You are on page 1of 9

Elitist Non-Dominated Sorting GA-II (NSGA-II) as

a Parameter-less Multi-Objective Genetic Algorithm


Khoa Duc Tran
Graduate School of Computer and Information Sciences
Nova Southeastern University
khoa@nova.edu

Abstract instead of a set of good trade-off solutions from which the


decision maker can choose one.
Genetic Algorithms (GAs) are general-purpose heuristic The potential of Genetic Algorithms (GAs) in multi-
search algorithms that mimic the evolutionary process in objective optimization was initially hinted by Rosenberg in
order to find the fittest solutions. The algorithms were 1960s [4]. However, this research area remained
introduced by Holland in 1975. Since then, they have unexplored until recently it has received a lot of interest [5],
received growing interest due to their ability to discover [6], [7]. The reason for this increasing interest because a
good solutions quickly for complex searching and GA is capable of finding multiple optimum solutions in one
optimization problems. The traditional GAs have been single run.
converted to multi-objective GAs to solve multi-objective However, Genetic Algorithms (GAs) are not easy to use
optimization problems successfully. However, GAs require because they require parameter tunings in order to achieve
parameter tunings (such as population size, mutation the desirable solutions. The task of tuning GA parameters
probabilities, crossover probabilities, selection rates) in has been proven to be far from trivial. Many researchers
order to achieve the desirable solutions. The task of tuning have been trying to understand the interdependencies of
GA parameters has been proven to be far from trivial due GA parameters [8], [9]. The parameter-less GA [10], [11]
to the complex interactions among the parameters. It takes was recently proposed as a technique to make GAs easier to
trial and error experiments to obtain the optimal GA use. The objective of this research is to develop the elitist
parameter settings for an arbitrary real-world problem. non-dominated sorting GA (NSGA-II) for multi-objective
Many researchers have been trying to understand the optimization [6] as a parameter-less multi-objective GA.
interdependencies of GA parameters in order to determine The research then will evaluate and discuss the
their optimal settings. The objective of this research is to performance of the parameter-less NSGA-II against the
develop the elitist non-dominated sorting GA (NSGA-II) for original NSGA-II with optimal parameter settings using the
multi-objective optimization as a parameter-less multi- experiment result on a test problem borrowed from the
objective GA. The research then will evaluate and discuss literature.
the performance of the parameter-less NSGA-II against the The remainder of this paper is organized as follows:
original NSGA-II with optimal parameter settings using the Section 2 presents prior research related to the topic of
experiment result on a test problem borrowed from the parameter setting in GAs and the Fast Elitist Non-
literature. Dominated Sorting GA-II for Multi-Objective
Optimization: NSGA-II. Section 3 presents the modified
1. Introduction version of NSGA-II as a parameter-less NSGA-II. Section
4 describes the test problem with available published
Multi-objective optimization (also known as multi- results used to evaluate the parameter-less NSGA-II.
criteria optimization, multi-performance, or vector Section 5 presents the experimental results showing the
optimization) is the task of finding one or more optimum performance of the parameter-less NSGA-II against the
solutions which would give the values of all the objective original NSGA-II and provides some comments. Section 6
functions acceptable to the decision maker [1], [2], [3]. suggests the improvements for the parameter-less NSGA-II.
Thus, multi-objective optimization is a component of a Finally, the paper ends with the summary and conclusion.
Decision Support System to assist a practitioner in his
decision-making. Many approaches have been suggested 2. Prior Research
for solving multi-objective optimization problems: from
naively combining objectives into one, to the use of game 2.1. Empirical Studies
theory to coordinate the relative importance of each
objective [2]. However, these classical approaches can find De Jong [12] conducted one of the first empirical studies
only one optimum solution in a single simulation run to understand the complex interactions and
interdependencies of GA parameters in his doctoral
dissertation “An analysis of the behavior of a class of and two-point crossover will be used 25% of the time. The
genetic adaptive systems.” Based on his studies, De Jong technique was tested using a set of n-Peak and
introduced a good set of parameter settings that have been Halmiltonian Circuit problems. The results have shown that
adopted widely and sometimes referred to as “standard” this adaptive crossover operator out-performs non-adaptive
settings: population size of 50-100, crossover probability of crossover operator especially with large population sizes.
0.6, and mutation probability of 0.001. However, the Smith and Smuda [16] introduces an algorithm for
theoretical studies described below have proven these adaptively resizing GA populations. This algorithm
“standard” settings problematic. requires users to supply a desired target value for
acceptable selection loss. The algorithm then estimates
2.2. Theoretical Studies schema fitness variances and population sizes so that the
selection loss approximates the users’ loss requirement.
The theoretical models on population sizing [9], [13] However, there are some limitations with this algorithm.
identifies the shortcomings of the “standard” settings and First, it is hard to match users’ specified selection loss with
suggests that the population size should be proportional to the actual accuracy of the GA. Second; the estimation of
the problem’s difficulty. These theoretical models are schema fitness variances is very noisy because the GA
impractical because they reply on parameters that may not performs many samples simultaneously. Third, the nearly
be available for real-world problems and require knowledge perfect building block mixing may not be available in
of GA theoretic concepts and statistics in order to use. practice.
However, the theoretical models provide useful guidelines
for choosing population size parameter. Moreover, these 2.4. Parameter-less GA
theoretical models point out that setting the population size
to a fixed value (i.e. in the range of 50 to 100) regardless of Harik & Lobo [10] proposed a parameter-less genetic
the problem’s difficulty is problematic. Since problem’s algorithm to make GAs easier to use and available to more
difficulty is usually unknown until solved, in practice users users. Later, Lobo elaborated this idea in his doctoral
will have to try a number of population size settings to dissertation [11]. According to Harik and Lobo, users do
guess the right setting. There are two problems associated not need to know the GA parameters such as population
with this guessing game: 1) setting the population size not sizes, crossover probabilities, mutation probabilities,
large enough will cause the GA to converge to sub-optimal selection rates, and other GA parameters in order to use a
solutions; 2) setting the population size above optimum will GA. Hence; a GA should have no parameters.
cause the GA to waste unnecessary computational Harik & Lobo chose to ignore mutation in the
resources. parameter-less GA for the sake of simplicity. They preset
The theory of genetic algorithms’ control maps suggests the selection pressure s and crossover probability pc to fixed
the guidelines for setting the values of two parameters values (s = 4, pc = 0.5) according the schema theorem [17]
selection rate and crossover probability in order to in order to ensure building block growth. Using the
constitute robust settings [14]. schema theorem equation [17], [18], they derived the
expression: s(1 – pc) where s substitutes the term that
2.3. Parameter Adaptation accounts for the selection operation, pc substitutes the term
that accounts for the crossover operation, and term that
Spears [15] applied self-adaptation mechanism for accounts for the mutation operation is ignored. Thus, by
selecting an optimal crossover operator: uniform crossover setting s = 4 and pc = 0.5, the expression gives a net growth
or two-point crossover. In this technique, one bit is of 2 which ensures the necessary building blocks will grow.
appended to the end of every individual string in the The parameter-less GA selects the right population size by
population (i.e. a ‘1’ is uniform crossover, and a ‘0’ refers running multiple populations of various sizes
to two-point crossover). All genetic operators can be simultaneously, and eliminating the smaller populations
applied to these extra bits because this approach is tightly that have lower average fitness, and the ones that drift.
coupled. Spear pointed out that there are two cases: local According to the experimental results on three test
adaptation and global adaptation. In local adaptation, the problems: the one-max problem, the noisy one-max
last bits of the two chosen individuals are examined to problem, and the bounded deceptive problem [10], [11], the
decide which crossover operator should be applied. If the parameter-less GA is a bit slower than a GA that starts with
two bits are ‘1’s, uniform crossover is applied. If the two optimal parameter settings. However, it takes trial and
bits are ‘0’s, two-point crossover is used. If the two bits are error experiments to obtain the optimal parameter settings
different, crossover operator is randomly selected. In global for an arbitrary real-world problem.
adaptation, the last bits of the whole population
probabilistically determine which crossover operator is 2.5. Elitist Non-Dominated Sorting GA-II for
applied on each individual. For example, if the population Multi-Objective Optimization (NSGA-II)
has 75% ‘1’s and 25% ‘0’s in the last bits of individuals
then uniform crossover will be applied 75% of the time, Deb et al. [6] proposed a fast elitist multi-objective GA
that is able to find much spread solutions over the Pareto- order of their ranking. To fill in the new population with
optimal front and requires low computational requirements exactly N individuals, the solutions of the last front is
( O(mN2), where m is the number of objectives and N is the sorted in descending order using the crowded comparison
population size). The NSGA-II has the following features: operator and the best solutions are chosen to fill up the
1) Use elite-preserving operator to preserve and use remain slots in the new population. The parameter-less
previously found best solutions in subsequent NSGA-II then applies binary tournament selection, and
generations; crossover on the new population and the current population
2) Use the non-dominated sorting concept in GAs [18], to produce a new offspring population. Thus, the selection
[19]; rate is always equal to the population size N.
3) Use the crowded tournament selection operator to As in the NSGA-II, the parameter-less NSGA-II has the
preserve the diversity among non-dominated following settings internally within the algorithm:
solutions in the later stage of the run in order to 1) Distribution indices for real-coded crossover
obtain a good spread of solutions. In this diversity- operator ηc = 20;
preserving mechanism, if the solutions are on the 2) Distribution indices for real-code mutation
same non-dominated front the selection will be operator ηm = 100.
based on their crowding distance which is a
measure of density of solutions in the
3.2. Establish a Race among Multiple Populations
neighborhood. The one residing in a less crowded
area (with a larger crowding distance) will be
with a Constant Number of Generations
selected.
According to the experimental results on five test The first attempt is to let the parameter-less NSGA-II
problems taken from the literature, the NSGA-II has out- execute multiple populations with different population sizes
performed the Pareto-Archived Evolution Strategy (PAES) simultaneously with a constant maximum number of
– another multi-objective EA with the explicit goal of generations equals to 500. The initial population size is set
preserving spread on the non-dominated front [6]. to 25 and maximum population size is set to 500. Similarly,
to the simple parameter-less GA, the parameter-less
3. Development of the Parameter-less NSGA- NSGA-II allows the smaller populations more generations
to run. The coordination of the array of populations is
II implemented with a counter. Overall, the nth population is
allowed to run 32 times more generations than the n+1th
As the name implied, the parameter-less NSGA-II has population. Each time a population yields its execution if
no parameter making the GA easier to use and available to the next population has yet not been created the algorithm
more users. The parameter-less NSGA-II integrates the creates a new population twice as large as the previous one
NSGA-II [6] into the parameter-less GA [10] making the until the maximum population size is reached. Harik and
NSGA-II parameter-less. Lobo [10] point out that the reason for doing this is that the
smaller populations expects to converge faster than the
3.1. Remove Parameter Settings larger one and the algorithm would terminate with a proper
population size. On the other hand, if at any point in time, a
As in the parameter-less GA, the parameter-less NSGA- larger population has an average fitness better than that of a
II has the following settings internally within the algorithm: smaller population then the algorithm would stop running
1) Ignores mutation for the time being pm = 0; the smaller population because it is very unlikely that the
2) Crossover probability pc = 0.5. This setting enables smaller population would produce better solutions than the
the parameter-less NSGA-II works more or less larger one.
well on both easy and difficult problems [10]; However, it turns out that this concept does not work
3) Selection rate s = 4; quite well for MOEAs because most MOEAs assign each
4) Population size: parameter-less NSGA-II will run solution a fitness equal to its a non-domination rank (1 is
multiple populations with different population size the best rank, 2 is the next best rank and so on) and all
simultaneously. solutions in the first non-dominated front (or rank) have the
The selection rate however is also ignored because the same fitness [1]. Hence, there is no discrimination to any
parameter-less NSGA-II uses different approach from the of these solutions [20]. Therefore, there is no simple way
simple GA to create an offspring population. The to determine the better solutions between two populations.
parameter-less NSGA-II combines parent population and As the result of this, the parameter-less NSGA-II defines
previous offspring population to create a combined the convergence and fitness of a population with the
population of size 2N. Since all previous and current following property:
population individuals are included in the combined Convergent property
population, the elitism is preserved. The combined 1) The number of solutions in the first non-dominated
population is sorted according to non-domination. The new front is equal to the population size N;
population is then filled with non-dominated fronts in the 2) All solutions in the first non-dominated front have
no constraint errors (for constrained problems). 3.3. Establish a Race among Multiple Populations
To determine the better solutions between two with Increasing Number of Generations
populations, the following criteria is naively defined:
Better Solutions Condition The test results on the original NSGA-II (see Table 2)
Population P1 has better solutions than population P2 if has shown that as the population size is increased and the
P1 has converged and P1 has more solutions in the first non- maximum number of generations is kept constant, the
dominated front than that of P2, and P1 has executed for performance declines. This result confirms Krishnakumar’s
number of generations less than or equal to that of P2. work [21]. The second attempt is to increase the maximum
The main loop of the parameter-less NSGA-II is number of generations for each population proportionately
described as below: with its population size (10*N). Each population is allowed
pc=0.5; pm=0.0; ηc=20; ηm=10; s = 4
Pop : array of 100 populations to run until it converges and the solutions are printed out to
Counter : base 32 on 100 files. The convergent property is modified as below:
Do Convergent property
// which population to run? 1) The number of solutions in the first non-dominated
ix = get Counter
If Pop[ix] exists front is equal to the population size N;
// run NSGA-II for one generation 2) All solutions in the first non-dominated front have
Run Pop[ix] for 1 generation no constraint errors (for constrained problems).
Else The main loop of the parameter-less NSGA-II is
If max population size reached modified as below:
// do not create new population
pc=0.5; pm=0.0; ηc=20; ηm=10; s = 4
Reset Counter Pop : array of 100 populations
// continue running existing populations
Counter : base 32 on 100
Continue
Do
Else // which population to run?
// new population with population size
ix = get Counter
// N =2 * population size of Pop[ix-1]
If Pop[ix] exists
// and filled with random individuals
// run NSGA-II for 1 generation
Pop[ix] = new population
Run Pop[ix] for 1 generation
EndIf
Else
EndIf If max population size reached
If Pop[ix] converges // do not create new population
For ii = 0 to ix-1 do Reset Counter
Kill Pop[ii] // kill smaller populations // continue running existing populations
EndFor Continue
EndIf Else
For ii = 0 to ix-1 do // new population with population size
// according to the above definition // N = 2 * population size of Pop[ix-1] and
If Pop[ix] has better solutions than Pop[ii] // max. number of generations = 10*N
Kill Pop[ii] // and filled with random individuals
EndIf Pop[ix] = new population
// according to the above definition EndIf
If Pop[ii] has better solutions than Pop[ix] EndIf
// smaller population has better solutions If Pop[ix] converges
found_best_ever_solutions = true Print out the solutions
break // the loop For ii = 0 to ix-1 do
EndIf // kill smaller populations
EndFor Kill Pop[ii]
increase Counter EndFor
Until found_best_ever_solutions or all populations converge EndIf
Print out the solutions For ii = 0 to ix-1 do
// according to the above definition
It has been observed that the code to find better If Pop[ix] has better solutions than Pop[ii]
solutions among populations has never reached because the Kill Pop[ii]
larger populations always catch up with the smaller ones EndIf
with more solutions in the first non-dominated front. The // according to the above definition
If Pop[ii] has better solutions than Pop[ix]
last population with the maximum population size always // smaller population has better solutions
wins the race and runs until maximum number of found_best_ever_solutions = true
generations reached. Therefore, the technique does not break // the loop
work for the parameter-less NSGA-II. EndIf
EndFor
increase Counter
Until found_best_ever_solutions or all populations converge
Print out the solutions
Again, the code to find better solutions has never
reached because the larger populations always catch up
with the smaller ones with more solutions in the first non- carried out and it produces the same result as depicted in
dominated front. However, in this modified version, the Figure 1. Five more experiments are performed with
parameter-less NSGA-II runs each population until it different settings for population size and maximum number
converges according to the modified convergent property of generations. Table 2 presents the summary of these
above and it prints the solutions to files. The algorithm runs experimental results.
until it finishes with the last population (population with the Table 2
maximum population size), or the user is happy with the Summary of experimental results with NSGA-II on
solutions quality and stops the program. This approach the constrained multi-objective test problem used
mimics what the user typically does in trial and error in this study (pc=0.9; pm=0.5; ηc=20; ηm=100)
experiments to find the optimal parameter settings, except Population Max. Comments Plotting
Size (N) Number of
this time, the parameter-less NSGA-II is doing it in a Generations
systematic way. 100 500 Wide spread of solutions Fig. 1
As population size is increased and max.
500 500 number of generations is constant, performance Fig. 2
4. Test Problem declines.
As population size is increased and max.
1500 500 number of generations is constant, performance Fig. 3
The test problem used in this study is a constrained declines.
multi-objective problem which was originally suggested by As population size is increased and max.
Tanaka, Watanabe, Furukawa, & Tantrio [22] and used by 1500 1500 number of generations is also increased, NSGA- Fig. 4
II is able to find good solutions.
Deb et al. [6] to test the NSGA-II. The Pareto-optimal With larger number of generations yields
100 1500 Fig. 5
region for this problem is discontinuous and falls entirely better solutions.
the first constraint boundary [6]. The results for this test Number of generations is larger than needed,
100 3000 Fig. 6
thus computing resource is wasted.
problem are published in both papers. The problem is given
in the Table 1 below.
Table 1
The constrained multi-objective test problem used
in this study
Number Variable Minimize Constraints
of bounds Objective
Variables functions
(m = 2)
x
g 1 ( x ) = − x 12 − x 22 + 1 + 0 . 1 cos( 16 arctan ) ≤ 0
xi ∈ [0, f1(x) = x1 y
2
π] f2(x) = x2
g 2 ( x ) = ( x1 − 0 .5 ) + ( x 2 − 0 .5 ) ≤ 0 .5
2 2

5. Experimental Results Figure 1. Obtained non-dominated solutions with


NSGA-II on the test problem (parameter settings:
This section presents the experimental results and N=100; pc=0.9; pm=0.5; ηc=20; ηm=100; max.
comments on the performance of the parameter-less generations=500).
NSGA-II on the test problem. First, the experiments are
performed on the original NSGA-II in order to understand
the behavior of this algorithm on the test problem. Second,
the experiments are performed on the parameter-less
NSGA-II. Finally, the section presents comments on the
performance of the parameter-less NSGA-II against the
original NSGA-II.

5.1. Experimental Results with the Original


NSGA-II

This section presents the experiments performed on the


original NSGA-II in order to understand the behavior of
this algorithm on the test problem. Deb et al. [6] use a Figure 2. Obtained non-dominated solutions with
population size (N) of 100, crossover probability (pc) of 0.9, NSGA-II on the test problem (parameter settings:
mutation probability (pm) of 0.5, distribution indices for N=500; pc=0.9; pm=0.5; ηc=20; ηm=100; max.
real-coded crossover and mutation operators (ηc and ηm) of generations=500).
20 and 100, respectively, and run the NSGA-II (real-coded)
for a maximum of 500 generations. The same experiment is
Figure 3. Obtained non-dominated solutions with Figure 6. Obtained non-dominated solutions with
NSGA-II on the test problem (parameter settings: NSGA-II on the test problem (parameter settings:
N=1500; pc=0.9; pm=0.5; ηc=20; ηm=100; max. N=100; pc=0.9; pm=0.5; ηc=20; ηm=100; max.
generations=500). generations=3000).

5.2 Experimental Results with Parameter-less


NSGA-II

The experiments with the parameter-less NSGA-II run


multiple populations simultaneously with the initial
population size is set to 25 and maximum population size is
set to 500. Each population has a maximum number of
generations equals to 10*N. These settings are coded
internally within the algorithm. Each population runs until
it converges. Table 3 presents the summary of these
experimental results.
Table 3
Figure 4. Obtained non-dominated solutions with Summary of experimental results with parameter-
NSGA-II on the test problem (parameter settings: less NSGA-II on the constrained multi-objective
N=1500; pc=0.9; pm=0.5; ηc=20; ηm=100; max. test problem used in this study
generations=1500). Population Max. Comments Plotting
Size (N) Number of
Generations
The population size is too small for a
25 250 complex problem therefore the algorithm Fig. 7
cannot find the solutions
The performance starts to improve as population
50 500 Fig. 8
size is increased
Good enough solutions are found. The NSGA-II
100 1000 obtains better solutions with the same population Fig. 9
size running for 500 generations
Solutions are refined but two boundary solutions
200 2000 fall within constraint boundary of value less than Fig. 10
1
Solutions are more refined but two boundary
solutions fall within constraint boundary of
400 4000 Fig. 11
value less than 1. The GA wastes computational
resources
Solutions are more refined with one boundary
Figure 5. Obtained non-dominated solutions with 500 5000
solution falls within constraint boundary of
Fig. 12
NSGA-II on the test problem (parameter settings: value less than 1. The GA wastes computational
resources
N=100; pc=0.9; pm=0.5; ηc=20; ηm=100; max.
generations=1500).
Figure 7. Obtained non-dominated solutions with Figure 10. Obtained non-dominated solutions with
parameter-less NSGA-II on the test problem (when parameter-less NSGA-II on the test problem (when
N=25 is reached at max. generations=250). N=200 is reached at max. generations=2000).

Figure 8. Obtained non-dominated solutions with Figure 11. Obtained non-dominated solutions with
parameter-less NSGA-II on the test problem (when parameter-less NSGA-II on the test problem (when
N=50 is reached at max. generations=500). N=400 is reached at max. generations=4000).

Figure 9. Obtained non-dominated solutions with Figure 12. Obtained non-dominated solutions with
parameter-less NSGA-II on the test problem (when Parameter-less NSGA-II on the test problem (when
N=100 is reached at max. generations=1000). N=500 is reached at max. generations=5000).

5.3 Comments on Experimental Results

The parameter-less NSGA-II performs quite well


compare to the original NSGA-II. The parameter-less
NSGA-II is able to obtain a good set of 100 non-dominated
solutions after 1000 generations while the original NSGA-
II can find better distribution set of solutions after 500
generations. The boundary solutions found by the
parameter-less NSGA-II have a tendency to fall within
constraint boundary of value less than 1 because the
algorithm ignores mutation. During the latter stages of the determine the better solutions between two populations in
run, it is likely that most solutions in the population lie in order to terminate with a proper population size.
the first non-dominated front; therefore, a small mutation Another improvement would be integrating mutation
probability could be use to maintain diversity in the into the parameter-less NSGA-II adaptively to get better
population. The parameter-less NSGA-II has been tested results. For example, use high mutation probability for
with a small mutation probability (0.001) and it yields smaller populations and use very low mutation probability
better results, which the boundary solutions found spread when there are N solutions in the first non-dominated front.
beyond the constraint boundary of value greater than 1.
It has been observed that the distribution of solutions 7. Summary and Conclusion
gets better with generations. Deb [1] points out that when
the entire population converges to the Pareto-optimal front, This paper presented the development of the parameter-
the continuation of crowding distance sorting ensures a less NSGA-II that removes the need of parameters to make
better spread among the solutions. Therefore, both the GAs easier to use and to free users from the task of
original NSGA-II and the parameter-less NSGA-II run a selecting and tuning GA parameters. There has been
population until the maximum number of generations increasing interest in the use of GAs in multi-objective
reached. This approach introduces genetic drift when the optimization recently because a GA is capable of finding
population has already converged to Pareto-optimal front multiple optimum solutions in one single run. However,
with diversity of non-dominated solutions. Genetic drift GA's are not easy to use because they require parameter
occurs when the algorithm cannot discriminate between tunings in order to achieve the desirable solutions. There
two or more distinct solutions. The roaming around of the has been a lot of research related to the topic of parameter
population is called drift and this condition will cause the setting in GAs. This research integrates the parameter-less
algorithm to waste unnecessary computational resources. GA approach into a successful MOEA named NSGA-II.
The parameter-less NSGA-II is a crossover-based multi-
6. Improvements objective based GA which use a crossover probability pc =
0.5. This setting enables the parameter-less NSGA-II works
It is necessary to stop the population from running as more or less well on both easy and difficult problems. The
soon as the population converges to Pareto-optimal front elitism approach used in the parameter-less NSGA-II also
with diversity of non-dominated solutions. This would ignores the selection rate. Regarding the population size,
prevent the algorithm from wasting unnecessary the first attempt is to let the parameter-less NSGA-II
computational resources. To achieve this a performance executes multiple populations with different population
metric can be integrated into the parameter-less NSGA-II to sizes simultaneously with a constant maximum number of
measure the quality of the obtained solutions. There are two generation equals to 500 with hope that the smaller
different measures: measure the extent of convergence to a populations would converge faster than the larger one and
known set of Pareto-optimal solutions and measure the the algorithm would terminate with a proper population
extent of diversity among the obtained solutions [6]. The size. However, it turns out that this concept does not work
first metric requires a known set of Pareto-optimal quite well for MOEAs because most MOEAs assign each
solutions therefore it cannot be used for any arbitrary solution a fitness equal to its a non-domination rank and all
problem. Deb et al. [6] use the following metric to compute solutions in the first non-dominated front have the same
the non-uniformity in the distribution: fitness. Hence, there is no discrimination to any of these
_ solutions. Moreover, it has been observed that the larger

N −1
d + dl + |di − d | populations always catch up with the smaller ones with
∆ = i =1
f
_
d f + d l + ( N − 1) d more solutions in the first non-dominated front. The second
attempt makes a compromise by letting the parameter-less
(1)
NSGA-II executes multiple populations with different
Here, the parameter df and dl are the Euclidean distance
population sizes simultaneously with increasing maximum
between the extreme solutions and the boundary solutions
number of generations proportionately with its population
of the obtained non-dominated set. The parameter d is the size (10*N). This would allow larger populations enough
average of all distance di, where i = 1, 2, … , (N – 1), number of generations to converge to quality solutions. The
assuming that there are N solutions on the first non- parameter-less NSGA-II runs each population until it
dominated front. For the most widely and uniformly spread converges and it prints the solutions to files, or the user is
out set of non-dominated solutions, the value of ∆ is zero. happy with the solutions quality and stops the program.
According to Deb et al., this diversity metric can be used on This approach mimics what the user typically does in trial
any set of solutions. Therefore, the parameter-less NSGA-II and error experiments to find the optimal parameter
can use this metric to measure the quality of the obtained settings, except this time, the parameter-less NSGA-II is
solutions. If the value of ∆ falls within a certain threshold doing it in a systematic way.
the algorithm can stop running the population. Likewise, The parameter-less NSGA-II performs quite well
the parameter-less NSGA-II can use the metric to compare to the original NSGA-II on a difficult test problem
borrowed from the literature. As expected, it is a bit slower [10] G. R. Harik and F. G. Lobo, "A parameter-less genetic
than the original NSGA-II. However, it takes trial and error algorithm," presented at In Banzhaf, W., Daida, J., Eiben, A.
experiments to obtain the optimal parameter settings for an E., Garzon, M. H., Honavar, V., Jakiela, M., & Smith, R. E.
arbitrary real-world problem. Two improvements are (Eds.), GECCO-99: Proceedings of the Genetic and
Evolutionary Computation Conference. San Francisco, CA:
proposed. A performance metric can be integrated into the Morgan Kaufmann., 1999.
parameter-less NSGA-II to measure the quality of the
obtained solutions. The algorithm can use the metric to [11] F. G. Lobo, "The parameter-less genetic algorithm: Rational
prevent genetic drift and to determine the better solutions and automated parameter selection for simplified genetic
between two populations in order to terminate with a proper algorithm operation," Universidale Nova de Lisboa, Portugal,
population size. Mutation can be integrated into the 2000.
algorithm adaptively to get better results.
[12] K. A. De Jong, "An analysis of the behavior of a class of
8. References genetic adaptive systems." Ann Arbor: University of
Michigan, 1975.

[1] K. Deb, Multi-Objective Optimization Using Evolutionary [13] D. E. Goldberg, K. Deb, and J. H. Clark, "Genetic
Algorithms: John Wiley & Sons, 2001. algorithms, noise, and the sizing of populations," Complex
Systems, vol. 6, pp. 333-362, 1992.
[2] C. A. Coello, "An Updated Survey of GA-Based
Multiobjective Optimization Techniques," ACM Computing [14] D. E. Goldberg, K. Deb, and D. Thierens, "Toward a better
Surveys, vol. 32, 2000. understanding of mixing in genetic algorithms," Journal of
the Society of Instrument and Control Engineers, vol. 32, pp.
[3] A. Osyczka, "Multicriteria optimization for engineering 10-16, 1993.
design," in Design Optimization. New York, NY: Academic
Press, Inc., 1985, pp. 193-227. [15] W. M. Spears, "Adapting crossover in a genetic algorithm,"
presented at Proceedings of the 4 th International Conference
[4] C. A. Coello, "A Short Tutorial on Evolutionary on Genetic Algorithms, 1991.
Multiobjective Optimization," presented at First International
Conference on Evolutionary Multi-Criterion Optimization, [16] R. E. Smith and E. Smuda, "Adaptively resizing populations:
2001. Algorithm, analysis, and first results," Complex Systems, vol.
9, pp. 47-72, 1995.
[5] C. A. Coello and G. T. Pulido, "Multiobjective Optimization
using a Micro-Genetic Algorithm," presented at Proceedings [17] J. H. Holland, Adaptation in Natural and Artificial Systems,
of the Genetic and Evolutionary Computation Conference 2nd ed. Ann Arbor: MI: University of Michigan Press, 1975.
(GECCO'2001), San Francisco, California, 2001.
[18] D. E. Goldberg, Genetic Algorithms in Search, Optimization
[6] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, "A Fast and Machine Learning. Reading, MA: Addison-Wesley,
and Elitist Multiobjective Genetic Algorithm: NSGA-II," 1989.
IEEE Transactions on Evolutionary Computation, vol. 6, pp.
182-197, 2002.
[19] N. Srinivas and K. Deb, "Multiple objective optimization
using nondominated sorting in genetic algorithms,"
[7] D. Thierens and P. A. N. Bosman, "Multi-Objective Mixture- Evolutionary Computation, vol. 2, pp. 221-248, 1994.
based Iterated Density Estimation Evolutionary Algorithms,"
presented at Proceedings of the Genetic and Evolutionary
[20] V. Khare, X. Yao, and K. Deb, "Performance Scaling of
Computation Conference (GECCO'2001), San Francisco,
Multi-objective Evolutionary Algorithms," School of
California, 2001.
Computer Science, The University of Birmingham,
Edgbaston 2002009, 2002.
[8] K. Deb and S. Agrawal, "Understanding interactions among
genetic algorithm parameters," In Foundations of Genetic
[21] K. Krishnakumar, "Microgenetic algorithms for stationary
Algorithms 5, pp. 265-286, 1998.
and non-stationary function optimization," presented at SPIE
Proceedings on Intelligent Control and Adaptive Systems,
[9] G. R. Harik, E. Cantu−Paz, D. E. Goldberg, and B. L. Miller, 1989.
"The gambler's ruin problem, genetic algorithms, and the
sizing of populations," presented at Proceeding of the Fourth
[22] M. Tanaka, H. Watanabe, Y. Furukawa, and T. Tantrio, "GA-
International Conference on Evolutionary Computation, New
based decision support system for multi-criteria,
York, 1997.
optimization," presented at Proceedings of the International
Conference on Systems, Man and Cybernetics-2, 1995.

You might also like