Anda di halaman 1dari 6

R.Sivaraj et al.

/ International Journal of Engineering Science and Technology (IJEST)

A REVIEW OF SELECTION METHODS


IN GENETIC ALGORITHM
R.SIVARAJ
Research Scholar, Assistant Professor (Senior Grade),
Department Of Computer Science and Engineering,
Velalar College of Engineering and Technology,
Erode, Tamil Nadu, India.
Dr.T.RAVICHANDRAN
Principal and Research Supervisor,
Hindusthan Institute of Technology
Coimbatore-641032, Tamil Nadu, India.

Abstract:
Genetic Algorithms are optimization algorithms that maximize or minimize a given function. Selection operator
deserves a special position in Genetic algorithm since it is the one which mainly determines the evolutionary
search spaces. It is used to improve the chances of the survival of the fittest individuals. There are many
traditional selection mechanisms used and many user specified selection mechanisms specific to the problem
definition. This paper gives an overview of the selection mechanisms commonly available and that are used.
Keywords: Genetic Algorithms; selection mechanisms; survival

1. Introduction
Genetic Algorithms (GAs) are adaptive heuristic search algorithm based on the evolutionary ideas of natural
selection and genetics. They represent an intelligent exploitation of a random search used to solve optimization
problems. Although randomized, GAs are by no means random, instead they exploit historical information to
direct the search into the region of better performance within the search space. The basic techniques of the GAs
are designed to simulate processes in natural systems necessary for evolution; especially those follow the
Charles Darwins principles of "Survival of the Fittest. This principle is similar to competition among
individuals for scanty resources in nature results in the fittest individuals dominating over the weaker ones.
It is better than conventional AI in that it is more robust. Unlike older AI systems, GAs does not break easily
even if the inputs changed slightly, or in the presence of reasonable noise. Also, in searching a large state-space,
multi-modal state-space, or n-dimensional surface, a genetic algorithm may offer significant benefits over more
typical search of optimization techniques. (Linear programming, heuristic, depth-first, breath-first, and praxis.)

2. Genetic Algorithms Overview

GAs simulates the survival of the fittest individuals among all the individuals in the population over successive
generations for solving a problem. Each generation consists of a population of character strings that are
analogous to the chromosome that are seen in DNA. Each individual represents a point in a search space and a
possible solution. The individuals in the population are then made to go through a process of evolution. GAs are
based on an analogy with the genetic structure and behavior of chromosomes within a population of individuals
using the following foundations: Individuals in a population compete for resources and mates. Those individuals
most successful in each 'competition' will produce more offspring than those individuals that perform poorly.
Genes from `good' individuals propagate throughout the population so that two good parents will sometimes
produce offspring that are better than either parent. Thus each successive generation will become more suited to
their environment [Goldberg (2003)].

2.1. Selection Operator

The key idea of selection operator is to give preference to better individuals (those that are nearer to the
solution) by allowing them to pass on their genes to the next generation and prohibit the entrance of worst fit
individuals into the next generations. The selection operator mainly works at the level of chromosomes. The

ISSN : 0975-5462 Vol. 3 No. 5 May 2011 3792


R.Sivaraj et al. / International Journal of Engineering Science and Technology (IJEST)

goodness of each individual depends on its fitness. Fitness value may be determined by an objective function or
by a subjective judgement specific to the problem. As the generations pass, the members of the population
should get fitter and fitter (i.e. closer and closer to the solution.)

2.2. Selection mechanisms

A critical parameter to be improved in GA is the selection pressure which is the process of selecting the best
individuals for the next generation. If it is set too low, then the rate of convergence towards the optimum
solution is too low. If the selection pressure is set too high, the system is likely to be stuck in a local optimum
due to the loss of diversity in the population. Thus the selection methods controls the selection pressure, which
in turn determines how fast the algorithms coverage. The speed of convergence of different selection schemes
was first studied by Goldberg and Deb [Goldberg and Deb (1990)]. The selection mechanism should be chosen
such that it converges to the global optimum solution by without being caught into local optimum and also
should encompass the knowledge of the existing data.
The general classification of genetic algorithm is as follows:

1. Traditional mechanisms:
o Proportionate Selection methods
Roulette Wheel selection
Deterministic Sampling
Stochastic Remainder Sampling
Stochastic Remainder selection with replacement
Stochastic remainder selection without replacement
Stochastic universal selection
o Ranking Selection
Linear Ranking selection
Truncate selection
o Tournament Selection
Binary Tournament Selection
Larger Tournament Selection
Boltzmann tournament select ion
Correlative Tournament Selection
2. Alternative selection mechanisms
o Range selection
o Gender-Specific Selection
Genetic algorithm with chromosome differentiation. (GACD)
Restricted mating
Coorelative Family- based selection
o GR based selection
Fitness Uniform selection scheme (FUSS)
Reserve selection

2.3. Proportionate Selection methods

2.3.1. Roulette Wheel selection


In the Roulette wheel selection method [Holland, 1992], the first step is to calculate the cumulative fitness of the
whole population through the sum of the fitness of all individuals. After that, the probability of selection `is
calculated for each individual as being pseli = fi/Pfi. Then, an array is built containing cumulative probabilities
of the individuals. So, n random numbers are generated in the range 0 to Pfi and for each random number an
array element which can have higher value is searched for. Therefore, individuals are selected according to their
probabilities of selection.

2.3.2. Deterministic sampling


In the Deterministic Sampling, the average fitness of the population Pselecti is calculated as the fitness
associated to each individual divided by the average fitness but only the integer part of this operation is used to
select individuals for the next generation population. The whole fraction is used to sort the individuals. The
remaining free places in the new population will be fulfilled with individuals which have greatest fraction value
in the sorted list.

ISSN : 0975-5462 Vol. 3 No. 5 May 2011 3793


R.Sivaraj et al. / International Journal of Engineering Science and Technology (IJEST)

2.3.3. Stochastic Remainder Sampling


The Stochastic Remainder Sampling has identical concepts used in the Deterministic Sampling and the
population must be formed with the integer part of the expression result api/apavg. In this case, free places were
filled based on the Roulette method. In Remainder Selection Mechanism, expected number of copies of a string
is calculated as mi = fi/f. It assigns parents deterministically from the integer part of each individual's scaled
value (mi), and then uses roulette selection on the remaining fractional part. For example, if the scaled value of
an individual is 2.3, that individual is listed twice as a parent because the integer part is 2. After parents have
been assigned according to the integer parts of the scaled values, the rest of the parents are chosen
stochastically. The probability that a parent is chosen in this step is proportional to the fractional part of its
scaled value.

2.3.4. Stochastic Remainder sampling with replacement


If done with replacement, the remainders are used to size the slots of a roulette wheel selection process. The
algorithm when performed with replacement takes on the complexity of the roulette wheel, because O (n) of the
individuals is likely to have fractional parts to their m values.

2.3.5. Stochastic Remainder sampling without replacement


If done without replacement, each remainder is used to bias the flip of a coin that determines whether the part
receives another copy or not. The algorithm without replacement is O(n), because the deterministic assignment
requires only a single pass (after the calculation of f, which is also O( n), and the probabilistic assignment is
likely to terminate in 0(1) steps. On the other hand, the algorithm when performed with replacement takes on the
complexity of the roulette wheel, because O (n) of the individuals are likely to have fractional parts to their m
values.

2.3.6. Stochastic universal selection


Stochastic universal selection is performed by sizing the slots of a weighted roulette wheel, placing equally
spaced markers along the outside of the wheel, and spinning the wheel once; the number of copies an individual
receives is then calculated by counting the number of markers that fall in its slot. The algorithm is O (n),
because only a single pass is needed through the list after the sum of the function values is calculated

2.4. Ranking Selection

2.4.1. Linear Ranking selection


In linear ranking selection (Baker, 1985), first the individuals are ranked according to the fitness values. Those
with high fitness values will be ranked high and those with low fitness values will eventually have lower ranks.
Then the individuals are selected with a probability that is linearly proportional to the rank of the individuals in
the population.

2.4.2. Truncate Selection

In truncation selection the candidate solutions are ordered by fitness, and some proportion, p, (e.g. p=1/2, 1/3,
etc.), of the fittest individuals are selected and reproduced 1/p times. Truncation selection is less sophisticated
than many other selection methods, and is not often used in practice.

2.5. Tournament Selection


2.5.1. Binary Tournament selection

In this variant, two individuals are chosen at random and the better of the two individuals is selected with fixed
probability p, 0.5 < p ~ 1.

2.5.2. Larger Tournament selection


To analyze the performance of tournament selection with any size tournament, it is easier to consider the
doughnut hole (the complementary proportion) rather than the doughnut itself (the primary proportion).
Considering a deterministic tournament3 of size s and focusing on the complementary proportion Qi, a single

ISSN : 0975-5462 Vol. 3 No. 5 May 2011 3794


R.Sivaraj et al. / International Journal of Engineering Science and Technology (IJEST)

copy will be made of an individual in this class only when all s individuals in a competition are selected from
this same lowly group.

2.5.3. Boltzmann tournament selection


Boltzmann tournament selection (BTS) is a select ion procedure for genetic algorithms that is motivated by
simulated annealing. BTS maintains diversity in a populate ion by sampling in such a way that the samples, over
time, become Boltzmann dist ributed. However, the factor of genetic drift works against BTS by limiting the
amount of diversity it can maintain. A population contains n individuals. In each generation, a three-way
competition or "tournament" is held for each population slot. First individual is chosen uniformly at random,
second individual with an objective function value different from the first by a threshold amount x.Half of the
time, choose a third individual with an objective function value different from the first and second individuals
by the threshold amount (this is called strict choice), and the other half of the time, choose the third individual
with an objective function value different by the threshold amount from the first individual alone (this is called
relaxed choice).If the algorithm is successful, it will create a steady proportion of individuals according to the
Boltzmann distribution. If a uniform random sample is drawn from a Boltzmann -distributed population, it will
be biased toward better individuals.

2.5.4. Correlative Tournament Selection


It is an extension of the ordinary tournament selection. Instead of selecting two parents randomly for crossover,
this method selects a pair of parents which are highly correlative in nature thereby realizing an effective
crossover. This is similar to the Seduction method proposed by, but the selection of the second parent is closely
dependent on the first parent unlike the former where the seduce function is independent of the first parent. The
hamming distance is used as the correlation function that is the similar individuals have a correlation value to
each other. Many works on Tournament selection has been discussed in (Goldberg (1989), Muhlenbein (1989)].

2.6. Range selection


Each individual in the population is assigned a numerical rank based on their fitness, and selection is based on
these ranking rather than absolute differences in fitness. The advantage of this method is that it can prevent very
fit individuals to gain dominance at first at the expense of the less fit, which would reduce the genetic diversity
of the population and could hamper the search for an acceptable solution.

2.7. Gender-Specific Selection


The inspiration for this type of selection comes from the concept of sexual selection in the population genetics
where the process of choosing a mating partner is modulated differently for male and female individuals. Sexual
selection uses two different selection schemes for the selection of the two parents required for each crossover.
So it is possible to simulate the concept of male vigor and female choice by using random selection as the first
selection schemes and another selection strategy with far more selection pressure as the second one. The second
selection can be Roulette Wheel selection or linear rank selection. By using two different selection concepts
simultaneously a GA user can influence the selection pressure level of a GA run more precisely.

2.7.1 Genetic algorithm with chromosome differentiation.(GACD)


Nature generally differentiates the individuals in the species into more than one class. Sexual differentiation is
the difficult example where the individuals of a species generally belong to either male or female classes. The
prevalence of this form of differentiation indicates an associated advantage which appears to be in terms of
cooperation between two dissimilar individuals who can at the same time specialize in the own fields. This
cooperation and specialization has lead to the rise of healthier and fit offspring. In order to make the two
populations most dissimilar these class of populations are initially generated in such a way that the Hamming
distance between them is maximized. Crossover is allowed between individuals belonging to the distinct
population only.

2.7.2 Restricted mating


A species, according to the biological definition is a class of organisms that are capable of interbreeding
amongst them, but that typically do not breed with individuals outside their class as said by Cook in. This
concept was exploited in genetic algorithms by researchers like Booker where individuals were allowed to mate
only with individuals with a common tag representing some similarities in their genes.

ISSN : 0975-5462 Vol. 3 No. 5 May 2011 3795


R.Sivaraj et al. / International Journal of Engineering Science and Technology (IJEST)

2.7.3 Correlative Family- based selection


In this method, an individual xi that has the highest fitness in each family is selected. Family here implies two
parents and their offsprings. The hamming distance between the individuals and the other three individuals in
the family are calculated. The one with the maximum hamming distance is chosen for the survival to the next
generation along with xi. This is considered to maintain the population diversity of genotype as well as select
individuals with higher fitness value.

2.8. GR based selection


The proposed selection schemes are performed in the environment where all individuals are endowed with
memory to remember the value of its best related individuals. The concept of genetic relatedness is introduced
to define the relationship existing between individuals that have a common best ancestor. An ancestor is an
individual that has existed in one of the previous generations and has contributed to the creation of the current
one: The parents and the parents of the parents and so on. Here two such operators have been proposed. One is
the GR based proportional selection and the other is the GR based Tournament selection.

2.8.1 Fitness Uniform selection scheme (FUSS)


It is well suited for difficult optimization problems as it focuses the selection pressure towards fitness levels
which have relatively few individuals rather than on the higher fitness levels. It is based on the concept of the
preserving genetic diversity in the population by using the fitness of individuals to estimate their similarity. The
similarity is found by calculating the distance between two individuals I and j with fitness f(i) and f(j) and is
defined as To evolve a population after having calculated the fitness values for the members of the current
population. A uniform random number is chosen in the interval (f min, f max) where f min and f max are the
maximum and minimum fitness values in the current population. Then the fitness nearest to this number is
chosen to be the fitness level for this generation. The individuals for the next generations are selected randomly
from each of the fitness level.

2.8.2 Reserve selection


This selection mechanism is based on the technique called population segmentation a process of splitting into
a Non-reserved Area and a Reserved area [Chen et al(2007)]. The Non-reserved area consists of high fit
individuals which are subjected to normal applications of operators such as elite selection, tournament selection,
crossover and mutation operations and it focus on the exploitation of the search space. The Reserved area
consists of low fit individuals which focus on the exploitation of the search space as it performs mutation on the
low fit individuals, aiding the algorithm to escape from the local optima.

3. Improvement techniques in selection mechanisms

3.1 Pecking order


The individuals go through multiple rounds of selection in each generation. Evaluations of the first levels are
faster and less discriminatory, while those who survive to higher levels are evaluated more rigorously. The
advantage of this method is that it reduces the total time of calculation by using a more rapid and less selective
to eliminate most of the individuals who show little or no promise, and undergoing a rigorous fitness evaluation
and computationally more costly only to those who survive this initial test.

3.2 Selection generation


The offspring of selected individuals in each generation becomes all the next generation. Individuals are not
retained between generations.

3.3 Selection climbing


To increase the average fitness of the population, the strength of selective pressure is increased and the fitness
function becomes more discriminating. This method may be useful in selecting later, when all individuals have a
relatively high fitness and only small differences distinguish them in fitness.

4. Discussion

GAs that use selection alone are incapable of generating solutions that are not currently in the population.
Aggravating this situation is the fact that some good solutions in the population eventually disappear due to the
variance of the selection process. In fact, if the typical selection scheme is run long enough, all but one solution

ISSN : 0975-5462 Vol. 3 No. 5 May 2011 3796


R.Sivaraj et al. / International Journal of Engineering Science and Technology (IJEST)

will disappear, even without selection pressure (high temperatures in BTS result in little to no selection
pressure) which is called as genetic drift.
As a selection scheme runs, the number of alternative solutions declines. Of course, one does not run a GA
using selection alone. Crossover and mutation (along with many other operators) are used to explore solutions
not currently present. However, these operators are limited in their abilities to explore the solution space. When
operating on identical or nearly identical strings, crossover will yield results similar or identical to the parent
strings; thus, it cannot be relied up on to reintroduce diversity. Mutation, though it can potentially explore the
full solution space, is also of dubious value when it comes to reintroducing diversity: too high a mutation rate
will lead to a more random search procedure, and random search is of very little use on sufficiently hard
problems.

Variable Population Size


Varying the population size between two successive generations affects only the selection operator of the GA.
Let x denote the population size of the current and the subsequent generation, respectively. The selection of the
individuals can be considered as a repetitive process of selection operations, with being the probability of
selection of the individual. For most of the selection operators, such as fitness proportionate selection and
tournament selection with replacement, the selection probability remains constant for the selection operations.
Therefore, the portion of the population related to the individual after selection is independent of the population
size of the subsequent generation, provided that the variation of the population size is not significant enough to
modify the probability. However, a GA with decreasing population size has bigger initial population size and
smaller final population size, as compared to a constant population size GA with the same computing cost (i.e.,
equal average population size). This is expected to be beneficial, because a bigger population size at the
beginning provides a better initial signal for the GA evolution process; whereas, a smaller population size is
adequate at the end of the run, where the GA converges to the optimum [Grefenstette and Baker (1989)].

5. Conclusion

All of these selection mechanisms have the same purpose of creating more copies of the individuals with higher
fitness than those with lower fitness. However the selection mechanisms differ in the manner in which they
allocate copies to the fittest individuals. A selection method has the higher selection measure than the other if it
makes more copies of the best individuals thereby eliminating low fit individuals rapidly. A strong selection
mechanism reaches equilibrium faster than a weaker method. But it also sacrifices genetic diversity that may be
needed to find an adequate solution. Different selection mechanisms work well under different situations.
Appropriate method has to be chosen for the specific problem to increase the optimality of the solution.

6. References
[1] Baker, J.E. (1985).Adaptive selection methods for Genetic algorithms. In Proceedings of the 1st International Conference on Genetic
Algorithms and their Applications, pp 101-111
[2] Chen, Y., Hirasawa, K., and Yu, S. (2007). GARS. An Improved Genetic Algorithm with Reserve Selection for Global Optimization.
In proceedings of Genetic and Evolutionary Computation Conference (GECCO07), pp 1173-1178
[3] Goldberg D.E.(1990),A Note on Boltzmann Tournament Selection for Genetic Algorithms and Population - Oriented Simulated
Annealing International Conference on Complex Systems, 4, 445-460
[4] Goldberg D.E. and Deb Kalyanmoy (1991), A Comparative Analysis of Selection Schemes Used in Genetic Algorithms,
Foundations of Genetic Algorithms, Morgan Kaufmann, California
[5] Goldberg D.E. (2003), Genetic Algorithms in Search, Optimization, and Machine Learning (Addison Wesley, Reading g, MA,)
[6] Grefenstette, J. J. & Baker, J. E. (1989). How genetic algorithms work : A critical look at implicit parallelism. Proceedings of the
Third International Conference on Genetic Algorithms, 20-27.
[7] Holland, J.H (1992), Adaptation in Natural and Artificial Systems, 2nd Ed, MIT Press
[8] Muhlenbein, H. (1989). Parallel genetic algorithms, population genetics and combinatorial optimization. Proceedings of the Third
International Conference on Genetic Algorithms, 416-421

ISSN : 0975-5462 Vol. 3 No. 5 May 2011 3797

Anda mungkin juga menyukai