (
0 6000 ) x ( ) x (
0 ) x (
0 ) x ( 30000 ) x (
0 ) x ( 13600 ) x (
/ 1952 . 2 ) x (
) 14 ( 04811 . 0 10471 . 1 ) x (
tb t P
b t
t h l hl
t h l l
hl
t h l l
P C
h b C
C
C
b t f
l tb l h f
c
c
=
=
+ +
+ + +
=
=
+ + + + =




=
+ + =

3.2 Parameter Settings

Each optimization run was carried out with similar parameter settings for all the
methods. The following are the parameters for the three GAs.
Let ndim be equal to the number of dimensions of the problems.
1. Population size: For OEGADO and OSGADO the population size was set to
10*ndim. For NSGA-II the population size was fixed to 100 as recommended in
[19].
2. Number of objective evaluations: Since the three methods work differently the
number of objective evaluations is computed differently. The number of objective
evaluations for OEGADO and OSGADO according to Section 2.1 and 2.2 is given
as Objective evaluations for OEGADO and OSGADO = 2*500*ndim + sum of
feasible points found by each GA in OEGADO model
NSGA-II is a generational GA, therefore for a two-objective NSGA-II:
Total number of objective evaluations =2*population size * number of generations
Since we did not know exactly how many evaluations would be required by
OEGADO before hand, to give fair treatment to NSGA-II, we set the number of
generations of NSGA-II to be 10*ndim. In effect NSGA-II ended up doing
significantly more evaluations than OEGADO and OSGADO for some problems.
We however did not decrease the number of generations for NSGA-II and repeat
the experiments as our methods outperformed it in most domains anyway.
3.3 Results
In the following section, Figures 1-4 present the graphical results of all three methods
in the order of OEGADO, OSGADO and NSGA-II for all problems. The outcomes of
five runs using different seeds were unified and then the non-dominated solutions
were selected and plotted from the union set for each method. We are using graphical
Constrained Multi-Objective Optimization Using Steady State Genetic Algorithms
representations of the Pareto-optimal curve found by the three methods to compare
their performance.
It is worth mentioning that the number of Pareto-optimal solutions obtained by
NSGA-II is limited by its population size. Our methods keep track of all the feasible
solutions found during the optimization and therefore do not have any restrictions on
the number of Pareto-optimal solutions found.
The BNH and the SRN (figures not shown) problems are fairly simple in that the
constraints may not introduce additional difficulty in finding the Pareto-optimal
solutions. It was observed that all three methods performed equally well within
comparable number of objective evaluations (mentioned in Section 3.2), and gave a
dense sampling of solutions along the true Pareto-optimal curve.
Fig. 1 Results for the benchmark problem TNK
Fig. 2 Results for the benchmark problem OSY
The TNK problem (Fig. 1) and the OSY problem (Fig. 2) are relatively difficult. The
constraints in the TNK problem make the Pareto-optimal set discontinuous. The
constraints in the OSY problem divide the Pareto-optimal set into five regions that
can demand a GA to maintain its population at different intersections of the constraint
boundaries. As it can be seen from the above graphs for the TNK problem, within
comparable number of fitness evaluations, the OEGADO model and the NSGA-II
model performed equally well. They both displayed a better distribution of the Pareto-
Deepti Chafekar, Jiang Xuan, Khaled Rasheed
optimal points than the OSGADO model. OSGADO performed well at the extreme
ends, but found very few Pareto points at the mid-section of the curve. For the OSY
problem, it can be seen that OEGADO gave a good sampling of points at the mid-
section of the curve and also found points at the extreme ends of the curve. OSGADO
also performed well, giving better sampling at one of the extreme ends of the curve.
NSGA-II however did not give a good sampling of points at the extreme ends of the
Pareto-optimal curve and gave a poor distribution of the Pareto-optimal solutions. In
this problem OEGADO and OSGADO outperformed NSGA-II while running for
fewer objective evaluations.
Fig. 3 Results for the Two-bar Truss design problem
For the Two-bar Truss design problem (Fig. 3), within comparable fitness evaluations,
NSGA-II performed slightly better than our methods in the first objective. OEGADO
showed a uniform distribution of the Pareto-optimal curve. OSGADO however gave a
poor distribution at one end of the curve, but it achieved very good solutions at the
other end and converged to points that the other two methods failed to reach.
Fig. 4 Results for the Welded Beam design problem
In the Welded Beam design problem (Fig. 4), the non-linear constraints can cause
difficulties in finding the Pareto solutions. As shown in Fig. 4, within comparable
fitness evaluations, OEGADO outperformed OSGADO and NSGA-II in both
distribution and spread [15]. OEGADO found the best minimum solution for f
1
with a
value of 2.727 units. OSGADO was able to find points at the other end that the other
two methods failed to reach. NSGA-II did not achieve a good distribution of the
Pareto solutions at the extreme regions of the curve.
Constrained Multi-Objective Optimization Using Steady State Genetic Algorithms
4 Conclusion and Future Work
In this paper we presented two methods for multi-objective optimization using steady
state GAs, and compared our methods with a reliable and efficient generational multi-
objective GA called NSGA-II. The results show that a steady state GA can be used
efficiently for constrained multi-objective optimization. For the simpler problems our
methods performed equally well as NSGA-II. For the difficult problems, our methods
outperformed NSGA-II in most respects. In general, our methods demonstrated
robustness and efficiency in their performance. OEGADO in particular performed
consistently well and outperformed the other two methods in most of the domains.
Moreover, our methods were able to find the Pareto-optimal solutions for all the
problems in fewer objective evaluations than NSGA-II. For real-world problems, the
number of objective evaluations performed can be critical as each objective
evaluation takes a long time. Based on this study we believe that our methods can
outperform multi-objective generational GAs for such problems. However, we need
to experiment more and find out whether there are other factors that contribute to the
success of our methods other than their steady state nature.
In the future, we would like to experiment with several steady state GAs as the base
method. We would also like to improve both of our methods. Currently they do not
have any explicit bias towards non-dominated solutions. We therefore intend to
enhance them by giving credit to non-dominated solutions. OEGADO has shown
promising results and we would like to further improve it, extend its implementation
to handle more than two objectives and further explore its capabilities. The current
OSGADO implementation can already handle more than two objectives. We would
also like to use our methods for more complex real-world applications.
Acknowledgement
This research is sponsored by the US National Science Foundation under grant CTS-
0121058. The program managers are Drs. Frederica Darema, C. F. Chen and Michael
Plesniak.
References
1. Khaled Rasheed. GADO: A genetic algorithm for continuous design optimization.
Technical Report DCS-TR-352, Department of Computer Science, Rutgers, The State
University of New Jersey, New Brunswick, NJ, January 1998. Ph.D. Thesis,
http://webster.cs.uga.edu/~khaled/thesis.ps.
2. Khaled Rasheed and Haym Hirsh. Learning to be selective in genetic-algorithm-based
design optimization. Artificial Intelligence in Engineering, Design, Analysis and
Manufacturing, 13:157-169, 1999.
3. Deb, K., S. Agrawal, A. Pratap, and T. Meyarivan (2000). A fast elitist non-dominated
sorting genetic algorithm for multi-objective optimization: NSGA-II. In Proceedings of
the Parallel Problem Solving from Nature VI, pp.849-858.
Deepti Chafekar, Jiang Xuan, Khaled Rasheed
4. Khaled Rasheed and Haym Hirsh. Informed operators: Speeding up genetic-algorithm-
based design optimization using reduced models. In Proceedings of the Genetic and
Evolutionary Computation Conference (GECCO'2000), pp. 628--635, 2000.
5. K. Rasheed., S. Vattam, X. Ni. Comparison of Methods for Using Reduced Models to
Speed up Design Optimization. In Proceedings of the Genetic and Evolutionary
Computation Conference (GECCO'2002), pp. 1180--1187, 2002.
6. Khaled Rasheed. An incremental-approximate-clustering approach for developing
dynamic reduced models for design optimization. In Proceedings of the Congress on
Evolutionary Computation (CEC2002), pp. 986993, 2002.
7. K. Rasheed., S. Vattam, X. Ni. Comparison of methods for developing dynamic reduced
models for design optimization. In Proceedings of the Congress on Evolutionary
Computation (CEC2002), pp. 390395, 2002.
8. William H. Press, Saul A. Teukolsky, William T. Vetterling, and Brian P. Flannery.
Numerical Recipes in C: the Art of Scientific Computing. Cambridge University Press,
Cambridge [England]; New York, 2
nd
edition, 1992.
9. J.D.Schaffer .Multi-objective optimization with vector evaluated genetic algorithms. In
Proceedings of an International Conference on Genetic Algorithms and Their
Applications, J.J. Grefenstette, Ed., Pittsburg, PA, July 24-26 1985, pp. 93-100, sponsored
by Texas Instruments and U.S. Navy Center for Applied Research in Artificial Intelligence
(NCARAI).
10. Binh and Korn. MOBES: A multi-objective Evolution Strategy for constrained
optimization Problems. In Proceedings of the 3
rd
International Conference on Genetic
Algorithm MENDEL 1997, Brno, Czech Republic, pp.176-182.
11. Srinivas, N. and Deb, K. (1995). Multi-Objective function optimization using non-
dominated sorting genetic algorithms. Evolutionary Computation (2), 221-248.
12. Tanaka, M. (1995). GA-based decision support system for multi-criteria, optimization. In
Proceedings of the International Conference on Systems, Man and Cybernetics-2, pp.
1556-1561.
13. Osycza, A. and Kundu, S. (1995). A new method to solve generalized multicriteria
optimization problems using the simple genetic algorithm. Structural Optimization (10).
94-99.
14. Deb, K. Pratap, A. and Moitra, S. (2000). Mechanical Component Design for Multiple
Objectives Using Elitist Non-Dominated Sorting GA. KanGAL Report No. 200002.
15. Ranjithan, S.R., S.K. Chetan, and H.K. Dakshina (2001). Constraint method-based
evolutionary algorithm (CMEA) for multi-objective optimization. In E.Z. et al. (Ed.),
Evolutionary Multi-Criteria Optimization 2001, Lecture Notes in Computer Science 1993,
pp.299-313. Springer-Verlag.
16. Zitzler, E., Laumanns, M., and Thiele, L. (2001). SPEA2: Improving the Strength Pareto
Evolutionary Algorithm. Technical Report 103, Computer Engineering and Networks
Laboratory (TIK), Swiss Federal Institute of Technology (ETH) Zurich, Gloriastrasse 35,
CH-8092 Zurich, Switzerland.
17. Crone, D. W., Knowles, J.D., and Oates, M.J. (2000). The Pareto Envelope-based
Selection Algorithm for Multi-objective Optimization. In Schoenauer, M., Deb, K.,
Rudolph, g., Yao, X., Luton, E., Merelo, J.J., and Schewfel, H.-P., editors, Proceedings of
the Parallel Problem Solving from Nature VI Conference, pp.839-848, Paris, France.
Springer. Lecture Notes in Computer Science No. 1917.
18. K. Deb. Multi-objective optimization using evolutionary algorithms. Chichester, UK:
John Wiley, 2001.
19. K. Deb. S. Gulati (2001). Design of truss-structures for minimum weight using genetic
algorithms, In Journal of Finite Elements in Analysis and Design, pp.447 465, 2001.