You are on page 1of 6

A New Real-valued Diploid Genetic Algorithm for

Optimization in Dynamic Environments

Amineh Omidpour Kamran Alagheband


Department of electronic, Computer and IT Department of Mechanical Engineering
Qazvin Branch, Islamic Azad University Shiraz University
Qazvin, Iran Shiraz, Iran
Omidpour_amineh@yahoo.com Kam_alagheband@yahoo.com

Babak Nasiri Mohammad Reza Meybodi


Department of electronic, Computer and IT Department of Computer engineering and IT
Qazvin Branch, Islamic Azad University Amirkabir University of Technology
Qazvin, Iran Tehran, Iran
Nasiri_babak@yahoo.com mmeybodi@aut.ac.ir

Abstract—Many real-world problems are dynamic, requiring models the process of natural evolution. G.A.’s have been
an optimization algorithm which is able to continuously track a applied to a diverse field of problems with promising results.
changing optimum over the time. Using a diploidy and Most of the approaches found in literature address stationary
dominance is one method to enhance the performance of problems, whereas problems with changing environments
genetic algorithms in dynamic environment. Diploid genetic present a challenging area of research.
algorithm has two chromosomes in each individual. In this
paper, for the first time, a real-valued diploid genetic The main reason why traditional G.A’s don’t perform
algorithm is proposed. Its new dominance mechanism is based well in dynamic environments is that they quickly converge
on a simple function with homogeneous outputs. In addition, a to a solution and the population loses its genetic diversity
new dominance change mechanism is added to the algorithm. consequently, when a change occurs, it is hard for the
Hence, when environment change occurs, it can increase converged population to adapt. One of the main approaches
diversity to respond more quickly to the changes. Other for coping with different types of changing environments is
diploid genetic algorithms in literature are discrete and they memory. Memory may be provided in two general ways.
have never been tested by Moving Peak Benchmark (MPB) Implicitly by using redundant representations, or explicitly
which is continuous and dynamic. For the first time, the by introducing an extra memory and formulating strategies
proposed approach is tested by MPB. Results are compared to deposit and retrieve solutions from it [2-6].
with other diploid genetic algorithms showing that proposed
algorithm significantly outperforms previous approaches. Diploidy and dominance have been studied for
improving genetic algorithms performance in dynamic
Keywords-diploid genetic algorithm; dynamic environment; environments. A diploid genetic algorithm has two
moving peak benchmark. chromosomes in each individual, and has the advantage of
being able to adapt rapidly to changes in the environment.
I. INTRODUCTION
Researchers have developed several genotypic representation
Evolutionary algorithms (EA’s) are a class of meta- and dominance schemes [7,8].
heuristic algorithms inspired by principles of natural
evolution, such as selection and population genetics. While In 1987, Goldberg and Smith, report on experiments
most of the optimization problems discussed in the scientific with using diploidy and dominance. They use a triallelic
literature are static, many real world problems are dynamic, scheme where an allele can take on one of three values “0”,
i.e., they change over time. In such cases the optimization “recessive 1”, and “dominant 1”. In 1995, this approach has
algorithm has to track a moving optimum as closely as been reviewed critically by Ng and Wong. They argued that
possible, rather than just find a single good solution.[1] the triallelic is biased and proposed a new diploid scheme
with four possible alleles (0 and 1, each dominant and
Genetic algorithms (G.A’s) 1 are stochastic, global recessive). In 1997, Ryan uses additive multiploidy. The
optimization methods which model the biological principles phenotypic trait becomes 1 when a certain threshold is
of Darwin’s theory of evolution and Mendelian of exceeded, and is 0 otherwise. In 1998, Lewis et al. observed
inheritance. We can call genetic algorithms as the most that a simple dominance scheme is not sufficient to track the
widespread and universal optimization method, which optimum reasonably well. If the diploid approaches are
extended with a dominance change mechanism, much better
1
978-1-4799-3351-8/14/$31.00 ©2014 IEEE
results can be obtained. They extended additive approach the most famous schemes are Ng-Wong and Additive. Ng-
with dominance change mechanism [5-12]. Wong dominance scheme uses four genotypic alleles:
dominant “1” and “0”, and recessive “i” and “o”. The
In 1998, an interesting comparative study has been dominant allele is always expressed in the phenotype. If
performed by Lewis et. al. They observed that a simple contention exists between two dominant or two recessive
dominance scheme is not sufficient to track the optimum alleles, one is randomly chosen to be expressed. The
reasonably well. If the diploid approaches are extended with occurrence of “1i” or “0o” is prohibited. If it occurs the
a dominance change mechanism, much better results can be recessive gene is promoted to be dominant. The genotype to
obtained [13,14]. phenotype mapping is shown in Fig. (1). Ng and Wong also
All the diploid genetic algorithms being reviewed are incorporated a dominance scheme: when the environment
discrete and they have tested by discrete dynamic change is detected, the dominance values of all allele-pairs
environments like oscillating knapsack. Unlike other diploid are invented such that “11” becomes “ii”, “00” become “oo”,
genetic algorithms, in this paper, we explore a novel real- “10” becomes “1o” becomes “i0”, and vice versa. The
valued diploid genetic algorithm that is especially designed genotype to phenotype mapping keeps unchanged. Additive
for continuous dynamic environment. In order to test dominance scheme uses four genotypic alleles A, B, C and D
different algorithmic approaches to dynamic problems, and allocates the values 2, 3, 7, and 9 to them respectively.
several benchmarks have been proposed in the literature. An addition is performed on the values allocated with the
Among the most used ones are the Moving Peak Benchmark two genotypic alleles for each gene locus. If the sum is
(MPB). In Addition, for the first time, a diploid genetic greater than 10, the phenotypic allele becomes 1; otherwise,
algorithm will be tested by MPB. It makes it possible to it becomes 0. The resulting dominance map is shown in
compare the result of real-valued proposed algorithm with Fig.(2).
other obtained result of discrete ones to observe if they can A B C D
be generalized. A 0 0 0 1
The paper is structured as follows. The next section, B 0 0 0 1
C 0 0 1 1
provide a brief introduction on diploid genetic algorithms
D 1 1 1 1
and their dominance schemes. Section III describes in detail
the proposed real-valued diploid genetic algorithm. In Figure 1. Ng-Wong dominance mechanism
section IV the proposed approach is empirically evaluated on
0 o 1 i
the moving peak benchmark and compared with other
0 0 0 0/1 1
approaches from literature. Finally, section V concludes the o 0 0 1 0/1
paper and some idea for future work. 1 0/1 1 1 1
i 0 0/1 1 1
II. DIPLOID GENETIC ALGORITHM
Figure 2. Additive dominance mechanism
An Evolutionary algorithm that uses representations
containing more information than necessary to define B. Reproduction operations
phenotype, i.e. redundant representations, basically has some
For the second aspect, in diploid genetic algorithms,
memory. This kind of memory is called implicit. Redundant
crossover is divided into two steps. In the first step, two
representations slow down convergence and increase
parents are exchange their chromosomes randomly to create
diversity. The most prominent approach to redundant
two temporary offsprings. Each offspring has one
representations seems to be diploidy. A diploid evolutionary
chromosome from each parent and hence the genotypic
algorithm is usually an algorithm whose chromosomes
materials from the parents are mixed and propagated to the
contain two alleles at each locus.
offsprings.
Genetic algorithms with diploidy representation and
In the second step, two chromosomes in each offspring
dominance schemes are called diploid genetic algorithms.
undergo the uniform crossover operation with the probability
Diploid genetic algorithms differ from traditional genetic
pc, which is the same as in traditional genetic algorithms.
algorithms mainly in two aspects: the representation and
After crossover, each of the two genotypic chromosomes of
evaluation scheme and the reproduction operations.
an offspring is independently subject to a bitwise mutation
A. Representation and evaluation scheme with a probability pm [4-7].
For the first aspect, each individual in diploid genetic III. THE PROPOSED ALGORITHM
algorithms has two genotypic chromosomes. In order to
evaluate an individual we need first map the diploid There are several discrete diploid genetic algorithms
genotype into a haploid phenotype according to some developed in the literature. In this paper, for the first time, a
dominance mechanism. Then, the phenotype is evaluated real-valued diploid genetic algorithm is proposed. It aims to
according to the external environment to give fitness to the overcome the deficiencies of other previously diploid
individual. As mentioned before, genotypic representation approaches and make it possible to compare the result of
and dominance schemes play a significant role in the real-valued proposed algorithm with other discrete ones to
performance and researchers have developed several observe if results can be generalized. In this algorithm, the
genotypic representation and dominance schemes that two of individuals have diploid chromosomes and a global
domination map is used to determine the phenotype of them. The individuals that take part in the reproduction step are
The general pseudocode of proposed algorithm is given in determined using tournament selection.
Fig. (3). The different parts of the algorithm will be
explained in the following parts. 2) Crossover
At first stage of crossover phase, one chromosome from
Proposed Diplid Genetic Algorithm: each parent is randomly selected to be copied into each one
Initialize population P(0) // randomly of the two offspring. At the second stage, each offspring
Dominace Mechanism(P(0)) undergoes an arithmetic crossover that occurs between its
{ for each individual own chromosomes. This ensures that genotypic materials
α i = min( xi , y i ), β i = max( xi , y i ) from the parent are mixed for propagation into the next
⎧if β i ≥ α i + (length / 2) ⇒ β i generations.
phenotypei = ⎨
⎩otherwise ⇒ α i 3) Mutation
end The effect of uniform mutation is the same as in standard
} genetic algorithms. At this part, mutation acts directly on the
Evaluate population P(0) genotype, i.e. on each gene of the two strings of
Repeat chromosomes. Each mutation at each location is independent
P’(t)=Select for Reproduction(P(t)) // Tournament selection of those in other locations and the probability of its
Crossover(P’(t),Pc) // Pc is the crossover probability occurrence is the same for all genes.
Mutate( P’(t),Pm) // Pm is the mutation probability
Dominace Mechanism(P’(t)) 4) Dominance Mechanism
{for each offspring Since the phenotype of the individual determines the
α i = min( xi , y i ), β i = max( xi , y i ) actual characteristics and the fitness, mapping the genotype
to the phenotype, play a crucial role in the performance of
⎧if β i ≥ α i + (length / 2) ⇒ β i diploid genetic algorithm. Subsequently, there has been some
phenotypei = ⎨ research in this area as outlined in the previous sections.
⎩otherwise ⇒ α i
end In this study, instead of strictly dominant or recessive
} discrete alleles, for the first time, continuous ones are
Evaluate offsprings(P’(t)) considered. They are calculated based on a function with
Replacement(P’(t),P(t)) // Elitism of size 1 homogeneous outputs. In the other words, the method
If ( Environment Change=true) involves a function that makes the probability of occurrence
Dominance Change Mechanism(P(t)) any real-valued numbers in the interval [L,U] equal. One of
For individuals whose fitness have gotten 20% worse the main advantages of this function is its simplicity. For
{ determining the ith location phenotype, the genotype
genotypei ,t = genotypei ,t −1 + (Rand (−1,1) × Severity) elements corresponding to that location are considered as the
} inputs of the function. Its output, express the allele of
end phenotype corresponding to that location. This function is
end explained in more details as follows:
End if terminated = true
Initially, the minimum and maximum of two alleles
Figure 3. Pseudocode of proposed diploid genetic algorithm
corresponding to the ith location are defined. Next, based on
A. Initialization and evaluation the function condition, one of them is chosen to be expressed
as the allele of phenotype regarding to that location:
In this algorithm, a diploid representation for individuals
is used and each individual has two chromosomes. In the - If the maximum is greater than
initialization step, individuals are initialized with random [minimum+(length/2)], then the maximum is chosen,
real-valued numbers in the interval [L,U]. The L and U are otherwise, the minimum is selected for the
lower and upper bound of real-valued numbers respectively. phenotype.
Choosing the value of L and U is arbitrary and depends on Note that length is range of alleles, i.e. (U-L). As
the problem. There is a global domination mechanism being mentioned before, L and U are lower and upper bound of
used to determine the phenotype of them that will be real-valued alleles respectively and they are arbitrary
discussed in more detail in the following sections. The chosen.
fitness value is calculated based on the phenotype of the
individual. The proposed function of dominance mechanism is given
in Esq. (1).
B. Reproduction offsprings
α i = min (xi , yi )
The loop runs for a number of predefined generations and β i = max(xi , y i )
is composed mainly of selection, reproduction, dominance (1)
mechanism and dominance change mechanism steps. ⎧if β i ≥ α i + (length / 2) ⇒ β i
phenotypei = ⎨
1) Selection ⎩otherwise ⇒ α i
The xi and yi represent two alleles of ith location in environment change, all individuals whose fitness values
genotypes. The α i and β i depict the minimum and have gotten worse by a threshold percentage 20% between
maximum of the two corresponding alleles of ith location (xi successive evaluation cycles are determined. Then they
and yi) respectively. uniformly and randomly distributed in the vicinity of
genotypic alleles before change. After the environment
To demonstrate how the alleles of phenotype are change, the new genotypic alleles of individual i in the t
calculated, consider an example. Assume that each iteration, can be calculated by Eq. (2):
individual from the current population has two chromosomes
with 5 genes and the value of alleles is in the interval [0,100] genotype i ,t = genotype i ,t −1 + (Rand ( −1,1) × Severity ) (2)
(Fig. (4)). For this example, phenotype of the individual
becomes: (83.698 …, 28.005 …, 79.359 …, 42.741 …, in which Rand function produces a vector of random
89.396 …) (Fig. (5)). The required calculations are shown in numbers with a uniform distribution in the range of [-1,1]
table (I). and severity is the shift length of peaks in the changing
environment.
83.698 … 41.029 … 11.082 … 42.741 … 93.657 …
20.443 … 28.005 … 79.359 … 53.817 … 89.396 … IV. PARAMETERE SETTING AND RESULTS
Figure 4. An example of an individual’s real-valued genotypes To assess correctness and efficiency of the proposed
algorithm, it is tested on Moving Peak Benchmark (MPB)
83.698 … 28.005 … 79.359 … 42.741 … 89.369 … and compared with various known diploid genetic
Figure 5. An example of an individual’s real-valued phenotype algorithms. MPB is one of the most famous benchmarks of
dynamic environments and it is the first time that MPB is
TABLE I. AN EXAMPLE OF CALCULATION OF used for testing diploid genetic algorithms. Experiments
PHENOTYPE
have been done with respect to MPB parameters being given
Calculation phenotypei in table (II) [15,16].
83.698 …>=(20.443 …+50) 83.698 … (Max)
41.029 …<(28.005 …+50) 28.005 … (Min) TABLE II. MOVING PEAK PARAMETER SETTING
79.359 …>=(11.082 …+50) 79.359 … (Max)
53.817 …<(42.741 …+50) 42.741 … (Min) Parameter Value
93.657 …<(89.396 …+50) 89.396 … (Min) Number of peaks Variable between 1 to 100
After using dominance mechanism, based on obtained Change frequency Variable between 500 to 10000
Height change 7.0
phenotypes, the fitness values of offsprings are calculated. Width change 1.0
5) Replacement Peaks shape Cone
Basic function No
For replacement, consider union of produced offsprings Shift length 1.0
and the current population. Next, copy the elites from union Number of dimension 5
for the next generation as a population. Correlation coefficient 0
Peaks location range [0-100]
6) Detecting Change Peak height [30.0-70.0]
Detecting changes by re-evaluating solutions is by far the Peak width [1-12]
most common change-detection approach. To detect Initial value of peaks 50.0
environment change in the proposed algorithm, a point with To compare different approaches, in a dynamic
a random position called test point is selected in the search environment it is not sufficient to simply compare the best
space at the beginning of algorithm execution and its fitness solution found, because the optimum is changing over time.
value is stored. In each iteration of proposed algorithm The main metric for evaluating the performance of
execution, the fitness value of test point is evaluated and algorithms is offline error, indicating the average of the best
compared to its previous result. If the two results in two found position’s fitness during the running optimization
subsequent executions are the same, it means that the process. The value of offline error is equal or greater than
environment has not changed. On the other hand, their zero, where zero indicates the ideal situation.
difference illustrates a change in the environment. This
algorithm, regularly re-evaluate the current best solution to For all diploid genetic algorithms, parameter settings are
detect changes in the environment. as follows: population size n=100, standard tournament
selection with tournament size of 2 and elitism of size 1for
7) Dominance Change Mechanism replacement.
In natural systems, dominance can change over time, as a
result of the presence or absence of particular enzymes. We Uniform crossover for discrete diploid genetic algorithms
have extended proposed diploid genetic algorithm by adding with increase accuracy to a certain degree with Pc=0.6 and
a similar dominance change mechanism. After detecting arithmetic crossover for proposed real-valued diploid genetic
change, there will be the problem of diversity loss. To solve algorithm with the same crossover probability is considered
this problem a dominance change mechanism is executed to and bit flip mutation with Pm=0.01 for discrete ones and
increase diversity. In this mechanism, the genotypic alleles uniform mutation for proposed real-valued algorithm with
of all individuals will change according to their genotypic the same probability is applied.
alleles before the environment change. After the
For all diploid genetic algorithms with dominance As it is observed, the proposed algorithm has the better
change scheme, change mechanism is triggered whenever efficiency than the other algorithms. In addition, the
the environment changes are detected. For each experiment proposed algorithm with dominance change mechanism has
of diploid genetic algorithms on MPB, 30 independent runs the better performance than the proposed algorithm without
were executed and each time, they were performed by change mechanism. Extending the proposed algorithm with a
different random seeds. Each run continued until change mechanism improves it considerably, since it can
environment changes 10 times. In table (III) to (VII), the respond quickly to changes in the environment.
proposed algorithm efficiency is compared with other known
diploid genetic algorithms. two well known diploid Although the performance of all diploid genetic
algorithms have been chosen for the study: Ng-Wong with algorithms decreases by increase of peak numbers, the
dominance change mechanism and Additive with dominance efficiency of proposed algorithm is still better than the
change mechanism. others. In other words, the proposed algorithm efficiency is
higher than the others both when there are a few number of
peaks and when there are many of them.

TABLE III. COMPARISON OF OFFLINE ERROR OF 4 DGAS ON MPB PROBLEM WITH DIFFERENT NUMBER OF PEAKS IN 500 FREQUENCY

Number of Ng-Wong (with Additive (with Continuous Diploid Genetic Continuous Diploid Genetic
peaks dominance change) dominance change) Algorithm(without dominance Algorithm (with dominance
change) change)
1 59.0231(0.21) 49.3014(0.06) 39.2914(0.01) 37.8907(0.12)
5 63.7970(0.04) 55.7586(0.05) 38.9997(0.08) 38.9258(0.15)
10 61.0486(0.12) 57.3101(0.05) 45.4607(0.11) 39.1627(0.07)
20 62.6846(0.11) 58.7142(0.05) 33.9244(0.07) 38.0419(0.08)
30 59.9374(0.04) 59.7608(0.06) 42.0238(0.76) 40.9180(0.06)
50 63.4169(0.17) 61.2444(0.07) 41.7898(0.08) 41.7549(0.17)
100 68.2784(0.10) 62.7261(0.06) 43.9885(0.03) 42.5675(0.03)

TABLE IV. COMPARISON OF OFFLINE ERROR OF 4 DGAS ON MPB PROBLEM WITH DIFFERENT NUMBER OF PEAKS IN 1000 FREQUENCY

Number of Ng-Wong (with Additive (with Continuous Diploid Genetic Continuous Diploid Genetic
peaks dominance change) dominance change) Algorithm(without dominance Algorithm (with dominance
change) change)
1 53.6435(0.06) 37.5436(0.02) 29.4836(0.10) 28.1412(0.03)
5 57.0909(0.21) 48.7586(0.06) 32.2897(0.06) 33.5334(0.02)
10 59.0860(0.10) 50.8833(0.17) 33.0824(0.03) 35.6446(0.07)
20 59.3099(0.17) 52.2097(0.13) 35.8220(0.19) 33.8648(0.06)
30 62.6387(0.08) 52.0025(0.09) 36.0639(0.04) 35.6192(0.01)
50 65.6844(0.06) 61.2444(0.62) 34.0704(0.05) 36.4688(0.02)
100 71.2784(0.01) 62.7261(0.05) 37.0734(0.05) 37.0628(0.05)

TABLE V. COMPARISON OF OFFLINE ERROR OF 4 DGAS ON MPB PROBLEM WITH DIFFERENT NUMBER OF PEAKS IN 2500 FREQUENCY

Number of Ng-Wong (with Additive (with Continuous Diploid Genetic Continuous Diploid Genetic
peaks dominance change) dominance change) Algorithm(without dominance Algorithm (with dominance
change) change)
1 40.4848(0.76) 54.1001(0.38) 19.8517(0.21) 16.5960(0.06)
5 42.9442(0.56) 46.6150(0.32) 22.4296(0.19) 22.6373(0.03)
10 44.1571(0.26) 39.8061(0.16) 28.5634(0.12) 28.6494(0.27)
20 44.7853(0.31) 38.3590(0.08) 25.1881(0.11) 25.0247(0.22)
30 45.9699(0.11) 39.7983(0.09) 27.4953(0.05) 26.6119(0.14)
50 49.3867(0.18) 47.2633(0.10) 29.1899(0.08) 29.0448(0.10)
100 53.0158(0.09) 52.8635(0.21) 26.9202(0.06) 28.0621(0.16)
TABLE VI. COMPARISON OF OFFLINE ERROR OF 4 DGAS ON MPB PROBLEM WITH DIFFERENT NUMBER OF PEAKS IN 5000 FREQUENCY

Number of Ng-Wong (with Additive (with Continuous Diploid Genetic Continuous Diploid Genetic
peaks dominance change) dominance change) Algorithm(without dominance Algorithm (with dominance
change) change)
1 41.0384(0.17) 38.8207(0.06) 14.8440(0.12) 11.6838(0.16)
5 40.8552(0.19) 35.2645(0.08) 23.2127(0.05) 21.6591(0.31)
10 43.2642(0.14) 40.8552(0.09) 23.9921(0.04) 22.9517(0.17)
20 46.3564(0.16) 38.3887(0.02) 26.9870(0.10) 27.0733(0.16)
30 51.5379(0.04) 42.3138(0.08) 27.9030(0.02) 27.8398(0.17)
50 50.7719(0.21) 43.5692(0.09) 27.0894(0.01) 22.5450(0.06)
100 50.7970(0.11) 47.2099(0.26) 31.0293(0.31) 25.5878(0.11)

TABLE VII. COMPARISON OF OFFLINE ERROR OF 4 DGAS ON MPB PROBLEM WITH DIFFERENT NUMBER OF PEAKS IN 10000 FREQUENCY

Number of peaks Ng-Wong (with Additive (with Continuous Diploid Genetic Continuous Diploid
dominance change) dominance change) Algorithm(without dominance Genetic Algorithm (with
change) dominance change)
1 32.2672(0.20) 29.2927(0.14) 9.6922(0.15) 8.5307(0.08)
5 38.9451(0.32) 27.8454(0.10) 19.0228(0.38) 19.2818(0.05)
10 34.2337(0.18) 35.2690(0.03) 21.1115(0.12) 20.5169(0.02)
20 35.2773(0.17) 31.7640(0.10) 23.6263(0.09) 22.4366(0.09)
30 39.8805(0.06) 34.4421(0.07) 24.0374(0.04) 25.2383(0.09)
50 44.9072(0.15) 39.9801(0.02) 24.9321(0.10) 24.7480(0.05)
100 46.0750(0.13) 43.6175(0.39) 25.1383(0.18) 23.9844(0.47)

V. CONCLUSION [4] Y. Jin, J. Branke, “Evolutionary Optimization in Uncertain


Environments-A Survey”, IEEE Transaction on Evolutionary
Redundant coding using diploid genomes are the most Computation, Vol. 9, no. 3, pp. 303-317, 2005.
common implicit memory used in evolutionary algorithms [5] J. Branke, Evolutionary Optimization in Dynamic Environments,
for solving dynamic problems. In this paper, for the first springer, 2001.
time, a real-valued diploid genetic algorithm was proposed [6] J. Branke, Schmeck H. “Designing Evolutionary Algorithms for
for optimization in dynamic environments. The new Dynamic Optimization Problems”, Advances in Evolutionary
Computing, Springer, pp. 239-262, 2003.
dominance mechanism consists of a simple function which
[7] M. Kominami, Hamagami T., “A New genetic Algorithm with
has homogeneous outputs. Results on MPB problem, being Diploid Chromosomes Using Probability Decoding for Adaptation to
the most famous dynamic and continuous benchmark, were Various Environment”, Wiley Publishing, Vol. 93, pp. 38-46, 2010.
compared with other known discrete diploid genetic [8] S. Yang, “On the Design of Diploid Genetic Algorithms for problem
algorithm with increased accuracy to a certain degree. Also a Optimization in Dynamic Environments”, IEEE Congress on
new dominance change mechanism is proposed and added to Evolutionary Computation, pp. 1362-1369, 2006.
new diploid genetic algorithm to improve its performance for [9] S. Yang, “Learning the Dominance in Diploid Genetic Algorithms for
responding to changes of the environment. It is clear from Changing Optimization Problems”, ACM, 2nd International
Symposium on Intelligence Computation and Applications, pp. 157-
the result of experiments that there is a significant difference 162, 2007.
between the performance of proposed algorithm and other
[10] S. Yang, “Dominance Learning in Diploid Genetic Algorithms for
discrete diploid genetic algorithms. Dynamic Optimization Problems”, ACM, pp.1435-1436, 2006.
There is some relevant works to pursue in the future. [11] S. Uyar, Harmanci E., “A new Population based Adaptive
Domination Change Mechanism for Diploid Genetic Algorithms in
Combining the explicit memory schemes with the proposed Dynamic Environments”, Springer Soft Compute, pp. 803-814, 2005.
algorithm will further improve the performance of new
[12] K. V. Makhotilo "Diploid Genetic Algorithm with Mortality," vol. 43,
diploid genetic algorithm in dynamic environment. pp. 60-73, 2011.
[13] J. Lewis, Hart E., “A Comparison of Dominance Mechanisms and
REFERENCES
Simple Mutation on Non-Stationary Problems”, Springer, pp. 139-
[1] X. Yu, Introduction to Evolutionary Algorithms, Springer, 2010. 148, 1998.
[2] I. G. Amo, D. A. Petla, J. R. Gonzalez, A. D. Masegosa, “An [14] M. Kominami and T. Hamagami, "A new genetic algorithm with
Algorithm Comparison for Dynamic Optimization Problems”, diploid chromosomes by using probability decoding for non-
Applied Soft Computing, Vol. 12, no. 10, pp. 3176-3192, 2012. stationary function optimization," in Systems, Man and Cybernetics,
[3] S. Yang, “Population-Based Incremental Learning With Associative 2007. ISIC. IEEE International Conference ,pp. 1268-1273, 2007.
Memory for Dynamic Environments”, IEEE Transaction on [15] T. T. Nguyen, S. Yang, J. Branke, “Evolutionary Dynamic
Evolutionary Computation, Vol. 12, no. 5, pp.542-561, 2008. Optimization: A Survey of the state of the art”, Swarm and
Evolutionary Computation, Vol. 6, pp. 1-24, 2012.
[16] J. Branke, Available at: http://people.aifb.kit.edu/jbr/MovPeaks/

You might also like