Multi-Population Methods with Adaptive Mutation for Multi-Modal Optimization Problems

International Journal on Soft Computing, Artificial Intelligence and Applications (IJSCAI), Vol.2, No.2, April 2013
DOI : 10.5121/ijscai.2013.2201 1
Multi-Population Methods with Adaptive
Mutation for Multi-Modal Optimization Problems
Imtiaz Korejo1
, Shengxiang Yang2
(IEEE Member), Kamran Brohi3
, Z.U.A.
Khuhro4
1
Institute of Mathematics & Computer Science, University of Sindh, Jamshoro,
Pakistan
cskorejo@gmail.com
2
School of Computer Science and Informatics, De Montfort University, Leicester,
United Kingdom
syang@dmu.ac.uk
ABSTRACT
This paper presents an efficient scheme to locate multiple peaks on multi-modal optimization problems by
using genetic algorithms (GAs). The premature convergence problem shows due to the loss of diversity,
the multi-population technique can be applied to maintain the diversity in the population and the
convergence capacity of GAs. The proposed scheme is the combination of multi-population with adaptive
mutation operator, which determines two different mutation probabilities for different sites of the
solutions. The probabilities are updated by the fitness and distribution of solutions in the search space
during the evolution process. The experimental results demonstrate the performance of the proposed
algorithm based on a set of benchmark problems in comparison with relevant algorithms.
KEYWORDS
Multi-population approaches, adaptive mutation operator, multi-modal function optimization, genetic
algorithms.
1. INTRODUCTION
Genetic algorithms (GAs) are adaptive heuristic search techniques inspired by genetic
inheritance and natural evolution of species. Over the years, GAs have been successfully used
for solving many optimization problems due to the properties of easy-to-use and robustness for
determining candidate solution for ambitious problems. There are a lot of applications of GAs in
different areas of real world problems, such as science, engineering, business, and social
science. The interested reader may wish to consult [11]. The basic concept of GAs was
introduced by John Holland in the 1960s in the USA. There are different components of GAs,
such as the encoding scheme, the population size, the selection approaches, the crossover and
mutation operators; the performance of GAs depends on good choice of operators and relevant
parameters. However, it is difficult to choose proper operators and relevant parameters for the
optimal performance. A limited amount of research has been done on GAs to find multiple
optima for multi-modal optimization problems. However, the implicit parallism of GAs is able
to find several optima (or multiple optima including global and local optima) in the search space
instead of just a single solution. Recently, researchers have shown a great interest in optimizing
multimodal functions with GAs. Several GA techniques have been proposed to optimize multi-
modal problems.
In recent years, various techniques have been suggested into standard evolutionary
algorithms (EAs) to divide the population proposed to optimize multi-modal problems. In recent
years, various techniques have been suggested into standard evolutionary algorithms (EAs) to
divide the population into different sub-populations, including k-means clustering algorithm
[16], self-organizing scouts (SOS) [5], multinational GAs [38], clustering scheme [27], niching

2
method [28], and a hierarchical clustering [46]. These techniques have considered sub-
populations as a means of increasing the diversity of GAs. The main motivation behind these
schemes is to maintain multiple populations on different peaks, the highest peak in the search
space is the global optimum and other peaks with a lower height than the global optimum are
local optima.
There are different schemes for controlling the values of relevant parameters and
genetic operators of an EA. Parameter control or adaptive adaptation techniques can be
classified into three categories: deterministic adaptation, adaptive adaptation, and self-adaptive
adaptation approaches. Deterministic adaptation adjusts the values of parameter according to
some deterministic rule without using any feedback information from the search space. The
strategy variables are modified by using the feedback information from the search space. This is
called adaptive adaptation. The best example of this type of adaptive adaptations is
Rechenberg’s “success rule” in evolutionary strategies (ESs) [31] (p. 110). In the case of self-
adaptive adaptation, parameters are altered by EAs themselves. Strategy parameters to be
altered are encoded into chromosomes and undergo variation with the rest of the chromosome.
The self-adaptive adaptation was proposed by Back and Schlutz [3] into GAs.
In this paper, a new multi-population with adaptive mutation operator is proposed for
GAs. Two different multi-population approaches are considered in this study, one is species
with adaptive mutation (SWAM) and another is basin with adaptive mutation (BWAM). The
proposed technique can maintain the exploration and exploitation in the evolution process,
which can be considered into the category of adaptive adaptation EAs. This paper gives the
comparison of several adaptive mutation operators, including adaptive GA, self-adaptive GA,
and site-specific rate GA with our proposed technique. The experimental results demonstrate,
the proposed algorithm exhibits better performance in terms of the quality of candidate solutions
found on the benchmark problems.
The rest of this paper is outlined as follows. Section 2 explains related work, the
classification of adaptation in EAs, and some sub-population approaches. Section 3 presents the
SWAM and BWAM adaptive mutation operators in detail. Section 4 describes the experimental
study of comparing our proposed techniques with other algorithms. Finally, conclusion and
discussions on further directions of research are stated in Section 5.
2. RELATED RESEARCH
Generally speaking, a lot of research has been done on selecting appropriate genetic
operators and relevant parameters for GAs to find optimal or near-optimal solutions. Usually,
these genetic operators and relevant parameters are fixed before the execution of GAs. For
example, some constant mutation probabilities are derived from experience or by trial-and-error.
It is very difficult, if not impossible, to find suitable parameter setting for the optimal
performance of GAs. Different values of parameters and operators might be optimal at different
levels of the evolutionary process. Due to this reason, researchers are turning to adaptive
operators and parameters for GAs to find the optimal results.
2.1. Adaptation in GAs
The main motivation of adapting a GA is to adjust genetic operators and relevant
parameters in order to find better solutions of a problem. Adaptation in GAs is one of the most
important and promising research areas. The efficiency of GAs is dependent on not only the
algorithms, representation, and operators for the problem, but the set of the parameter values for
GAs to find a good solution. There are different methods for controlling the values of various
parameters and genetic operators of GAs. Generally, there are two classification schemes [1],
[8], [9]. The classification of adaptation based on these two distinct schemes are explained as
follows:
1) Type of adaptation: the type of adaptation means how a parameter is modified.

3
2) Level of adaptation: the level of adaptation means where changes occur.
The above classification schemes were first introduced by Angeline in [1]. According to
Angeline, the type of adaptation is further divided into absolute update rules and empirical
update rules. The level of adaptation is also further classified into three levels, i.e., population
level, individual level, and component level. In [8], [9], the authors extended the classification
schemes and broadened the idea introduced by Angeline in [1]. They further classified the type
of adaptation in two subfields: parameter tuning and parameter control. This is bases on the
technique of adaptation applied in the process. It is explained as follows
1) Parameter Tuning or Static Adaptation: Parameter tuning (or static adaptation) means
to set the suitable parameters before the run of EAs and they remain constant during the
execution of EAs. In other words, no change in the parameters occurs. Static adaptation is done
by an external process (e.g., a person or a program) for selecting appropriate values. Many
researchers have suggested different static appropriate values for key parameters of GAs. These
values are derived from experiences or by trial-and-error methods. In [18], De Jong proposed
static single-point crossover and mutation probabilities for finding a good solution for problems
using a traditional GA. Some other researchers also suggested well known heuristics with other
static parameters for finding a good solution for problems [2], [11], [30], [25], [23], [24]. The
different values of parameters and operators might be optimal at different levels of the
evolutionary process. Due to this reason, static adaptation (parameter tuning) is gradually
discouraged in the EA community. So, there is no common optimal parameter setting that can
be found a prior (consult to No Free Lunch Theorem [42]). According to this problem,
researchers diverted their attention towards adaptive adaptation (parameter control) methods.
2) Parameter Control or Adaptive Adaptation: Strategy parameters are adjusted on the
basis of various methods during the execution of EAs. This class of adaptation can be further
divided into three classes, described as deterministic adaptation, adaptive adaptation, and self-
adaptive adaptation. Deterministic adaptation adjusts the values of parameter according to some
deterministic rule without using any feedback information from the search space. Adaptive
adaptation modifies strategy variables by using feedback information from the search space.
The best example of adaptive adaptations is Rechenberg’s “1/5 success rule” in ESs [31] (p.
110). In the case of self-adaptive adaptation, the parameters are altered by EAs themselves.
Strategy parameters to be altered are encoded into chromosomes and undergo variation with the
rest of the chromosome. The self-adaptive adaptation was first proposed by Schwefel [32] into
ESs.
There are three distinct levels at which adaptation can take place in adaptive EAs [34].
For the population-level adaptation, strategy parameters are modified globally for the whole
population. Various examples are available for population-level adaptation in ESs and GAs
[43], [44]. For the individual level adaptation, an EA modifies the strategy parameters of an
individual and these modifications occur independently with each individual. In [41], the
crossover probability in a GA is changed at the individual level. For the component level
adaptation, parameters are changed for some genes of an individual in the population. Self-
adaptation is a well-known example of component-level adaptation.
Fogarty [10] proposed a dynamic mutation rate control scheme for GAs, in which the
mutation rate decreases exponentially over the number of generations. In papers [6], [7], Devis
proposed an effective algorithm that updates operator probabilities according to the performance
of the operators. This scheme specifies for the modification of operator rates in associated to the
fitness of solutions generated by the operators. It has been introduced for the steady state GA.
Srinivas and Patnaik [37] suggested an adaptive GA that uses an adaptive crossover and
mutation probability scheme. The rates of crossover and mutation are modified according to the
fitness of the solution. Uyar el al. proposed an asymmetric gene based adaptive mutation
(GBAM) technique [39]. In GBAM, each gene locus has two different mutation probabilities:
is used for those loci that have the value of “1” and is used for those loci that have
the value of “0”. The probabilities of and are automatically updated based on the

4
feedback information from the search space, according to the relative success or failure of those
chromosomes having a “1” or “0” at that locus for each generation. Yang and Uyar [45]
suggested another gene based adaptive mutation with fitness and allele distribution correlation
(GBAM FAD). The mutation rates of each gene locus are adaptively modified based on the
correlated feedback information from the search process, according to the relative success or
failure of solutions.
A dynamic mutation GA (DMGA) was proposed by [15]. This approach simultaneously
uses four mutation operators in solving optimization problems. These mutation operators are
adaptively applied in the GA. In [19], the authors introduced an adaptive mutation technique,
which is closely related to Hong’s method, simultaneously applying multiple mutation operators
in GAs [15]. The authors proposed a directed mutation (DM) operator for real coded GAs [17],
which is to explore the promising area in the search space by using the feedback information
from the current population.
This is an altered version of the standard mutation. In the DM scheme, solution moving
is not only based on the feedback information of average fitness of intervals, but also on the
population distribution. By taking into account the information of population distribution, DM
avoids the premature convergence problem.an adaptive mutation technique, which is closely
related to Hong’s method, simultaneously applying multiple mutation operators in GAs [15].
The authors proposed a directed mutation (DM) operator for real coded GAs [17], which is to
explore the promising area in the search space by using the feedback information from the
current population. This is an altered version of the standard mutation. In the DM scheme,
solution moving is not only based on the feedback information of average fitness of intervals,
but also on the population distribution. By taking into account the information of population
distribution, DM avoids the premature convergence problem.
2.2. Multi-population Approaches
Several researchers have considered multi-population schemes for locating multiple
peaks in a multi-modal fitness landscape. These approaches can be applied to cover solutions on
different peaks with different multi-populations. In this paper, we focus on more than one peak
in the fitness landscape. A global optimum is the best peak in the fitness landscape and local
optima are referred to the rest of the lower height peaks. Some other techniques have been
proposed to maintain the diversity of the population. These approaches work just like multi-
population schemes, which are crowding [21], fitness sharing [12], island and parallel GAs [13],
[4]. Multi-population techniques can improve the diversity in the GA and avoid the premature
convergence problem.
Branke et al. [5] introduced a self-organizing scout (SOS) algorithm, which has shown
promising results on evaluated problems. In this scheme, the population consists of a parent
population that explores through the whole search space and child populations that locate
previously detected optima. The parent population is continuously examined. If the condition is
satisfied, then a child population is split off from the parent population. The size of each child
population is updated regularly, although the total number of solutions are fixed and no new
solutions are proposed.
Ursem [38] proposed a multi-population scheme referred to as the multinational GA,
which employs a multinational GA technique to find multiple peaks of multi-modal problems in
dynamic environments. Each nation corresponds to the policy and best representatives. In
multinational GAs, a “hill-valley detection” approach is applied to sample points on a line
marked between policies. This procedure uses migration of solutions from one nation to
another, to merge nations, and to create a new nation in a newly located region.
A clustering approach for multi-modal functions was proposed in [27]. This approach
determines the number of clusters within the k-means particle swarm optimization (PSO)
algorithm. It has been applied for the optimization of a criterion function in a probabilistic
mixture-model framework. In this algorithm, particles are supposed to be generated by a mix of

5
several probabilistic distributions. Each different cluster is associated to a different distribution.
Then, locating the optimum number k is equivalent to adapting the model with the determined
data while optimizing some criterion.
The niching method [28] and the species conserving GA (SCGA) [20] have been used
for multi-modal optimization problems. Both the clearing (niching) procedure and SCGA
adapted a scheme for splitting population according to the nation of species, which is
incorporated under the evolution process of a standard GA. The two methods have shown to
provide best results on evaluated multi-modal problems.
Yang and Li [46] developed a hierarchical clustering based PSO algorithm which is
specially introduced for finding and tracking multiple optima in dynamic environments. This
approach can distribute particles in separate promising subregions and the main contribution of
this approach is adaptively modifying the number of sub-swarms required, and automatically
compute the search region for each sub-swarm. A fast local search method is incorporated in
this approach , which specifies the optimal solutions in promising subregion detected by the
clustering scheme.
3. MULTI-POPULATION SCHEMES WITH ADAPTIVE
MUTATION OPERATOR
3.1 Motivation
In order to explain the problem of premature convergence is due to the loss of diversity,
the multi-population approach can be applied to maintain the diversity and retain the
convergence capacity of GAs. The suggested technique is composed of multi-population with
adaptive mutation operator, which identifies two different mutation rate vectors for each
subpopulation. These vectors are updated by using the fitness and distribution of individuals in
the fitness landscape. Each site of solution having two different mutation rates. This approach
can be able to improve the performance of the algorithm and to avoid the premature
convergence to multi-modal optimization problems.
3.2 Multi-population Schemes
1) Species Based Multi-population approach: A species can be a collection of
individuals sharing with common characteristic regarding to the similarity of the individuals.
The measurement of two individuals similarity is calculated by the hamming distance. The
smaller hamming distance between two individuals, the more similar they are. Species-based
particle swarm optimization(SPSO) scheme was proposed by Parrot and Li [26].
2) An Other Multi-population approach: In this paper [40], the authors has recently
proposed a partitioned technique, which divides the current population into different
subpopulation according to the fitness and distribution of individuals in the landscape. Each
subpopulation can be able to ideally incorporate the basins of attraction of similar optima. The
basic idea of proposed partition scheme is to distribute best individuals away from each
other(according to the distance measure) under the different peak in the search space.
3.3 Adaptive Mutation Operator with Multi-population approaches
Traditionally, mutation operator considered as a background operator, it affects the
alteration of the value of each bit of individual with small mutation probability ∈ [0,1].
The main motivation of the mutation in the GAs has been that of insurance policy against
permanent loss or unexplore genetic material in the population to avoid the premature
convergence of the GAs to local optima.
The adaptive techniques are used to accelerate the convergence speed and preserving
GAs from being trapped into local optima. The important contribution of adaptive mutation
operator to improve the performance of GAs. All above explained adaptive schemes are used to

6
a single population in the process of evolution, either using steady state GA or generational
GAs.
In this paper, multi-population with adaptive mutation scheme is proposed. Basically
adaptive mutation operator is combining with two different multi-population approaches, one is
species based multi-population and the other one is basin (fitness of solution and distribution of
individual in the search space) based approach. BWAM and SWAM represents basin with
adaptive mutation and species with adaptive mutation respectively. GAs has two properties
exploration and exploitation but the problem is that how to balance these properties during the
evolution. In order to address this problem, the researchers have been used separately multi-
population approaches and adaptive probabilities of genetic operators to increase the diversity
during the evolution. Our proposed scheme is going to combine both of them for finding the
better solution in the multi-modal problems. This technique is applied to maintain the balance of
explorative and exploitative properties of genetic operator.
The framework of the multi-population GAs with the adaptive mutation scheme is
represented in Algorithm 1. This algorithm starts from randomly generated population which is
called p population. Then, sub-populations are created either using BWGA or SWGA scheme.
When sub-populations have been generated then, calculate mutation probability, apply mutation
and selection approaches on them. At the end, overlapping and convergence techniques are
applied on each sub-population before starts the next generation. Some of the task in this
algorithm are similar to the classical SGA but here we are going to focus on the main operations
of this approach, including partition, Statistics, Mutation, overlapping and convergence
methods. These methods determine the general theme of technique. The detail description of the
these functions are given in the following:
Algorithm 1: Multi-population scheme with adaptive mutation operator
1: Randomly create an initial population ;
2: Evalute the fitness of each solution of ;
3: 0;
4 partition( );
5: while ( < max gen) do
6: for ( to ) do
7: Statistics( );
8: Mutation( );
9: Selection( );
10: end for
11: Overlapping( );
12: Convergence( );
13: if (p < MinInds) then
14: regeneratep:=Re-initialization(removedSols);
15: := partition( );
16: Merge and into
17: end if
18: + 1;
19: end while
1) partition(pt):There are several techniques available in the literature to generate multi-
population, two approaches are considered in this paper, detail explanation of these techniques
are mentioned in the section 3.2. represents the current population, which is passed to the
partition function as a parameter. The functionality of this method is to divide the current
population into different subpopulation. The different sub-population have different position in

7
the fitness landscape. The multi-population scheme can be applied to maintain the diversity
according to the convergence. This approach can be helpful towards getting the better results on
multi-modal problems. Each sub-population has assigned and update probabilities separately. In
the following section will compute the probability of each site of individual of each sub-
population.
2)Statistics: In the statistics method, the mutation rates are computed by using the
feedback information of current sub-population. Each sub-population has defined two different
probability vectors, where and are the mutation rates and of subpopulation at
generation . Each vector contains set of element, which is associating to the probabilities of
keeping a particular allele at each site. Initially each element of vector is set to an initial value
within its determined boundaries. These vectors are generated real-valued probabilities of the
each sub-population, which determines high quality sub-population with high probability. For
each generation vectors are updated based on the fitness and distribution of solutions in the
specified sub-population, it means that each sub-population has altered own vectors
independently. The two different mutation rates are specified for each site of solution
of each sub-population. The estimation of both vectors are updated on the explication shown in
equation 2 and 3.
( , + 1) =
( , ) + ,
1
>
( , ) − ,
(1)
( , + 1) =
( , ) − ,
1
>
( , ) + ,
(2)
where is the updated value of mutation probabilities of the vectors, is the average fitness
with allele “1“ of those individuals which are contained in subpopulation for site at
generation , and is the average fitness of the subpopulation at iteration .
3) Mutation : This function is performed mutation on each sub-population at every
iteration. It is distinctly employed on each subpopulation using the associated vectors. Suppose,
an offspring ( ́ = { ́ , … … ́ })generated by mutation of parent ( = { , … … })
according to corresponding vectors, which is the member of subpopulation. The mutation is
performed on whole solution site by site, which is mentioned in the Algorithm 2, where Rnd is
the uniform random number, which is generated within the interval(0,1).
Algorithm 2 Mutated each Individual by using adaptive mutation
1: (j 0 to 1 )
2: ( [ ] == 0)
3: Rand(0,1) < M [j]
4: ́[ ] = 1;
5:
6: else
7: Rand(0,1) < M [j]
8: ́[ ] = 0;
9: end if
10: end if
11: end for

8
The proposed mutation scheme is applied with two different partition approaches (See
Section 3.2). The contribution of the adaptive mutation with multiple population is summarized
in the following distinct characteristics:
1) It describes the two different vectors for each subpopulation.
2) Each site of subpopulation having two different mutation rates and different mutation
probabilities for the different locus of the encoded individual.
3) The rate of each locus are updated iteration-by-iteration to modify the current state of the
evolution.
4) This is the one of the important property of this approach, which investigates the undetected
regions around the search space.
5) The suggested method is focused on the locating multiple optima of the multi-modal
optimization problems.
3.4 Overlapping and Convergence of sub-populations
After mutation and selection operations, all the subpopulations are checked regarding
overlapping and convergence. Normally, the overlapping check between two subpopulations
can determined by comparing the distance of the best solutions of the sub-populations. If the
search radius of best individuals of sup-populations is less than threshold value , than
delete the sub-population, which is related to less fitness best solution and keep the total
individuals, which are associated to the removed sub-population. Algorithms 3 is shown the
overlapping search of sub-populations.
Algorithm 3: The algorithm for overlapping search
1: while ( . () > 0) ( < . ()) do
2: 1 ( );
3: 0;
4: + 1;
5: while ( . () > 0) ( < . ()) do
6: if ( ( 1, 2) < ) then
7: if ( 1 < 2) then
8: + ( ). ();
9: ( );
10: else
11: + ( ). ();
12: ( );
13: end if
14: end if
15: end while
16: end while
After overlapping search, all sub-populations will undergo the convergence process, which is
determined to see weather sub-population has converged. If the total similar solutions of
associated sub-population is greater than threshold value _ , the value of threshold
(0.95*size of subpopulation) is fixed in this paper. If sub-population is converged on one of the
available peaks then sub-population is removed from , which is converged. Convergence
of subpopulation is given in the algorithm 4.
Algorithm 4 Algorithm for determine convergence of subpopulation
1: for _ ∈ do
2: if ( > (0.95 ∗ . ())) then

9
3: + . ();
4: ;
5: end if
6: end for
4. EXPERIMENTAL STUDY
In this section, there are three different group of experiments were conducted on well-known
decision problems, which has been evaluated in the popular area of maximum satisfiability
(MAX-SAT) problems. Boolean satisfiability problem involves finding an assignment of
variables that maximizes the number of satisfied clauses of a given constraints. The general
form of these constraints are presented in conjunctive normal form (CNF) or product-of-sum
form. MAX-SAT is also well known NP-hard optimization problems. In addition, Boolean
satisfiability expression has applied to introduced a approach for generating problems with
controllable degree of ”multimodality” and ”epistasis”.
In the rest of this section, first introduce the mapping between GAs to MAX-SAT
problems, spear [36] for generating constraints with controllable degree of multi-modality and
epistasis and then investigate the result of experiments.
4.1 Boolean Satisfiability and Genetic Algorithms
In order to employ GAs to any particular problem, that problem consider two critical
components: 1) specify the appropriate representation for the solution space (search space) and
2) define an external evolution function which determines utility of individuals (fitness of
individuals). MAX-SAT has a simple string representation which is highly compatible to the
GAs. Each individual shows binary strings of length in which bit determines the truth value
of the boolean variable of the boolean variables shown in the boolean expression.
Any candidate solution is evaluated by the fitness function, which delegates fitness
value 1 to that individual (string) boolean values of that string by which the boolean expression
is evaluated to 1 and 0 to other solutions.
Spear proposed another scheme [36] that has been assigned fitness to individual
subexpressions in the original expression and aggregate them in some way to produce a
complete individual fitness value. According to this context the general and natural way is to
specify the value of TRUE to be 1 and the value of FALSE to be 0 by Smith [35].
4.3 Comparison of proposed scheme with other approaches
The performance of proposed algorithm is compared with some existing GAs
techniques, which are chosen from the GAs literature. According to the performance, these
algorithms were tested on different multi-modal boolean satisfiability problems, which are
widely applied in the literature. Initially our scheme has been compared to standard GAs with
random fixed mutation and crossover probabilities. In addition, the proposed technique
compared with some other parameter control approaches such as adaptive [37], self-adaptive [3]
and SSRGA [40] schemes. These all comparable approaches are considered, which briefly
described as follows:
The simple genetic algorithm (SGA) is collection of individuals, these solutions of SGA
are represented as a bit string of length , mutation and crossover operators are used in the SGA
as a variation operator. The probabilities of these variation operators are initially assigned by
random values (0,1) in our experiments. The initial values are fixed during the whole evolution
process.
Srinivas and Patnaik introduced a adaptive genetic algorithm (AGA) [37] which is
efficient adaptive algorithm for multimodal optimization problems. The main motivation of this
approach is to maintain the diversity in the population and retain the convergence capacity of

10
the GA by using adaptive probabilities of crossover and mutation. The rate of mutation and
crossover are updated depending on the fitness values of the solutions. In order to address the
convergence property, to determines and for each solution to protect highly-fit solutions
from high disruption. The mutation and crossover rates are updated applying following
equations.
=
( − ) ( −⁄ ), ≥
,
(3)
=
( − ) ( −⁄ ), ≥
,
(4)
where and represents the maximum fitness and average fitness of the population,
′
is
the larger of the fitness values of the individual to be crossed and is the fitness of solution
which is being mutated. 1, 2, 3, 4 ∈ [0,1] are predefined constants. These predefined
constant values are taken from [37], same values are considered for our experiments.
Self-adaptive genetic algorithm (SAGA) was developed by Back and Schutz [3]. which
uses single mutation rate separately per solution. For example an individual composes of a bit
binary string of length and associated mutation rate μ ∈ [0, 1]. The new mutation probability is
adjusted according to the following equation.
́ = 1 +
1 −
exp (0,1) (5)
where = 0.22 is the learning rate and N(0, 1) is the uniform distributed value with mean 0.0
and standard deviation 1.0. The new solution is specified through bit wise mutation of l bits
applying the mutated mutation rate value ́. The mutation rate is not less than 1/ .
Vafaee and Peter recently proposed a site-specific rate genetic algorithm (SSRGA) [40]
scheme. Which introduced mutation scheme to determines different mutation rate for different
sites of the individuals. The main motivation of this approach to face both explorative and
exploitative responsibilities of variation operators. This scheme starts from partitions of
population into number of sub-populations is defined based on fitness and distribution of
solutions contained in the search space. Then, in order to identify the representative individual
associating to each sub-population. Finally, the result of this procedure is a probability vector
(SSR vector) with elements associating to the rate of possessing a specific allele at each locus of
every solution included in a sub-population.
4.4 Experimental setting
In order to investigate our proposed scheme with other four approaches on problems
with different level of difficulty, the difficulty of the problem means to modify the length of
individuals and population size . Generally, the problem becomes more difficult if the length
of individuals would be increase. To avoid the premature convergence due to increase
population size and may also be the algorithms have a ability to find a better solutions during
the evolution process. We have been considered different levels of problem difficulty by using
the parameter pairs ( , ) with (20,30), (30,50), (50,50) and (60,70). These parameter setting is
used to multimodal landscape with different degree of multi-modality by assigning number of
peaks 1,5,10. Each peak is shown by a randomly created solution of length , which identifies
the specific position of peak in the search space. The default setting of parameters applied in the
experiments of this paper can be observed in the original papers [37], [3], [40]. To compare the
performance of multi-population with adaptive approach with other described above algorithms
in section 4.3. All algorithms were allowed same population size, length of individual, stop
criteria and maximum number of generations for each run.

11
4.4.1. Analysis of Parameter Sensitivity
The performance of included approaches depends on the operator and parameters applied. In
this paper, the key parameters are considered for the sensitivity analysis such as population size,
individual length
and radius( ). To locate the multiple optima by using the small population size, it is not
always good idea. Increasing the population size, however, to reduce the probability of being
trapped into local optimum and increasing the probability of locating better solution during the
evolution process. Solution is consist of length of bit binary string, if length of string
increases the problem become more challenging since the state space is exponentially expended
according to the length of solution increases.
Table 1
Parameter setting for all compared algorithms
Parameter Setting
l 20 30 50 60
n 30 50 50 70
r 0.5 0.6 0.7
Peaks 01 05 10
r( ) is used to determine the species and species seed in the SPSO [26] and to identify
the representative individual from population [40]. The radius (r) is the key parameter to both
approaches. If the radius is too small, there is potential problem that few isolated solutions
species may tarp into local optimum due to very small size of radius. In this context, the loss of
diversity then the algorithms can’t contribute further more during the process. If radius is too
large, there may be more than one optimum within the radius. It is very difficult to choose the
suitable value of radius. In this paper, we have considered different values of radius to
investigate the performance of algorithms. The different parameters and their settings of
different level of multi-modality problems applied in our experiments can be seen in Table 1. In
order to determine that which parameter setting gives the great performance among the
compared algorithms.
4.4.2 Experimental Results and Analysis
In order to investigate that multi-population with adaptive mutation schemes can
improve the performance of GAs on solution quality and convergence rate for multi-modality
problems. All algorithms were run independently 50 times on the different degree of multi-
modality problems. From the Table 2, It can be determined that the same value of r is
considered for first three algorithms because which is related to these approaches; the r is
applied with distinctly in different algorithms, first one is used a r as threshold (minimum
distance between two solutions), value in BWAM, another one is determined as a radius of the
each species SWAM and last one is considered of r values with same meaning as in first one.
The experimental results are shown in Table 2, which gives detail result of average and
standard deviation of the best-found fitness values during the evolution. Both values are
computed when the maximum number of iteration is completed or the global optimum is found.
The algorithm determine better results among the other algorithms which presents bold average
best fitness. In addition, those average best fitness values are underlined which is associated
with smallest standard deviation among the other approaches. The effect of smaller standard
deviation determine to produced reliable solutions.
1) Effect of varying the length of solution and population size: The algorithms have
been investigated on different problems with different level of difficulty. The difficulty means
to change the length of individual and population size. The problem would become more
difficult if increasing the length of solution. The size of population is increasing, however, less
probability of premature convergence to local optimum and increase probability to locate the

12
better solutions at the termination of optimization process [36]. The proposed algorithm have
been tested on different length of solution and population size.
Table 2
Comparison results of different algorithms with different parameter setting with different problems
Peak 1
(l,n) (20,30) (30,50) (50,50) (60,70)
r*l (0.5*l) (0.6*l) (0.7*l) (0.5*l) (0.6*l) (0.7*l) (0.5*l) (0.6*l) (0.7*l) (0.5*l)(0.6*l)(0.7*l)
BWAMavg
std
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
SWAMavg
std
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
SSRGAavg
std
1.0
0.0
1.0
0.0
1.0
0.0
0.998
0.007
1.0
0.0
0.998
0.006
0.922
0.016
0.916
0.019
0.921
0.022
0.891
0.016
0.893
0.014
0.897
0.016
SAGA avg
std
0.877
0.033
0.824
0.024
0.750
0.022
0.740
0.020
AGA avg
std
0.93
0.026
0.872
0.023
0.794
0.020
0.774
0.015
SGA avg
std
0.897
0.051
0.876
0.047
0.817
0.041
0.810
0.034
Peak 5
(l,n) (20,30) (30,50) (50,50) (60,70)
r*l (0.5*l) (0.6*l) (0.7*l) (0.5*l) (0.6*l) (0.7*l) (0.5*l) (0.6*l) (0.7*l) (0.5*l)(0.6*l)(0.7*l)
BWAMavg
std
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
SWAMavg
std
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
SSRGAavg
std
1.0
0.0
1.0
0.0
1.0
0.0
0.991
0.014
0.995
0.011
0.944
0.012
0.906
0.017
0.906
0.017
0.911
0.016
0.879
0.026
0.874
0.026
0.878
0.024
SAGA avg
std
0.909
0.022
0.850
0.021
0.779
0.018
0.760
0.012
AGA avg
std
0.943
0.026
0.878
0.022
0.802
0.017
0.779
0.014
SGA avg
std
0.905
0.038
0.886
0.039
0.806
0.034
0.807
0.028
Peak 10
(l,n) (20,30) (30,50) (50,50) (60,70)
r*l (0.5*l) (0.6*l) (0.7*l) (0.5*l) (0.6*l) (0.7*l) (0.5*l) (0.6*l) (0.7*l) (0.5*l)(0.6*l)(0.7*l)
BWAMavg
std
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
SWAMavg
std
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
1.0
0.0
SSRGAavg
std
1.0
0.0
1.0
0.0
1.0
0.0
0.987
0.017
0.988
0.015
0.987
0.016
0.876
0.040
0.882
0.027
0.884
0.030
0.831
0.031
0.837
0.030
0.833
0.034
SAGA avg
std
0.923
0.030
0.866
0.021
0.788
0.018
0.774
0.021
AGA avg
std
0.952
0.030
0.890
0.021
0.810
0.020
0.791
0.019
SGA avg
std
0.920
0.036
0.893
0.032
0.820
0.034
0.805
0.027

13
2) Effect of varying the peaks: The performance of BWAM and SWAM is equal on
specific setting of( , ) on the different number of peaks, but efficiency of SSRGA sightly
declines as number of peaks increases. However, the rest of the compared algorithms are
reversed.
3) Effect of varying the ( / ): Niche radius(r) is a very important role
regarding the performance of both approaches (BWAM and SWAM), which is determine
subpopulations. The many species could be generated if the value of r is very small. it is a big
problem due to this reason few isolated solutions of sup-population trap into local optimum very
quickly. Setting r to a too large will cause more than one optimum could be contained within the
circle (r). The experiments determine the effect of the different values on the performance of
those algorithms which are associated with the r in this paper.
The statistical results of comparing algorithms is investigated applying two-tailed t-test
with a 98 degree of freedom at 0.05 level of significance. Table 3 presents t-test results of
comparing pair of algorithms, where the result is determined as “s+” ,“s-”, “+”, “-”, and “ _ ” if
the first algorithm in pairs is significantly better than, significantly worse than the
insignificantly better than, insignificantly worse than or equivalent to the second algorithm,
respectively. The t-test result of comparing algorithms shows on peak 10 and is
chosen 0.6 * with different level of problem settings. From Table 3 it can be observed that the
BWAM and SWAM both approaches are statistically better than other comparing schemes to
locating the optimum value in the fitness landscape. Here we presents only statistical results
of compared techniques with different level of difficulty of different problems on peak 10.
According to the parameter setting of different level of multi-modality problems are
divided into three categories such as simple problems, bit harder problems, and difficult
problems. The performance of all algorithms varies on different categories. From Fig.1 it can be
determined that the convergence speed of involved algorithms, which were tested on different
level of difficulty in simple multi-modal problems. In Fig.1 BWAM, SWAM, and SSRGA have
a higher convergence rate than self-adaptive GA, adaptive GA, and Standard GA on simple
problems. The overall the convergence speed of BWAM and SWAM is initially better than
SSRGA on simple problems but with time going the performance of all three algorithms are
same except last two rows in Fig.1.
Table 3
t-test values of comparing algorithms on different problems with different level of difficulty
Problems (20,30) (30,50) (50,50) (60,70)
BWAM-SWAM ~ ~ ~ ~
BWAM-SSRGA ~ ~ s+ s+
BWAM-SAGA s+ s+ s+ s+
BWAM-AGA s+ s+ s+ s+
BWAM-SGA s+ s+ s+ s+
SWAM-SSRGA ~ ~ s+ s+
SWAM-SAGA s+ s+ s+ s+
SWAM-AGA s+ s+ s+ s+
SWAM-SGA s+ s+ s+ s+
SSRGA-SAGA s+ s+ s+ s+
SSRGA-AGA s+ s+ s+ s+
SSRGA-SGA s+ s+ s+ s+
SAGA-AGA s+ s- s+ s-
SAGA-SGA s+ s+ s- s-
AGA-SGA s+ s+ - -

14
Fig 2. presents the results of the evolutionary process of compared algorithms on bit harder
problems. It can be clearly noticed that the convergence speed of BWAM and SWAM are
quicker than other approaches but it is interesting to observe that different scheme have different

15
Fig 1. Evolutionary process of the comparison algorithms on simple problems
Fig. 2. Evolutionary process of comparison algorithms on bit harder problems.
convergence rate on different (bit harder) problems with different peaks. On few problems, the
BWAM is initially better than SWAM regarding convergence speed.
From Fig3. it can be seen that the convergence rate of BWAM initially is faster than
SWAM from beginning to 50 generations on first column, afterwards the both schemes have
same convergence curve. In this figure, the convergence curves are quiet clear among the
compared algorithms.
Analyzing the convergence rate of six comparing schemes. The very interesting things
is that the convergence speed of both (BWAM, SWAM) approaches is initially quick
convergence rate than compared techniques even on simple problems where SSRGA is going to
compete to the BWAM and SWAM. It can also be seen two properties of proposed approach
from above figures. The first one is the quick convergence rate against with other comparable
algorithms except simple problems. BWAM and SWAM have quick convergence speed than
other compared approaches on different levels of problem difficulty by setting the parameters.
Another goal of suggested scheme is the ability to explore the prominent area by avoiding the
premature convergence in the fitness landscape on multimodality problems. The proposed
technique reduce the less probability of trapping into local optima and increase the probability
of locating the better results than the other algorithms.

16
5. CONCLUSION AND FUTURE WORK
From the beginning, community of evolutionary algorithm has been done lot of work regarding
to increase the performance and locating multiple optima in multi-modal optimization problems
of GAs. The exploration and exploitation are becoming two main characteristic in the case of
performance. The researchers have applied multi-population scheme to increase the efficiency
of GAs in this context.
This paper proposed a new mutation approach, which is the combination of multi-
population and adaptive mutation genetic operator. It is determine two different probability
Fig. 3. Evolutionary process of comparison algorithms on difficult problems.
vectors for each sub-population. Each vector contains elements associating to the probabilities
of possessing a specific allele at each site. These vectors are updated during the evolution
process. The suggested mutation scheme has been considered two different multi-population
techniques in this paper.
In order to analysis the performance of the our proposed algorithms, experimental
results were conducted on different benchmark problems and to compare the results to standard
GA, and some other approaches which belongs to parameter control approaches. From the
experiments, it can be seen that suggested scheme is greatly improve the performance of GA
regarding to the locating and tracking the multiple optima in the fitness landscape by using the
multi-population with adaptive mutation operator.
In the future, we will use proposed mutation scheme to large and complex real-world
problems because GAs have a lot of applications in different area of real-world problems such
as science, engineering, business and social science, the interested reader may wish to refer [11]
p. 125. In addition, it would be interesting to combine clustering partition technique with
adaptive mutation operator to improve the performance of GAs. Finally, we will conduct further
experiments on epistatic [36] and deceptive[11] benchmark problems, which are similar to the
multi-modal problems.

17
REFERENCES
[1] P.J. Angeline, (1995) “Adaptive and Self-adaptive evolutionary computations,” In M. Palaniswami
and Y. Attikiouzel, editors, Computational Intelligence: A dynamic system perspective, pp. 152-
163. IEEE Press.
[2] T. B¨ack, (1992) “Self-Adaptation in Genetic Algorithms,” In Proc. of the 1st European Conf. on
Artificial Life, pp. 263-271.
[3] T. Back and M. Sch¨utz, (1996) “Intelligent mutation rate control in canonical genetic algorithms,”
In Foundation of Intellignet Systems, Springer, pp. 158-167.
[4] M. Bessaou, A. Paetrowski, and P. Siarry (2000) “Island model cooperating with speciation for
multimodal optimization,” In Proc. of the 6th Int.Conf. on Parallel Problem Solving from Nature,
pp. 16-20.
[5] J. Branke, T. Kaubler, C. Schmidt, and H. Schmeck, (2000) “A multi-population approach to
dynamic optimization problems,” In Proc. of the 4th Int. Conf. on Adaptive Comput. Des Manuf,
pp. 299-308.
[6] L. Davis, (1989) “Adapting operator probabilities in genetic algorithms,” Proc. of the 3rd Int.
Conf. on Genetic Algorithms, pp. 61-69.
[7] L. Davis, (1991) “Hankbook of Genetic Algorithms,” Van Nostrand Reinhold.
[8] A.E Eiben, R. Hinterding, and Z. Michalewicz, (1999) “Parameter Control in Evolutionary
Algorithms,” IEEE Trans. on Evol. Comput., 3(2): pp. 124- 141.
[9] A.E. Eiben, Z. Michalewicz, M. Schoenauer, and J.E. Smith. (2007) Parameter Control in
Evolutionary Algorithms, Parameter setting in Evolutionary Algorithms, Chapter 2, Springer
Verlag, pp. 19-46.
[10] T.C. Fogarty, (1989) “Varing the probability of mutation in genetic algorithms,” In Proc. of the
3rd Int. Conf. on Genetic Algorithms, pp. 104-109.
[11] D. E. Goldberg, (1989) Genetic Algorithms in Search, Optimization, and Machine Learning, New
York: Addison-Wesley.
[12] D. E. Goldberg, and J. Richardson, (1987) “Genetic algorithms with sharing for multimodal
function optimization,” In Proc. of the 2nd Int. Conf.on Genetic Algorithms, pp. 41-49.
[13] V.S. Gordon, D. Whitley, and A. Bohn, (1992) “Dataflow parallelism in genetic algorithms,” In
parallel Problem Solving from Nature, pp. 533-542.
[14] J. J. Greffenstette, (1986) “Optimization of Control Parameters for Genetic Algorithms,” IEEE
Trans. on Sys. Man and Cyber., 16(1): pp. 122-128.
[15] T. P. Hong, H. S. Wang, and W. C. Chen. (2000) “Simultaneously applying multiple mutation
operators in genetic algorithms,” Journal of Heuristics, 6: pp. 439-455.
[16] J. Kennedy. (2000) “Stereotyping: Improving particle swarm performance with cluster analysis,”
In Proc. 2000 IEEE Congr. Evol. Comput, pp. 1507- 1512.
[17] I. Korejo, S. Yang and C. Li, (2010) “A Directed Mutation Operator for Real Coded Genetic
Algorithms,” EvoApplications 2010, Part I, LNCS 6024,pp. 491-500.
[18] K. A. De Jong, (1975) “An Analysis of the Behavior of a Class of Genetic Adaptive Systems,”
PhD Thesis, Department of Computer and Communication Science, University of Michigan, Ann
Abor.
[19] C. Li, S. Yang, and I. Korejo, (2008) “An Adaptive Mutation Operator for particle swarm
optimization,” In Proc. of the 2008 UK Workshop on Computational Intelligence, pp. 165-170.
[20] J. P. Li, M. E. Balazs, G. Parks, and P. J. Clarkson, (2002) “A species Conserving genetic
algorithm for multimodal function optimization,” Evol. Comput., vol. 10, no. 3, pp. 207-234.

18
[21] S. W. Mahfound, (1992) “Crowding and preselection revisited,” In Proc. of the 2nd Int. Conf. on
Parallel Problem Solving from Nature, vol 2, pp. 27-36.
[22] H. M¨uhlenbein, (1992) “How Genetic Algorithms Really Work I. Mutation and Hillclimbing,”
In Proc. of the 2nd Int. Conf. on Parallel Problem Solving from Nature, pp. 15-29.
[23] G. Ochoa, (2002) “Setting the mutation rate: Scope and limitations of the 1/l heuristics,” Proc. of
the 2002 Genetic and Evol. Comput. Conf.,pp. 315-322.
[24] G. Ochoa, (2006) “Error thresholds in genetic algorithms,” Evol. Comput.,vol. 14, no. 2, pp. 157-
182.
[25] G. Ochoa, (1999) “Error thresholds and their relation to optimal mutation rates,” Springer-Verlag,
Berlin, Germany.
[26] D. Parrott and X. Li, (2006) “Locating and tracking multiple dynamic optima by a particle swarm
model using speciation,” IEEE Trans. Evol. Comput., vol. 10, no. 4, pp. 440-458, Aug.
[27] A. Passaro and A. Starita, (2008) “Particle swarm opotimization for multimodal functions: A
clustering approach,” J. Artif. Evol. Appl., vol. 2008, pp. 1-15.
[28] A. Petrowski, (1996) “A clearing procedure as a niching method for gentic algorithms,” In Proc.
1996 IEEE Int. Conf. Evol. Comput., pp. 798-803.
[29] R. Rechenberg, (1996) Evolutionary Algorithms in Theory and Practice. New York: Oxford
University, Press.
[30] J. D. Schaffer, R. A. Caruana, L. J. Eshelman, and R. Das, (1989) “A Study of Control Paramters
Affecting Online Performance of Genetic Algorithms for Function Optimization,” Proc. of the 3rd
Int. Conf. on Genetic Algorithms, pp. 51-60.
[31] H. P. Schwefel, (1995) Evolution and Optimum Seeking, Chichester, UK: John Wiley.
[32] H. P. Schwefel, (1981) “Numerical optimization computer models,” Chichester,Wiley.
[33] J.E. Smith and T.C. Fogarty, (1996) “Self-adaptation of mutation rates in a steady state genetic
algorithm,” In Proc. of the 3rd IEEE Conf. on Evol. Comput., pp. 318-323.
[34] J.E. Smith and T.C. Fogarty, (1997) “Operator and parameter Adaptation in genetic algorithms,”
Soft Comput., vol. 1, no. 2, pp. 81-87.
[35] G. Smith, (1979) “Adaptive genetic algorithms and boolean satisfiability problem,” Technical
Report, University of Pittsburgh, Pittsburgh.
[36] W. M. Spears, (2000) Evolutionary algorithms, the role of mutation and recombination, Natural
Computing, Speringer-Verlag, Berlin, Germany.
[37] M. Srinivas and L.M. Patnaik, (1994) “Adaptive probabilities of crossover and mutation in genetic
algorithms,” IEEE Trans. on System, Man and Cybern., vol. 24, no. 4, pp. 656-666.
[38] R. K. Ursem, (2000) “Multinational GAs: Multimodal Optimizaiton Techniques in Dynamic
Enviroments,” In Proc. 2nd Genetic Evol. Comput. Conf., pp. 19-26.
[39] S. Uyar, S. Sariel, and G. Eryigit, (2004) “A gene based adaptive mutation strategy for genetic
algorithms,” Proc. of the 2004 Genetic and Evol. Comput. Conf., pp. 271-281.
[40] F. Vafee and P. C. Nelson (2010) “An Explorative and Exploitative Mutation Scheme,” Proc.
2010 IEEE World Congr. on Comput. Intell., pp. 18-23.
[41] T. White and F. Oppacher, (1994) “Adaptive crossover using automata”, Proc. of the 3rd Conf. on
PPSN, pp. 229-238.
[42] D. H. Wolpert and W. G. Macready, (1997) “No free lunch theorems for optimization,” IEEE
Trans. on Evol. Comput., vol. 1, no. 1, pp. 67- 82.
[43] S. Yang, (2002) “Adaptive non-uniform crossover based on statistics for genetic algorithms,” Proc.
of the 2002 Genetic and Evol. Comput. Conf., pp. 650-657.

19
[44] S. Yang, (2003) “Adaptive mutation using statistics mechansim for genetic algorithms,” In F.
Coenen, A. Preece, and A. Macintosh (editors), Research and Devlelopment in Intelligent System
XX, pp. 19-32.
[45] S. Yang and S. Uyar, (2006) “Adaptive mutation with fitness and allele distribution correlation for
genetic algorithms,” Proc. of the 21st ACM Symp. on Applied Computing, pp. 940-944.
[46] S. Yang and C. Li, (2010) “A clustering particle swarm opotimizer for locating and tracking
multiple optima in dynamic environments,” IEEE Trans. on Evol. Comput. vol. 14, no. 6, pp. 959-
974, December.
Authors
Imtiaz Ali Korejo received his B.Sc.(Hons) and M.Sc.(Hons) in Computer Science
from University of Sindh, Jamshoro, Pakistan, in 1999, and 2000, respectively. He
worked as a research associate in the Institute of Mathematics and Computer
Science, University of Sindh, Jamshoro, Pakistan from 2001 to April 2003, as a
Lecturer in the same institute from April 2003 to 2012; He is currently working in
the same institute as an Assistant Professor since April 2012. He received his Ph.D.
from the Department of Computer Science, University of Leicester, United
Kingdom in 2012.
His research interests are evolutionary algorithms, genetic algorithms and adaptive approaches.
Shengxiang Yang received the B.Sc. and M.Sc. degrees in automatic control and
the Ph.D. degree in systems engineering from Northeastern University, China in
1993, 1996, and 1999, respectively. From October 1999 to June 2012, he worked
as a Postdoctoral Research Associate in the Department of Computer Science,
King's College London, U.K., a Lecturer in the Department of Computer Science,
University of Leicester, U.K., and a Senior Lecturer in the Department of
Information Systems and Computing, Brunel University, U.K., respectively. Since
July 2012, he has been appointed as a Professor at the Center for Computational
Intelligence, School of Computer Science and Informatics, De Montfort University, U.K.
His major research interests include evolutionary and genetic algorithms, swarm intelligence,
computational intelligence in dynamic and uncertain environments, artificial neural networks for
scheduling and relevant real-world applications. He has over 150 publications. He has given invited
keynote speeches in several international conferences and co-organized several symposiums, workshops
and special sessions in conferences. He serves as the area editor, associate editor or editorial board
member for four international journals. He has co-edited several books and conference proceedings and
co-guest-edited several journal special issues. He is the chair of the Task Force on Evolutionary
Computation in Dynamic and Uncertain Environments, Evolutionary Computation Technical Committee,
IEEE Computational Intelligence Society, and the founding chair of the Task Force on Intelligent
Network Systems, Intelligent Systems Applications Technical Committee, IEEE Computational
Intelligence Society.

Multi-Population Methods with Adaptive Mutation for Multi-Modal Optimization Problems

More Related Content

What's hot (19)

Viewers also liked (14)

Similar to Multi-Population Methods with Adaptive Mutation for Multi-Modal Optimization Problems (20)

Recently uploaded (20)

Multi-Population Methods with Adaptive Mutation for Multi-Modal Optimization Problems