Proceedings of the 3d World Congress on Intelligent Control and Automation June 2 C J u l y 2,2000, Hefei, P.R.

China

Study on Multi-Objective Genetic Algorithm

Ying Gao Lei Shi Pingjing Yao

Institute of Process Systems Engineering, Dalian University of Technology
Dalian, 116012, P.R China

, Abstract-Multi-objective genetic algorithm (MOGA) is an The basic structure of multi-objective genetic

effective approach in solving multi-objective optimization algorithm is similar to' the single objective genetic
problems. The current multi-objective genetic algorithms are algorithm. The main difFerence between MOGA and
reviewed in this paper, and a new form of MOGA,steady- SOGA is the fitness assignment strategy, while the
state non-dominated sorting genetic algorithm (SNSGA), is difference among MOGAs is also lies in the fitness
realized by combining the steady-state ideas in single- assignment and selection strategies. The ament MOGAs
objective genetic algorithm (SOGA) and the fitness are reviewed in section 2, and a new form of MOGA,
assignment strategy of non-dominated sorting genetic steady-state non-dominated sorting genetic al- is
algorithm. The fitness assignment strategy ir improved and a introduced in section 3.
new self-adaptive decision scheme of,,U is proposed This
algorithm is proved to be succeafd with some tert problems IL MULTI-OBJECTIVE GENETIC ALGORITHM
including GA difficult problem and GA deceptive problem.
A. MuWObjedive Oprimitatin
I. INTRODUCTION
Multi-objective optimization is a very i m p o m t
Genetic algorithm (GA)is a kind of self-adaptive research area because many practical engineering
global searching optimization algorithm Merent from problems are multi-objective optimization problem. Multi-
conventional optimization algorithms, genetic algorithm is objective optimization problem can usually be expressed
based on population, in which each individual is evolved as the following [2]:
parallelly, and the ultimate result is included in the last
population. The population evolution depends on some
genetic operators acted on the current population and is
realized by the generation of a new pupulation. Generally
the genetic operators are selection, crossover and mutation.
These genetic operatm have key effect on the
performance of the genetic algorithm. The intrinsic
parallel and the effective use of the global information are
the main character of genetic algorithm [l-31. Where V-min represents vector minimization, which

The basic approach of genetic algorithms is shown in

the following:
1. Determine the coding structure. simultaneously.
2. Determine the fitness function.
3. Determine the selection strategy. The simultaneous optimization of multiple, possibly
4. Select the conttol parameters. competing, objective function deviates from single
5. Design the genetic operators. function optimization in that it seldom admits a single,
6. Determine the terminate crithon. perfect solution, but rather a set of alternative solutions,
we call it nondominated solution set or Pareto set. These
Genetic algorithm is an interactive process. In the solutions are optimal in the wider sense that no other
course of each interactive process, selection operation acts solutions in the search space are superior to them when all
on the c m t population according to certain regulation. objectives are considered.
The operations of crossover and mutation are performed
for the selected individuals and then a new population is The main goal of multisbjective optimization
formed. The process repeats until the convergence techniques is to find the one or multiple acceptable
criterion is satisfied. solution in the Pareto set.

646
0-7803-5995-x/00/$ IO.00 Q2000 IEEE

Authorized licensed use limited to: UNIVERSITE DE SOUSSE. Downloaded on January 25,2023 at 08:12:41 UTC from IEEE Xplore. Restrictions apply.
 Conventional optimization techniques, such as for each objective separately. In detail, the mating pool is
gradient-based and simplex-based methods are difficult to divided into n parts of equal size, part i is filled with
extend to the multi-objective case. Genetic algorithms individuals that are chosen at random from the current
have been recognized to be possiily well-suited to multi- population according to objective i. Afterwards, the
gbjective optimization because individuals can search for mating pool is shuffled and crossover and mutation are
multiple solutions in parallel, eventually taking advantage performed as usual. This method is usually implemented
of any similarities available in the family of possible in combination with fitness proportionate selection. Each
solution to the problem. The ability to handle complex objective is effectively weighted proportionately to the
problems, involving features such as discontinuities, size of each sub-population and proportionately to the
multimdity, disjoint feasible spaces reinforces the inverse of the average fitness of the whole population at
potential effectiveness of GAS in multi-objective search each generation.
and optimization.
Population-based non-Pareto approaches, however,
Genetic operations act on the whole population, are able to evolve multiple nondominated solutions
emphasis on the integration of individuals, thus genetic parallelly, thexefore, the population is mostly monitored
algorithms is an effective way in solving multi-objective for nondominated solutions. But in contrast to the Pareto-
optimization problem. The combination of mdti-objective based approaches, they do not make direct use of the
programming and genetic algorithm to explore multi- concept of Pareto dominance.
objective genetic algorithm has significant meaning in
solving engineering problems. 3) Pareto-Based Approaches

B. MuiibObjedive Genaic Algorithm Pareto-based approaches make direct use of the

actual definition of Pareto-optimality. It operates on the
Mdti-objective genetic algorithm can be categorized principle of the non-dominance of the solutions.
into several evolutionary approaches like plain
aggregating approaches, population-based non-Pareto The idea of using pareto-based fitness assignment is
approaches, Pareto-based approaches and niche induction to use nondominated ranking and selection to move a
techniques [4-51. population toward the Pareto front in a multi-objective
optimization problem. The basic idea is to find a set of
1) Plain Aggregating Approaches individuals in the population that is Pareto nondominated
to the rest of the population. These individuals are then
Plain aggregating approaches apply weighting assigned the highest rank and eliminated from m er
aggregating method to convert the multi-objective contention. Another set of Pareto nondominated
problems into single objective problems, and then use the individuals are determined from the remaining population
single function genetic algorithm to get the solution. and are assigned the next highest rank. This process
continues until the population is suitably ranked.
Aggregation methods combine the objectives into a
higher scalar function used for fitness calculation, they Once the fitness is determined, the Pareto-based
produce one single solution and require profound domain approach can get the solution using SOGA process.
knowledge. Pareto-based MOGA can also be devided into parellel
selection approach, ranking selection approach and
Optimizing a combination of the objectives has the sharing function approach. All these approaches have their
advantage of producing a single compromise solution, own strengthes and weaknesses, thus the combination use
requiring no further interaction with the decision-maker. of these approaches help to develop their benefit.
The problem is, the setting of the weighting coefficients
relies on the preference of the decision-maker. If the Pareto ranking eliminates sensitivity to the possible
optimal solution cannot be accepted, either due to the nonconvexity of the trade-off. The other advantage of
function used excluding aspects of the problem which Pareto-ranking is that, because it rewards good
were unknown prior to optimization or to an inappropriate performance in any objective dimension regardless of the
setting of the coefficients of the combining function, new others, solutions which exhibit good performance in many
runs of the optimizer may be required until a suitable objective dimensions are more likely to be produced by
solution is found. recombination. Pareto optimal selection also eliminates
the needs to combine disparate objectives into a single
2) Population-Based Non-ParetoApproaches fitness as is usual in genetic algorithms.

Vector evaluated genetic algorithm (VEGA) is a kind 4) Niche Induction Techniques

of population-based non-Pareto approach. Vector
evaluated genetic algorithm carried out selection operation Pareto-based ranking correctly assigns all non-
647

Authorized licensed use limited to: UNIVERSITE DE SOUSSE. Downloaded on January 25,2023 at 08:12:41 UTC from IEEE Xplore. Restrictions apply.
 dominated individuals the same fitness, but it does not
guarantee that the Pareto set is uniformly sampled. When
presented with multiple equivalent optimhtion, finite A multi-objective genetic algorithm, which is based
populations tend to converge to only one of them, due to on the Pateto ranking, is the “domjnated salting genetic
stochastic errors in the selection proms. This algorithm. The fitness assignmat is carried out in several
phenomenon, known as genetic drift,has been observed in steps. Before selection is performed, the population is
natural as well as in artificial evolution, and can also occur ranked on the basis of “mation:allnondominated
in Pareto-based evolutionary optimization. individualsare classfied into one CategoIy (endowed with
a dummy fitness value, which is proportionate to the
Since preservation of diversity is crucial in the field population size, to provide an equal reproductive potential
of multi-objective optimization, many multi-objective for these individuals). To mainhin the diversity of the
GAS incorporate niching techniques [6], the mostly population, these classified individuals are shared with
implemented of which is fitness sbaring. Fitness sharing is their dummy fitness values. Then this group of classified
based on the idea that individuals in a particular niche individualsare ignored and another layer of nondominated
have to share the resources available, which is similar to indivictualsare considered. The dummy fitness is set to a
the nature. Thus, the fitness value of a certain individual is value less than the smallest shared fitness value m the
the more degraded if the more individuals are located in current nondominated front. The process continues untill
its neighborhood. Neighborhood is defined in terms of a all individuals in the population are classified.
distance measure and specified by the so-called niche
radius oh. A stochastic remainder proportionate selection was
used for this approch Since the individuals m the front
Niching technique is suggested to keep GA from have the maximum fitness value, they always get more
converging to a single point on the front and a niching coppies than the rest of the population. This allows to
mechanism such as sharing would allow the GA to search for nondominated re@ons,and results in quick
maintain individuals all along the nondominated frontier. convergence of the population toward such regions.
Thus the additional use of fitness sharing was proposed to Shareing helps to dishihte it over its regions.
prevent genetic drift and to promote the sampling of the
whole Pareto set. In this paper, according to the character of Pareto
basexi optimum fitness assignment approaches, a self-
EL STEADY-STATE NON-DOMINATED SORTING adaptive,,a determination scheme is realized to make
GENETIC ALGORITEIM the f- value of the majority population is 100, i;yB
ranging beteen 33.5-50. At this circwnStance, the
The two main tasks must be handled in exploring algorithm can keep a certain selection pressure and the
multi-objective genetic algorithms are: (1) induce the uniformity solution rlistributon can also realized.
population move to the pareto front; (2) keep the diversity
of the population in the current pareto front [7]. 3.Population Replackg 3haiegv

In this paper we explored steady-state nondominated Steady state ideas in the population replacing
sorting genetic algorithm (SNSGA) by the combination of strategy of genetic algorithm is applied in SNAGA [9].
steady-state ideas in single-objective genetic algorithm Here we adopt population overlap,u+Aselection
and the fitness assignment strategy of nondominated strategy. First, the parent individuals are selected in the
sorting genetic algorithm [8]. The individual fitness current population (the population size is p ) to form a
assignment strategy of SNSGA can apply Pareto-based temporary population (the population size is A , Ais
optimum approach, thus the individual fitness is assigned usual@smaller than ,U) using the stochastic remainder
according to its relative superority in the population. sampling selection approach. Individuals in the tempomy
Sharing technique is also applied to keep the uniformity population performc~ssoverand mutation operation, then
distribution of the solution. By applying self-adaption combined with the parent population to form a super
method of sharing operator,,a , the diversity of the population (the population size is p + A). The super
solution can be maintained in the pareto fiont. population perform fitness assignment and sorting
amrding to the nondominated sorting fitness assighent
Population overlap stmtegy is applied in SNSGA, its approach, the predominant individuals form the oftspings
realization depends on the selection and replacing steps. (the size is p ) . Thus the optimum solution in the current
The crossover individuals are selected according to the front must be preserved in the next generation, while the
stochastic remainder sampling selection approaches. The predomioant individuals in the parent population can also
selection of replacement indivaduals can adopt be preserved by larger prubability.
p+Aselection strategy, to preserve the predominant
individuals in the parent population and acceleme the N.SNSGA PERFORMANCE TEST
convergence of the algorithm.
648

Authorized licensed use limited to: UNIVERSITE DE SOUSSE. Downloaded on January 25,2023 at 08:12:41 UTC from IEEE Xplore. Restrictions apply.
 A group of test problems were introduced in this
paper to test the performance of SNSGA, here we take the
result of the two typical test problems as example to
0.2.
suggest the performance of SNSGA. Problem 1 is a two
0.15.
varible-three objective problem, it is usually used to test 0.1.
the ability of MOGA in searchmg the global Pareto front. a 0.05.
Problem 2 is the deception problem test function of
MOGA.

A. TedProblems

Test problem 1 [lo]:

min I; (2) = O.S(X: + x i +sin(x: + x,' 1,

(3x, - 2 x , +4)2 - x 2 +1y Fig.2 Distribution of the 100'' Population
min f i ( Z ) =
8
+ (XI 27
+ 15,
symmetry, three dimension curve. The binary coding
length of each viarable is 15. The population size p is 200,
replacing scale 2 is 120. The crossover probability p , is
0.8, and the mutation rate p , is 0.0 1. The value of qhe is
0.0125 (obtained from the self-adaption approach). Figure
1 shows the initial population distribution, while figure 2
Test problem 2 [7]: is the population distribution of the 100" generation. The
test result is accordant with the literature, which helps to
min f ; =l+u(l,), prove the ability of SNSGA in searching the Pareto front.

2) Result of Test Problem 2:

Here we apply binary coding approach, the coding

length of the first viable is 10, and the coding length of the
other three viable is 5 separately, thus the total coding size
is 25. The population size p i s 80, and the population
replacing scale 2 is 48. The crossover probabilityp , is 0.6,
Where function ~ ( 1 , )is the number of 1 in binary while the mutation rate p , is 0.001. In this problem the
wordstring fj. Hamming distance gene-type sharing operator is applied
and the value of C T , ~is, 3. ~ ~Figure
~ 3 shows the initial
B. Ted Result generation and the 100'' generation population distribution.
This test problem has 8 Pareto front lines. Only the global
1) Result of Test Problem I: Pareto front and one of the local Pareto front is shown in
this figure. The test result suggests that SNSGA has the
The Pareto front of this function is a non-liner, non- ability in solving GA deception problems

02s

0.16
O.1

Fig.1 Distribution of Initial Population

649

Authorized licensed use limited to: UNIVERSITE DE SOUSSE. Downloaded on January 25,2023 at 08:12:41 UTC from IEEE Xplore. Restrictions apply.
 V. CONCLUTION [3] Pan Zhengjun, Kang Lishan, Chen Yuping. Evaluation
Calculation, Tsinghua University Publishing Company,
Design problems are often formulated as multi- Beijing; 1998.
objective problem. In many cases such problem involve
tradeoffs among possibly conflicting objectives. MOGA [4] Carlos M. Fonseca and Peter J. Fleming. “An overview of
has been proved to be an effective approach to solve the evolutionaxy algorithms in multi-objective optimization”,
multi-objective optimizationproblems. In this paper a new EvolutionaryComputation,SpMg 1995, 3( 1): pp. 1-16.
form of MOGA, steady-state nondominated sorting [5] Carlos A Coello Coello. “A Comprehensive Survey of
genetic algorithm is realizedby combining the steady-state Evolutionazy-Based Multisbjective Optimization
ideas in SOGA and the fitness assignment strategy of non- Techniques”, Knowledge and Information Systems. An
dominated sorting MOGA. The individual fitness International Journal,1999.
assignment strategy of SNSGA can adopt the Pareto based
optimization approach, and sharing technique is used to [6] Jeffrey Hom and Nicholas Nafpliotis. “Multi-objective
keep the uniformity distribution of the solution. A sharing Optimization using the Niched Pareto Genetic Algorithm”,
operator o , self-adaptive approach is put forward to Technical Report IlliGAl Report 93005, Univemity of
keep the diversity of the solution in the Pareto front. Illinois at Urbana-Champaign, Udana, Illinois, USA,
Population overlap strategy is applied in SNSGA, the 1993.
selection of the crossover individual can apply the [7] Kalyanmoy Deb. “Multi-objective Genetic Algorithms:
stochastic remainder sampling approach, and the Problem Difficulties and Construction of Test Problems”,
replacement individuals are selected according to the Technical Report CI-49/98, Dortmund: Department of
p + A selection scheme. Thus the predominant ComputerSciencP/zTll. University ofDomund, Gemany,
individuals in the parent generation can be preserved and 1998.
the convergence of the algorithm can also be accelerated.
[8] Srinivas N. and Kalyanmoy Deb. “Multisbjective
The test results of GA dif€idt problem and GA Optimization Using Nondominated Sorting in Genetic
deceptive problem suggest that the SNSGA explored in Algorithms”,Evolutionary Computation,fall 1994,2(3): pp.
this paper shows its super performance in the convergence 221-248.
and the population distribution, This constitutes a stability [9] Syswerda G.“A Study of Reproduction in Generational and
foundationfor the practical application of SNSGA. Steady-state Genetic Algorithms, Foundations of h e t i c
Algorithm?, Morgan Kau/“n. San Mateo, CA, 1991,
VI. REFJZRENCES pp. 94-10 1.
[I] Wei Quanling, Wang Rishuang, Xubing. Mhematical [IO] David A. Van Veldhuizen and Gary B. Lamont. “Multi-
Pmgmmming and Optimization Design, National Defense objective Evolutionary Algorithm Test Suites”, In Janice
Industry Publishing Company, Beijing; 1984. Camll, Hisham Haddad, Dave Oppenheim, Barren Byant,
[2] Zhou Ming, Sun Shudong. Theory ofGenetic Algorithm and and Gav B. Lamont, editors, Proceedings of the 1999
Itr Application, National Defense Industry P u b l h n g ACM Symposium on Applied Computing, San Antonio,
Company, Beijing; 1999. Texas, 1999, pp. 351-357.

650

Authorized licensed use limited to: UNIVERSITE DE SOUSSE. Downloaded on January 25,2023 at 08:12:41 UTC from IEEE Xplore. Restrictions apply.