Upload
dervis
View
218
Download
0
Embed Size (px)
Citation preview
Accepted Manuscript
A novel binary artificial bee colony algorithm based on genetic operators
Celal Ozturk, Emrah Hancer, Dervis Karaboga
PII: S0020-0255(14)01053-6
DOI: http://dx.doi.org/10.1016/j.ins.2014.10.060
Reference: INS 11228
To appear in: Information Sciences
Received Date: 25 November 2013
Revised Date: 20 October 2014
Accepted Date: 27 October 2014
Please cite this article as: C. Ozturk, E. Hancer, D. Karaboga, A novel binary artificial bee colony algorithm based
on genetic operators, Information Sciences (2014), doi: http://dx.doi.org/10.1016/j.ins.2014.10.060
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers
we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting proof before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
A Novel Binary Artificial Bee Colony Algorithm BasedOn Genetic Operators
Celal OZTURK*, Emrah HANCER, and Dervis KARABOGAErciyes University, Engineering Faculty, Computer Engineering Department, Kayseri, Turkey
{celal,emrahhancer,karaboga}@erciyes.edu.tr
Abstract—This study proposes a novel binary version of the artificial bee colony algorithm
based on genetic operators (GB-ABC) such as crossover and swap to solve binary optimization
problems. Integrated to the neighborhood searching mechanism of the basic ABC algorithm,
the modification comprises four stages: 1) In neighbourhood of a (current) food source,
randomly select two food sources from population and generate a solution including zeros
(Zero) outside the population; 2) apply two-point crossover operator between the current, two
neighborhood, global best and Zero food sources to create children food sources; 3) apply swap
operator to the children food sources to generate grandchildren food sources; and 4) select the
best food source as a neighbourhood food source of the current solution among the children and
grandchildren food sources. In this way, the global-local search ability of the basic ABC
algorithm is improved in binary domain. The effectiveness of the proposed algorithm GB-ABC
is tested on two well-known binary optimization problems: dynamic image clustering and 0-1
knapsack problems. The obtained results clearly indicate that GB-ABC is the most suitable
algorithm in binary optimization when compared with the other well-known existing binary
optimization algorithms. In addition, the achievement of the proposed algorithm is supported by
applying it to the CEC2005 benchmark numerical problems.
Keywords; Binary optimization, Dynamic clustering, Knapsack problem, Artificial bee
colony, Genetic algorithm.
1 Introduction
Concentrating on local interactions of the swarm individuals (including bird
flocks, fish schools, ants, bees and etc.) with each other and with their environments
[5], Swarm Intelligence (SI) has become a significant research area among computer
scientists, engineers, economists, bioinformaticians and several other disciplines on
account of the fact that the remarkable ability of natural intelligent swarms on solving
their related problems (finding food, building nests etc.) can be simulated to deal with
real world problems [32]. It is known that an intelligent swarm should have the
following properties [47]: 1) the ability of proximity, 2) the ability of receiving and
responding to quality factors, 3) the ability of protecting behaviours against
fluctuations and 4) the ability of adapting to diverse situations. In the 1990s, two
important SI based algorithms which are ant colony optimization (ACO) [18] and
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
particle swarm optimization (PSO) [20] were developed and both of them have
attracted researchers’ attention.
After realizing the above four mentioned properties of SI in honey bee swarms,
researchers, especially from the beginning of the 2000s, started to concentrate on
modelling various intelligent behaviours of these swarms; for instance, the behaviours
of dance and communication, collective decision, task allocation, nest site selection,
mating, foraging, marriage, floral and pheromone laying [29]. Some well-known bee
swarm intelligence based algorithms are the virtual bees algorithm (VBA) [70], the
bees algorithm (BA) [57], BeeAdHoc [64], honey bee mating optimization (HMBO)
[22], the BeeHive [65], bee system (BS) [43], bee colony optimization (BCO) [61]
and the artificial bee colony (ABC) algorithm [27, 28]. According to [32], ABC is the
most widely used algorithm among bee swarm based algorithms presented in the
literature. Some successful applications of the ABC algorithm are optimization of
numerical problems [31], data clustering [35], neural networks training [34], wireless
sensor network deployment [54] and routing [33], and image analysis [24, 25, 44, 52,
53]. It is clearly seen that there exist various appreciated studies of the ABC algorithm
on continuous space [1, 6, 7, 13, 14, 30, 60, 66, 68, 72]. However, that situation is not
the same for the studies of ABC on binary problems. Through this gained
information, the basic aim of this study is to propose an improved binary model of the
ABC algorithm to solve binary optimization problems.
1.1 Related Works
It is known that the standard ABC algorithm can optimize continuous problems. In
other words, the structure of the basic ABC cannot be directly adapted to the binary
optimization problems. Thus, some modifications are required on the ABC algorithm
for binary optimization applications. Pampapa and Engelbrecht [56] considered
binary solution as four dimensional real vector through angle modulation schema.
Kashan et al. [37] introduced a binary version of the ABC algorithm (DisABC) which
uses the dissimilarity measure of binary structures instead of the arithmetic
subtraction operator. To measure the similarity of binary structures, Jaccard’s
coefficient of similarity is employed and to generate a new solution, two selections
are applied in a probabilistic manner; random selection depending only on
neighbourhood and greedy selection depending on both neighbourhood and the best
solution. The performance of the algorithm was verified with the implementation of
the uncapacited facility location problem (UFLP). The same methodology was also
applied to the differential evolution algorithm to solve UFLP problem [36]. Kıran and
Gunduz [39] proposed a binary version of the ABC algorithm based on XOR logic
operator (XOR-ABC) and its performance was compared with the DisABC algorithm
in terms of UFLP.
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
1.2 Motivations
For decades, researchers have attempted to develop algorithms in optimization
satisfying superior performance with respect to the other algorithms in literature [49].
To design a suitable algorithm, it is crucial to establish a connection between the
algorithm and a given optimization problem i.e. analysing optimization problem is the
starting point. There exist two mostly preferred approaches by the researchers to
improve an existing algorithm, known as hybridization and modification. The former
is the process of mixing of at least two heterogeneous particles through conscious
manipulation or as a natural progressive manipulation [71]. The latter is the process of
modifying some particles of the mechanism through internal or external forces. As
observed in nature, memes, which are the basic unit of cultural transmission and
imitation [17], can encounter modifications and combine with each other to generate
stronger memes. Language life cycle can be given as an example to this concept.
While new words are welcomed by languages on account of the needs and interests of
society or community, some words lose their popularity and hence become
disappeared from languages. From the algorithmic perspective, it is known that two or
more properly combined distinct algorithms and modified algorithms by the operators
of other algorithms can improve the ability of problem-solving mechanism. As the
importance of cultural information transmission has been realized by researchers,
memetic algorithms, “which are population-based metaheuristics composed of an
evolutionary framework and a set of local search algorithms activated within the
generation cycle of the external framework” [49], has gained popularity. That
motivates us to develop a genetically modified artificial bee colony algorithm for
optimizing binary search space which resembles to the memetic algorithms using
genetic operators.
1.3 Contribution
In this paper, a novel binary artificial bee colony algorithm (GB-ABC) based on
genetic operators is introduced. To our knowledge, this study is the first to use genetic
operators in a binary model of the ABC algorithm. Furthermore, the GB-ABC
algorithm is not intended to solve only a specific binary problem, it is also proposed
to overcome general binary optimization problems. This is demonstrated by applying
GB-ABC to well-known binary optimization problems; the dynamic image clustering
problem known as the process of automatically finding the optimum number of
clusters and the 0-1 knapsack problem known as maximizing the cost of a knapsack
with maximum weight capacity. For the dynamic clustering problem, the performance
of the GB-ABC algorithm is analysed by comparing it with the results of the binary
particle swarm optimization for dynamic clustering (known as DCPSO) [51], the
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
quantum inspired binary particle swarm optimization (QBPSO) [10], the genetic
algorithm for dynamic clustering, the discrete binary artificial bee colony (DisABC)
[37], the self-organizing map (SOM), the unsupervised fuzzy clustering approach
(UFA) and Snob algorithms. As for the knapsack problem, the performance of the
GB-ABC is evaluated by comparing with the results of the BPSO, QBPSO, GA, and
DisABC algorithms. It should be noted that the applied GA in binary problems is one
of the basic type of GA in which the population is represented in bit string form, the
crossover type is scattered, the mutation type is Gaussian, the reproduced population
consists of 2 elitist members, and selection type is stochastic uniform. Besides binary
problems, the suitability of the GB-ABC algorithm to the numerical problems
(CEC2005 benchmarks) is also considered and its results are compared with the same
existing algorithms as in the knapsack problem.
1.4 Organization
The rest of the paper is organized as follows: Section 2 summarizes the ABC
algorithm; Section 3 describes the GB-ABC algorithm; Section 4 explains the
considered problems concerning the perspective of an optimization problem; Section
5 discusses the results; and finally Section 6 concludes the paper.
2 ABC Algorithm
The artificial bee colony (ABC) algorithm is the simulation of the minimalistic
foraging model of honey bee in search process for solving real-parameter, non-
convex, and non-smooth optimization problems [30]. In this algorithm, a colony
consists of three types of artificial bees: employed bees, onlooker bees and scout bees;
Employed bee: Each food source is associated with an employed bee and each
employed bee tries to detect a new food source in the neighbourhood of its current
food source. The detected food source is memorized when the nectar amount of the
detected food source is higher than the nectar amount of current food source. After
completion of the search process, employed bees share their information concerning
the nectar amount of food sources with onlooker bees via waggle dance in the dance
area.
Onlooker bee: An onlooker bee evaluates the information gained from the
employed bees and tries to find a new food source in the neighbourhood of the
selected food based on this evaluated information. Thus, the tendency of onlooker
bees is to search around the food sources with high nectar amount; in this way, more
qualified food sources can be chosen for exploitation.
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
Scout Bee: The number of scout bees is not predefined in the colony. A scout bee
is produced according to the situation of a food source whether it is abandoned or not.
When an abandoned (fully exploited) food source is detected, its employed becomes a
scout bee, and then randomly searches a new food source in space. After that, the
scout bee returns to its earlier type i.e. employed bee again. The number of scout bees
is controlled by the parameter “limit” representing the number of trials before
defining a food source as “abandoned”. If a food source cannot be improved during
the predetermined number of unsuccessful trials being equal to limit parameter, that
food source is leaved and a new food source is randomly generated. Exploration and
exploitation procedures are carried out together for robust search process in the ABC
module [31].
In the ABC model, a food source denotes a possible solution of an optimization
problem and the nectar amount of a food source represents the quality of a solution.
As mentioned before, each food source is associated with an employed bee in the
colony. Thus, the number of food sources is equal to the number of employed bees.
Also, the number of employed bees is equal to the number of onlooker bees.
Therefore, the number of food sources is half of the population number.
The main steps of the ABC algorithm are as follows:
1. Initialize food sources ( ) by Eq. (1);
= + (0,1) − (1)
where = 1, … , (SN is the number of food sources); = 1, … , (D is the number
of parameters); is the minimum and is the maximum values of
parameter j.
2. Move employed bees to the food sources and determine neighbourhood food
sources by Eq. (2);
İ = +ϕ (x − x ) (2)
where and k are randomly chosen parameters and neighbourhoods, respectively
and ϕ is a random number within [-1,1].
3. Determine the new target food source as and remove from the memory, if the
nectar amount (fitness value) of is better than .
4. Move onlooker bees probabilistically to the food sources depending on the roulette
wheel using Eqs. (3) and (4) and determine new neighbourhoods as in Steps 2 and
3;
=∑
(3)
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
=1
1 + , ≥ 0
1 + ( ), < 0 (4)
where is the cost value (value of objective function) of the food source and
fitnessi the nectar amount (fitness value) of the food source .
5. Randomly initialize food source by a scout bee using Eq. (1), if a better food
source cannot be found in the neighborhood of present one after a number of
trials, called as “limit”.
6. Repeat steps (2) to (5) until Cycle=MCN
A basic pseudo code of the ABC algorithm can be seen in Fig. 1.
Fig. 1 The pseudo code of the ABC algorithm
3 The Proposed Algorithm
As mentioned in Introduction section, the structure of basic ABC is not suitable
for binary optimization. Thus, Eq. (1) and (2) need to be modified. In the proposed
genetic binary artificial bee colony algorithm (GB-ABC), the initial solutions are
generated by Eq. (5) instead of Eq. (1);
Xi: i=1...SN = 0 (0,1) ≤ 0.51 (0,1) > 0.5 (5)
where (0,1) is a uniformly generated value.
The genetic operators such as two-point crossover and swap operators are
integrated into ABC as a modification (referred as “genetic neighbourhood generator
(GNG)”) in place of Equation (2). For two-point crossover operator [46], two
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
positions on binary vectors are randomly chosen and then the parts of the vectors
between these positions are exchanged. For the swap operator, two positions are
randomly chosen with the values of 1 and 0 on the binary vector. Then, the position
with the value 1 is varied to 0, and the other with the value 0 is changed as 1.
The genetic neighbourhood generator (GNG) is as follows:
1. For a binary current food source = ( , , … . , ), randomly select two
neighborhoods = ( , , … . , ) and = ( , , … . , ) where is
the dimension of food source.
2. Determine a dimensional zero food source = 0 .
3. Define , , , and as parent food sources, where is the best
food source of the population.
4. Randomly match the parent food sources to each other based on one-to-one and
onto relation as seen in Fig. 2, and then apply a two-point crossover on the
matched parent food sources as seen in Fig. 3.
Fig. 2 An example of one-to-one and onto matching between parent food sources for crossover application.
Fig. 3 An illustrative example of exchanging positions between food sources through two-point crossover.
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
5. If any food source with zeros exists around children food sources, change ℎ
number of positions to 1 on the zero food source where ℎ is a predetermined
threshold value.
6. Apply swap operator to the children food sources as seen in Fig. 4 and define the
generated food sources as “grandchildren”.
Fig. 4 An illustrative example of applying swap operator on children food sources.
7. Select the best food source among the children and grandchildren food sources as
.
8. is determined as current food source instead of if the fitness value of is
better than the fitness value of , as in basic ABC.
In GB-ABC, the food source is determined to improve the global searching
mechanism since the food source satisfies diversity in children generation
through crossover. For crossover, one-to-one and onto functional relation is
determined between the parent food sources, an example of which can be found in
Fig. 2. In this way, the chance of getting the optimal solution is improved. As for the
determination of Th parameter, it is chosen as 3 for all the applications. The main aim
of using Th is to avoid possible errors resulting from empty solutions during the
objective function evaluation of children and grandchildren food sources. Any
number between 1 and D can be defined as Th parameter on account of the fact that
the proposed algorithm has an efficient search mechanism and its performance is not
adversely affected from this situation.
The pseudo code of the GB-ABC algorithm is presented in Fig. 5 and an
illustrative example of GNG implementation is given as follows:
Let = (1001010101) is the current solution; = (0010101011) and
= (1011101101) are randomly selected food sources; = (1101001100)
and = (0000000000). Assign these food sources as parents and make a
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
randomly one-to-one and onto relation them as seen in Fig. 2 and apply two-point
crossover to generate children between the matched food sources as in Fig. 3. After
the generation of children food sources, apply swap operator to create grandchildren
shown in Fig. 4. Finally, call the objective function to evaluate the quality of children
and grandchildren and assign the best food source as the food source.
Fig. 5 The pseudo code of the GB-ABC algorithm
4 Implementation of GB-ABC Algorithm
Three different problems such as dynamic image clustering, knapsack and
numerical optimization are chosen to show the effectiveness of the proposed
algorithm. The definition and implementation of problems through the GB-ABC
algorithm are given as follows;
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
4.1 Image Clustering
Image clustering is the process of identifying similar image primitives such as
pixels, regions, border elements etc. and collecting them together [59]. The act of
grouping is based on some similar measures like distance and intervals within
multidimensional data [26]. The groups satisfy “a concise summarization and
visualization of the image content that can be used for different tasks related to image
database management” [21]. In recent decades, image clustering has obtained a
significant role in biomedical imaging, remote sensing, geology, computer vision and
etc. all of which aim to achieve accurately partitioned and meaningful clustered
images.
Image clustering algorithms can be categorized into two groups: supervised and
unsupervised. Supervised classification is based on prior knowledge of the group or
members i.e. training examples [48]. A supervised learning clustering algorithm
analyzes the training data and produces an inferred output according to the test data
set (prior knowledge). Gaussian maximum likelihood, minimum-distance-to-mean,
artificial neural networks (ANN) and the naïve Bayes classifier are some of the most
widely used supervised algorithms. The main drawback of supervised algorithms is
the requirement of prior knowledge which may not be available in large unlabeled
datasets.
Contrary to the supervised clustering algorithms, unsupervised clustering
algorithms do not need prior information that makes these algorithms more acceptable
in the literature. Unsupervised clustering algorithms are classified into two groups:
hierarchical and partitional. Partitional clustering algorithms are more widely used in
image analysis and pattern recognition which lead us to use these algorithms in this
study. Partitional clustering algorithms can be accepted as optimization problems
since they attempt to partition the data into a set of disjoint clusters using some
similarity measures (eg. square error function) minimized by assigning clusters to
peaks in the probability density function, or the global structure. The most well-
known partitional clustering algorithm K-means [45] is applied by assigning data
members to the closest centroids via Euclidean distance measure. Various
modifications on K-means such as Fuzzy C-Means (FCM)[19], possibilistic C-means
(PCM) [41], robust fuzzy C-means (RFCM) [58] etc. are introduced to improve the
clustering performance. However, the number of clusters needs to be predefined in
these algorithms.
Clustering can be carried out in two different forms: crisp (hard) and soft (fuzzy).
In crisp clustering, all patterns should be assigned to only one cluster i.e. the clusters
are disjoint and have at least one pattern. As distinct from crisp clustering, a pattern
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
may be assigned to multiple clusters with fuzzy membership property in soft
clustering. For instance, K-means is a hard clustering algorithm and Fuzzy C-means
is a soft clustering algorithm.
Today’s technology allows us to store a large scale data set. Thus, clustering
algorithms requiring a predefined number have recently become disabled and the
demand for dynamic clustering has intensified. SNOB [63], the computer program
developed by Chris Wallace, applies the Minimum [Message|Description] Length
[Encoding] (MML|MDL) principle to estimate the number of classes, the relative
abundance of each class, the distribution parameters for each variate within each
class, and the class to which each thing in the sample most probably belongs.
Kohonen’s “Self Organizing Maps” (SOM) [40] is a single-layered unsupervised
neural network model and uses competitive learning to train the weight vectors and
provides a way for corresponding multidimensional data in lower dimensional spaces,
which is a data compression method known as vector quantization. Thus, SOM is
applied for automatically determining the number of clusters. However, SOM suffers
from the following drawbacks [26]: 1) it depends on initial conditions, 2) its
performance depends on the learning rate parameter and the neighborhood function,
3) it only works well with hyper-spherical clusters and 4) it uses a fixed number of
outputs. Lorette et al. [2] proposed a new fuzzy objective function in order to find the
optimum number of clusters (referred as UFA). UFA starts with a large number of
clusters and modifies the membership values and centroids until a user predefined
value where the performance of the algorithm is very sensitive.
Evolutionary computation based algorithms have also been applied to overcome
the dynamic clustering problem. The GA has been applied to the dynamic clustering
problem by using its standard bit-string structure [69]. Omran et al. [51] proposed a
binary PSO algorithm for dynamic image clustering in which the activation of
centroids is based on the selection of 1 and 0. Das and Konar [11] introduced a
modified differential evolution algorithm (ACDE) based on adaptive crossover rate
for dynamic clustering, which determines active clusters through the activation
threshold. In addition, The fuzzy version of ACDE was suggested to overcome
dynamic image fuzzy clustering [15]. Kuo et al. [42] applied a hybrid algorithm of
PSO and GA for dynamic clustering in order to prevent the local minima problem. In
this algorithm, PSO and GA correspondingly try to improve the solutions and then
elitist selection is performed between the solutions of PSO and GA. Although a great
research effort has been put forward to obtain an ideal, accurately constructed and
visually meaningful image clustering performance by researchers, it still remains a
challenge.
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
4.1.1 Dynamic Clustering Methodology
Clustering validity indexes are defined as a class of statistical-mathematical
functions depending on the similarity or dissimilarity between the data points [12].
Clustering validity indexes are used for two main purposes. The first is to measure the
quality of clustering relative to others created by other methods or by the same
methods using different parameter values. The second is to determine the number of
clusters. The traditional approach for determining the number of clusters is to
repeatedly run the algorithm as input with a different number of classes and then to
choose the clustered results with the best validity measure [23]. A validity index
needs to consider two important aspects [11]: patterns in the same cluster should be as
similar to each other as possible, known as cohesion (compactness), and clusters
should be well separated (separation).
Many various validity indexes have been proposed. For hard clustering, Dunn’s
index (DI) [19], the Calinski-Harabasz index [8], the DB index [16], the CS measure
[9], the VI index [62] etc. and for fuzzy clustering, Partition coefficient [4], Partition
Entropy [3] and the Xie-Beni [67] index are the most known indexes. According to
[12], the DB, PC, CE and XB indexes cannot obtain the optimal cluster number in
some images and according to [50, 62] the performance of the VI index is superior to
DB and Dunn’s index which leads us to use the VI index as an objective function.
The VI index consists of two issues:
a) Intra-Cluster Separation: Intra-cluster separation satisfies well collected groups
known as compactness by calculating the average of all distances between the
patterns and their related centroids. It should be minimized and defined by
Equation (6);
=1
−∈
(6)
where = , , … , , … , is whole dataset, is a pattern with
dimensional feature space and is the number of patterns in .
b) Inter-Cluster Separation: Inter-cluster separation, which is the minimum
Euclidean distance between any pairs of cluster centroids, provides well
separated clusters and is used to measure the cluster separation. It should be
maximized and defined by Eq. (7);
= min{‖ − ‖ } ℎ ≠ (7)
where is the cluster centroid of and is the cluster centroid of .
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
In terms of intra and inter-cluster separation, the VI index is defined by Equation (8);
= ( × (μ, ) + 1) × (8)
where c is a user specified constant (generally chosen as 20, 25 or 30) and N(μ, ) is a
Gaussian distribution with mean and standard deviation, defined by Eq. (9):
(μ, ) =1
√2[ ( ) ](9)
where K is the number of clusters.
Fig. 6 The way to activate centroids
1. Choose randomly Nc number of cluster centroids from image data set;
2. for t=1:TC (Termination criteria)
Initialize the initial solutions by Eq. (5);
for cycle=1:MCN(Maximum Cycle Number)
for each solution Xi
· Activate the centroids as seen in Fig. 6;
· Find the Euclidean distance of all patterns to the activated centroidsby Eq. (10);
, = ( , − , ) = − (10)
· Assign patterns to the closest clusters;
· Calculate VI index by Eq. (8);
· Improve the solution using GNG;
b. Apply K-means to the remaining centroids in order to decrease theadverse effects of initial conditions;
c. Generate new cluster centroids as in step 1;
d. Merge new cluster set with remaining centroids;
Fig. 7 The pseudo code of the GB-ABC dynamic clustering algorithm
The main steps of the GB-ABC dynamic clustering algorithm introduced in this study
are presented in Fig. 7. As seen in Fig. 7, a set of centroids (S) is first randomly
generated or chosen from dataset. After that, each food source is uniformly initialized
by {0,1}. Thirdly, each food source is employed to activate centroids for evaluating
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
the value of VI index as seen in Fig. 6 and then each food source is updated through
GNG. When the stopping criterion (MCN) is satisfied, K-means is applied to the
activated centroid subset (Sbest) of the global best food source to decrease the adverse
effects of initial conditions. After that, the corresponding positions of S cluster set are
updated with the centroids of Sbest. This procedure is repeated again with the updated
S until the termination criteria (TC) is satisfied.
4.2 Knapsack Problem
Searching for the best configuration or set of parameters has become a significant
point in order to accomplish some objective criteria in the fields of industry,
transportation, logistic, computer science etc. Thus, the knapsack problem plays
important role in these applications. The knapsack problem is defined as selecting the
most valuable items without extrapolating the weight capacity. Some types of
knapsack problem are given below [38]:
· Single knapsack problem: For the items, there is only one knapsack.
· Multidimensional knapsack problem: More than one knapsack.
· Multiple-choice knapsack problem: The items are subdivided into pre-defined
classes and for each class, exactly one item must be selected.
· Bounded knapsack problem: Only a limited number of items can be selected.
The basic knapsack problem can be defined by Eqs. 10 and 11;
max ( ) = (10)
− ≤ 0
∈ {0,1} ℎ = (0,1 … , )
(11)
where and are the ℎ item weight and cost and W is the capacity of the
knapsack.
In order to describe the knapsack problem as a binary optimization problem, is
the ℎ binary solution such that = [ , , … … … , ]. The number of n is the
dimension of the whole population and is the ℎ location of ℎ solution. If the
value of is 1, the ℎ corresponding item is loaded in the knapsack. The objective
function can be defined as;
= + 0, − (12)
where is a predefined penalty factor and Eq. (12) is a maximization problem.
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
4.3 Numerical Problems
To apply a binary optimization algorithm to the numerical problems, bit strings
are needed to be converted to numerical values within the given boundaries. The bit
string is represented in real form by Eq. (13);
= +( )
2 − 1 ∗ ( − )(13)
where lb and ub are lower and upper boundaries of reel number representations, b
is the length of bitstring and bittodec represents the decimal form of bitstring.
To evaluate the performance of the proposed algorithm, 25 special benchmark
functions, comprising unimodal, multimodal, shifted, rotated and hybrid composite
functions provided by CEC2005 [55] are employed. The number of variables (n) for
the search space is considered as 10 and the length of bit string for each variable is
chosen as 15. Thus, the length of a solution string is 10x15 (150).
5 Experimental Results
The experimental results are considered in three subsections: performance analysis
of the proposed algorithm on dynamic image clustering, on knapsack problem and on
numerical test problems. The results are presented over 30 simulations in terms of
mean values and standard deviations in the tables. The best results are denoted in
bold.
5.1 Dynamic Image Clustering
To compare the GB-ABC algorithm with other algorithms, including BPSO,
QBPSO, GA and DisABC in terms of dynamic image clustering on six benchmark
images such as Tahoe, Brain MRI, Jet, Lena, Mandril and Pepper, two simple criteria
are considered: 1) the quality of clustering determined by the VI index and 2) the
ability of obtaining optimal number of cluster. In addition to the evolutionary
computation based algorithms, the results of SOM and UFA from [50] and the results
of Snob algorithms from [62] are put forward to enhance the comparative analysis.
The optimal number of clusters for each image, Tahoe, Brain MRI, Jet, Lena, Mandril
and Pepper is demonstrated by a research group [62] in the range of [3,7], [4,8], [5,7],
[5,10], [5,10] and [6,10], respectively. However, that makes impossible to analyse the
clustering algorithms in terms of optimal number of clusters. In this work, the optimal
number of clusters is defined by Equation (14);
= + 1(14)
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
Therefore, the optimal number for Tahoe, Brain MRI and Pepper is determined as 4, 5
and 7, respectively; for the others, Jet, Lena and Mandril it is defined as 6 by Eq. (14).
It should be notified that the optimal number for the pepper image was also selected
as 7 in [11] and the number of clusters for each clustered image shown in Fig. 8 and 9
can be accepted as sufficient i.e. no more clusters need to be specified.
In the experiments of dynamic image clustering, the population size is chosen as
50 and the maximum number of evaluations is restricted to 20000. The other
parameters of the evolutionary computation based algorithms are as follows: the limit
values of DisABC and GB-ABC are chosen as 100 and 20, respectively; the
parameters of BPSO are selected as c = 2, c = 2, w = 0.9, w = 0.4 and
V = 6; the initeria and end weight parameters of QBPSO are selected as
w =0.5π and w =0.0001π; the values of crossover rate and mutation rate in
GA are selected as 0.8 and 0.2, respectively.
In the experiments, the following parameters are chosen for the VI index; c is
chosen 25; for Tahoe, µ is chosen 1, for the images brain MRI, jet, Lena and mandril,
µ is chosen as 2 and for pepper, µ is chose as 3 since it requires more than 6 clusters.
In other words, µ is set to 0 or 1 for fewer clusters (3 or 4), µ is set to 2 for 5 or 6
clusters and to 3 for more than 6 clusters. For all cases, σ is set to 1. Termination
criterion (TC) is set to 2 i.e. algorithms run two times to get optimal solution for each
simulation.
Tables 1 and 2 demonstrate the results of the VI index and the obtained number of
clusters. Table 1 reveals that the GB-ABC algorithm outperforms the others in terms
of minimizing the objective function except Mandril and as a result it can be extracted
that it dominates the other algorithms in satisfying clustering quality. In fact, the GB-
ABC algorithm can be accepted as the most robust algorithm by satisfying the
minimum standard deviation values almost in all cases in the values obtained from the
VI index. Table 2 indicates that the GB-ABC algorithm generally performs better than
the others in terms of finding the optimal number of clusters. Meanwhile, it is clearly
seen from Table 2 that the GA, SOM, UFA and Snob algorithms generally have a
much poorer performance in terms of finding the optimal cluster number. For
instance, the mean value of cluster number for jet is 6.733, 14, 20 and 22 for GA,
SOM, UFA and Snob, respectively; however, the optimal number of clusters is only 6
for the jet. For the images Jet and Pepper where the GB-ABC only cannot get the best
mean values of the obtained number of clusters, BPSO and QBPSO gets the better
results, but the obtained standard deviation values of BPSO and QBPSO are higher
than GB-ABC. It can be concluded that the GB-ABC algorithm performs well in
terms of minimizing the VI index and finding the optimal number of clusters.
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
Fig. 10 shows the convergence graphics of the algorithms for the selected images
Jet, Lena and Pepper where TC is set to 2 and MCN is set to 200 i.e. the number of
iterations becomes 400. It should be noted that the number of iterations are taken
similar in all algorithms only for the case study of convergence graphs since it is not
possible to establish a comparative study of algorithm convergence through the
number of evaluations. As seen in Fig. 10, the algorithms may have fluctuations after
200 iterations since TC is set to 2 i.e. the algorithms carry out their processes in two
stages: during the first stage from iteration 1 to 200, the algorithms search for the best
solution on randomly generated cluster centroid set and during the second stage from
iteration 201 to 400, the algorithms begin the search process again with the new
randomly generated population on the modified cluster centroid set obtained from the
first stage (see Fig. 7). It can be clearly inferred from the Fig. 10 that GB-ABC has
less fluctuations than the others, which shows its robustness.Table 1. The Obtained Mean VI Index Values for Image Clustering
Images GB-ABC DisABC QBPSO BPSO GATahoe 0.0569
(0.0067)0.0574
(0.0062)0.0604
(0.0073)0.0644
(0.0178)0.0884
(0.0377)MRI 0.0485
(0.0068)0.0499
(0.0069)0.0521
(0.0061)0.1339
(0.1181)0.1778
(0.2129)Jet 0.0945
(0.0145)0.0959
(0.0305)0.1238
(0.1000)0.1366
(0.1183)0.1517
(0.0595)Lena 0.0982
(0,0088)0.1032
(0.0141)0.1077
(0.0167)0.1126
(0.0130)0.1395
(0.0259)Mandril 0.1082
(0,0124)0.1045
(0,0111)0.1023
(0.0116)0.1077
(0,0115)0.1419
(0.0324)Pepper 0.1069
(0.0154)0.1201
(0.0150)0.1202
(0.0317)0.1323
(0.0460)0.1662
(0.0595)
Table 2. The Obtained Mean Cluster Numbers for Image Clustering
Images# of
clustersGB-ABC DisABC QBPSO BPSO GA SOM UFA Snob
Tahoe 4 4.066(0.253)
4.133(0.345)
4.566(0.954)
4.2(0.482)
4.791(1.178)
4 20 NA
MRI 5 5.133(0.434)
5.4(0.621)
5.724(1.130)
6.166(1.341)
7.29(2.123)
NA NA NA
Jet 6 5.7(0.836)
5.5(0.861)
6.366(1.299)
6.033(0.964)
6.733(1.680)
14 20 22
Lena 6 5.7(0.595)
5.666(0.547)
6.7(1.417)
6.695(1.063)
7.1(1.516)
20 20 31
Mandril 6 5.933(0.739)
6.166(1.019)
7.066(1.172)
6.9(1.471)
7.4(1.652)
20 20 42
Pepper 7 6.633(0.668)
6.566(0.568)
7.133(1.224)
7.333(1.212)
7.466(1.105)
20 20 39
5.2 Knapsack Problem
Six benchmark datasets, including 250, 500 and 750 items are chosen from [73] to
show the effectiveness of the GB-ABC algorithm over the BPSO, QBPSO, DisABC
and GA algorithms. In the experiments of the knapsack problem, population size is
chosen as 30 and maximum number of evaluations is set to 60000. The limit values of
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
DisABC and GB-ABC are selected as 100 and 500, respectively and the parameters
of BPSO and GA are selected as in the experiments of dynamic image clustering. It
should be noted that the convergence structure of the DisABC algorithm leads us to
use higher limit values.Table 3. The Obtained Results for Knapsack
ItemSet
Capacity GB-ABC DisABC QBPSO BPSO GA
250_1 6184
Best 9548 9023 9559 7926 9504
Mean 9520(13.913)
8824.67(117.26)
9518(22.368)
7767.6(86.300)
9448.23(31.854)
Worst 9491 8526 9463 7614 9353
p-val 0,9 2,9e-11 N/A 2,9e-11 4,9e-11
250_2 6536
Best 10217 9716 10219 8760 10172
Mean10177.3(23.111)
9539.37(90.805)
10187.1(17.961)
8600.53(78.171)
10123.8(30.001)
Worst 10126 9386 10152 8477 10067
p-val N/A 3,0e-11 0,09 2,9e-11 1,9e-08
500_1 13499
Best 20479 18872 20235 16547 20370
Mean 20383.9(33.703)
18531.8(172.92)
20108.6(75.477)
16261(132.60)
20244.1(60.248)
Worst 20331 18224 19951 15994 20102
p-val N/A 3.0e-011 3.0e-011 3.0e-011 1.5e-010
500_2 13743.5
Best 20633 18774 20500 16446 20582
Mean 20563.7(39.893)
18415.7(158.02)
20309.5(80.504)
16301.5(88.076)
20456.8(59.543)
Worst 20467 18086 20126 16129 20349
p-val N/A 3,0e-11 3,3e-11 3,0e-11 6,5e-09
750_1 20351.5
Best 30300 27243 29731 24174 30212
Mean 30097(95.417)
26630.3(297.91)
29475.4(154.61)
23644.7(197.49)
30011(95.943)
Worst 29957 26046 29133 23312 29816
p-val N/A 3,0e-11 3,0e-11 3,0e-11 0.0021
750_2 20450
Best 31641 28275 31008 24918 31660
Mean31430.1(95.149)
27838.6(285.36)
30767.3(114.08)
24528.6(162.61)
31361.4(125.72)
Worst 31197 27039 30584 24280 31077
p-val N/A 3.0e-011 3.0e-011 3.0e-011 0.0021
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
The maximum capacities and dimensions of knapsack datasets with the obtained
results in terms of worst, mean, best and Wilcoxon rank sum test (p-val) values at 5%
significance level are presented in Table 3. In this table, N/A indicates “non-
applicable” reflecting the corresponding algorithm cannot statistically compare with
itself in test. It can be clearly extracted from Table 3 that GB-ABC remains overall
commendable. In fact, it excels in knapsack problem by achieving higher results in
both worst and best cases compared to its contenders. ‘250_2’ is the only case where
GB-ABC cannot get the best results. It can be also clearly inferred from Table 3 that
the gap between the GB-ABC and the existing algorithms is increased when the size
of the problem becomes larger. Moreover, the difference between GB-ABC and the
others can be statistically demonstrated via p-vals. Consequently, the GB-ABC
algorithm is far superior to the other algorithms in terms of almost all properties such
as maximizing the cost of the knapsack and satisfying the best standard deviation.
5.3 Numerical Problems
In the experiments of numerical optimization, the population size is chosen as 40
and the maximum number of evaluations is restricted to 320000. The limit values of
GB-ABC and DisABC are set to 400 and 1000, respectively and the user-specified
parameters of the BPSO, QBPSO and GA are selected as in the experiments of
dynamic image clustering and knapsack problems.
Concerned with ten dımensional search space, Table 4 presents the obtained results of
the CEC2005 problems in terms of mean, standard deviation, rankings and Wilcoxon
rank sum test (p-val) values at 10% significance. The mean values in Table 4 clearly
reflect that GB-ABC performs well in unimodal functions (F1-F5), almost all cases of
composite functions (F15-F25), except F25. Although GB-ABC cannot get the first
degree in multimodal functions (F6-F14), except F6; it can be clearly seen that the
gap between the GB-ABC and the other algorithms is very narrow i.e. the results of
the algorithms in (F7-F14) are very close to each other. For instance, the mean value
difference between GB-ABC and QBPSO in F8 is only 0.06. Moreover, p-values also
mostly support the achievement of the GB-ABC algorithm except some cases
concerning the QBPSO algorithm. Concerned with the total rankings, it is seen that
GB-ABC gets the first position among all existing algorithms and the other
algorithms such as QBPSO, BPSO, GA and DisABC are ranked as second, third,
fourth and last, respectively.
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
Table 4. The Obtained Results of CEC2005 Benchmark Functions
GB-ABCMean(Std)
[Rank]p-val
DisABCMean(Std)
[Rank]p-val
QBPSOMean(Std)
[Rank]p-val
BPSOMean(Std)
[Rank]p-val
GAMean(Std)
[Rank]p-val
F1 -336.54(112.75)
[1]N/A
-103.56(142.39)
[5]7.6e-008
-334.94(78.78)
[2]0.0592
-271.06(131.36)
[3]0.0035
-190.99(171.40)
[4]8.1e-004
F2 -280.57(151.46)
[1]N/A
-74.43(170.08)
[5]5.8e-006
-270.45(173.75)
[2]0.1160
-263.83(123.33)
[3]0.0397
-143.43(253.23)
[4]0.0028
F3 -303.56(145.54)
[1]N/A
-44.47(161.15)
[5]5.1e-007
-291.45(153.19)
[2] 0.0991
-257.65(150.35)
[3] 0.0073
-124.02(265.35)
[4]1.9e-004
F4 -341.57(96.855)
[1]N/A
-64.12(154.92)
[5] 1.1e-008
-332.52(122.57)
[2] 0.4517
-290.42(108.15)
[3] 0.0025
-118.42(269.34)
[4]2.1e-005
F5 -136.87(177.21)
[1]N/A
83.19(155.51)
[5]4.4e-006
-115.02(166.90)
[2]0.1302
-112.83(119.63)
[ 3] 0.0107
13.991(254.74)
[4]6.4e-004
F6 533.92(127.21)
[1]N/A
740.04(183.25)
[5]2.4e-006
577.09(138.99)
[3]0.1668
567.83(101.27)
[2]0.1858
719.53(261.13)
[4]9.0e-004
F7 -167.25(9.06)
[3]N/A
-149.06(15.47)
[5] 7.6e-006
-171.42(8.82)
[1]0.0436
-169.07(9.07)
[2]0.4464
-160.71(12.02)
[4] 0.0163
F8 -119.73(0.09)
[2]N/A
-119.66(0.08)
[4] 0.0016
-119.79(0.07)
[1]0.0040
-119.71(0.06)
[3] 0.1031
-119.67(0.12)
[5] 0.0066
F9 -309.67(0.11)
[4]N/A
-309.69(0.07)
[3]0.0056
-309.81(0.06)
[1] 1.7e-006
-309.71(0.06)
[2]0.3555
-309.66(0.14)
[5]0.0877
F10 -309.66(0.12)
[4]N/A
-309.67(0.07)
[3]0.0010
-309.78(0.08)
[1]2.6e-004
-309.70(0.05)
[2]0.2226
-309.66(0.14)
[4] 0.4154
F11 110.32(0.10)
[4]N/A
110.29(0.09)
[3] 1.9e-004
110.19(0.09)
[1]1.0e-005
110.28(0.07)
[2]0.0798
110.36(0.17)
[5] 0.0877
F12 -439.75(0.09)
[2]N/A
-439.67(0.08)
[4]4.1e-008
-439.81(0.07)
[1]0.0184
-439.70(0.06)
[3]0.0224
-439.62(0.12)
[5]6.7e-005
F13 -109.68(0.12)
[3]N/A
-109.68(0.08)
[3]0.0030
-109.81(0.09)
[1]3.3e-004
-109.70(0.05)
[2] 0.2226
-109.64(0.13)
[4]0.0161
F14 -279.64(0.09)
[4]N/A
-279.66(0.08)
[3]0.0055
-279.75(0.08)
[1]1.0e-005
-279.71(0.05)
[2]3.5e-004
-279.64(0.11)
[4] 0.4260
F15 412.57(189.24)
[1]N/A
472.64(48.95)
[5] 0.0169
435.10(174.28)
[3]0.0565
458.75(105.22)
[4]0.0079
419.09(188.33)
[2]0.2489
F16 411.04(164.88)
[1]N/A
472.43(67.26)
[5]0.0070
449.07(160.37)
[4]0.0423
438.23(104.84)
[2]0.0179
444.86(184.35)
[3]0.0250
F17 367.16(165.15)
472.22(59.78)
398.99(183.57)
466.92(111.92)
446.07(171.13)
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
[1]N/A
[5]6.3e-005
[2]0.1302
[5]7.4e-005
[4]0.0216
F18 284.38(171.02)
[1]N/A
384.49(46.09)
[5]5.9e-004
288.25(201.08)
[2]0.2905
356.14(116.55)
[4]0.0133
332.12(188.89)
[3]0.0682
F19 291.25(152.52)
[1]N/A
372.68(51.72)
[5]0.0033
296.73(189.52)
[2]0.1868
336.29(112.08)
[3]0.0360
342.78(182.99)
[4]0.0019
F20 291.27(204.83)
[1]N/A
369.94(57.39)
[5]0.0066
331.57(173.24)
[3]0.1055
347.48(140.15)
[4]0.0077
314.62(181.76)
[2]0.0807
F21 632.88(178.25)
[1]N/A
696.78(63.13)
[4]0.0018
701.74(167.29)
[5]0.0151
691.12(121.87)
[3]0.0029
672.88(177.95)
[2]0.0743
F22 618.47(176.95)
[1]N/A
726.58(60.58)
[5] 9.3e-005
646.42(171.96)
[2]0.1206
694.21(103.99)
[3]4.2e-004
696.14(185.74)
[4]0.0086
F23 656.22(163.05)
[1]N/A
722.61(64.67)
[5]0.0193
694.85(164.65)
[4]0.0772
665.93(91.22)
[2]0.0345
672.48(184.78)
[3]0.0532
F24 505.23(169.10)
[1]N/A
604.26(52.72)
[5] 1.4e-004
561.22(176.22)
[3]0.0179
553.72(74.82)
[2] 3.7e-004
573.05(199.48)
[4]0.0066
F25 558.69(167.03)
[2]N/A
610.83(57.06)
[4]0.0070
522.70(172.31)
[1]0.0186
631.59(115.15)
[5] 0.0077
587.57(186.89)
[3]0.0842
6 Conclusion
Many engineering problems such as clustering, uncapacited facility location
problem (UFLP) and knapsack problem etc. are accepted as binary optimization
problems. In this paper, a novel binary artificial bee colony algorithm is proposed
based on genetic operators (GB-ABC). The main advantage of the proposed algorithm
is simple to implement. In other words, it does not include any complicated structure.
The other advantage of the proposed algorithm is that it is not only applicable for
specific binary problems and it can be also used for general binary optimization
problems. This is demonstrated by applying it to the different kind of binary
problems: dynamic image clustering and knapsack problems. Besides, its suitability is
tested on numerical optimization problems. From the simulation results, it is clear that
the GB-ABC algorithm outperforms all other algorithms in all handled problems.
Hence, it can be successfully applied to optimize general binary optimization
problems. As future works, it is planned to evaluate the performance of the proposed
algorithm on other binary optimization problems.
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
REFERENCES
[1] B. Akay, D. Karaboga, A modified Artificial Bee Colony algorithm for real-parameter optimization, Information Sciences, 192 (2012) 120-142.
[2] L. Anne, X. Descombes, J. Zerubia, Fully Unsupervised Fuzzy Clustering withEntropy Criterion, in: D. Xavier, Z. Josiane (Eds.) 15th International Conference onPattern Recognition (ICPR'00), 2000, pp. 3998-3998.
[3] J.C. Bezdek, Cluster validity with fuzzy sets, Journal of Cybernetics, 3 (1974) 58–72.
[4] J.C. Bezdek, Numerical taxonomy with fuzzy sets, Journal of Math. Biol., (1974)157–171.
[5] M. Birattari, Swarm intelligence, Scholarpedia, 2 (2007) 1462.
[6] S. Biswas, S. Das, S. Kundu, G. Patra, Utilizing time-linkage property in DOPs:An information sharing based Artificial Bee Colony algorithm for tracking multipleoptima in uncertain environments, Soft Comput, 18 (2013) 1-14.
[7] S. Biswas, S. Kundu, D. Bose, S. Das, P.N. Suganthan, B.K. Panigrahi, Migratingforager population in a multi-population Artificial Bee Colony algorithm withmodified perturbation schemes, in: IEEE Symposium on Swarm Intelligence (SIS),2013, pp. 248-255.
[8] R.B. Calinski, J. Harabasz, A dendrite method for cluster analysis, Commun. Stat.,3 (1974) 1–27.
[9] C.H. Chou, M.C. Su, E. Lai, A new cluster validity measure and its application toimage compression, Pattern Anal. Appl., 7 (2004) 205–220.
[10] C.Y. Chung, Y. Han, W. Kit-Po, An Advanced Quantum-Inspired EvolutionaryAlgorithm for Unit Commitment, IEEE Transactions on Power Systems, 26 (2011)847-854.
[11] S. Das, A. Abraham, A. Konar, Automatic Clustering Using an ImprovedDifferential Evolution Algorithm, IEEE Transactions on Systems, Man andCybernetics, Part A: Systems and Humans, 38 (2008) 218-237.
[12] S. Das, A. Abraham, A. Konar, Metaheuristic Clustering, Springer-Verlag,Berlin Heidelberg, 2009.
[13] S. Das, S. Biswas, S. Kundu, Synergizing fitness learning with proximity-basedfood source selection in artificial bee colony algorithm for numerical optimization,Applied Soft Computing, 13 (2013) 4676-4694.
[14] S. Das, S. Biswas, B.K. Panigrahi, S. Kundu, D. Basu, A Spatially InformativeOptic Flow Model of Bee Colony With Saccadic Flight Strategy for GlobalOptimization, IEEE Transactions on Cybernetics, 44 (2014) 1884-1897.
[15] S. Das, A. Konar, Automatic image pixel clustering with an improveddifferential evolution, Applied Soft Computing, 9 (2009) 226-236.
[16] D.L. Davies, D.W. Bouldin, A cluster separation measure, IEEE Trans. PatternAnal. Mach. Intell., 1 (1979) 224–227.
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
[17] R. Dawkins, The Selfish Gene, Oxford: Oxford Univ. Pres, 1976.
[18] M. Dorigo, Optimization Learning And Natural Algorithms, Ph.D Thesis,Politecnico Di Milano, Italy, 1992.
[19] J.C. Dunn, Well separated clusters and optimal fuzzy partitions, J. Cybern., 4(1974) 95–104.
[20] R. Eberhart, J. Kennedy, A new optimizer using particle swarm theory, in: 6thInternational Symposium on Micro Machine and Human Science, 1995.
[21] S. Gordon, Unsupervised Image Clustering using Probabilistic ContinuousModels and Information Theoretic Principles, Tel-Aviv University, Tel-Aviv 69978,Israel, 2006.
[22] O. Haddad, A. Afshar, M. Mariño, Honey-Bees Mating Optimization (HBMO)Algorithm: A New Heuristic Approach for Water Resources Optimization, WaterResources Management, 20 (2006) 661-680.
[23] M. Halkidi, M. Vazirgiannis, Clustering validity assessment: Finding the optimalpartitioning of a data set, in: IEEE International Conference on Data Mining (ICDM2001), San Jose, California, USA, 2001 pp. 187–194.
[24] E. Hancer, C. Ozturk, D. Karaboga, Artificial Bee Colony Based ImageClustering Method, in: IEEE Congress on Evolutionary Computation, CEC 2012,Brisbane, Australia, 2012, pp. 1-5.
[25] E. Hancer, C. Ozturk, D. Karaboga, Extraction of Brain Tumors from MRIImages with Artificial Bee Colony based Segmentation Methodology, in:ELECO'2013, Bursa, Turkiye, 2013.
[26] A.K. Jain, M.N. Murty, P.J. Flynn, Data Clustering: A Review, ACM ComputingSurveys, 31 (1999) 264-323.
[27] D. Karaboga, An idea based on honey bee swarm for numerical optimization, in,Technical Report-TR06, Erciyes University, Engineering Faculty, ComputerEngineering Department 2005.
[28] D. Karaboga, Artificial bee colony algorithm, Scholarpedia, (2010) 6915.
[29] D. Karaboga, B. Akay, A survey: algorithms simulating bee swarm intelligence,Artificial Intelligence Review, 31 (2009) 61-85.
[30] D. Karaboga, B. Basturk, A powerful and Efficient Algorithm for NumericalFunction Optimization: Artificial Bee Colony (ABC) Algorithm, Journal of GlobalOptimization, 39 (2007) 459-471.
[31] D. Karaboga, B. Basturk, On the performance of artificial bee colony (ABC)algorithm, Applied Soft Computing, 8 (2008) 687-697.
[32] D. Karaboga, B. Gorkemli, C. Ozturk, N. Karaboga, A comprehensive survey:artificial bee colony (ABC) algorithm and applications, Artificial Intelligence Review,42 (2014) 21-57.
[33] D. Karaboga, S. Okdem, C. Ozturk, Cluster Based Wireless Sensor NetworkRouting using Artificial Bee Colony Algorithm, Wireless Networks, 18 (2012) 847-860.
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
[34] D. Karaboga, C. Ozturk, Neural Networks Training By Artificial Bee ColonyAlgorithm On Pattern Classification, Neural Netw. World, 19 (2009) 279-292.
[35] D. Karaboga, C. Ozturk, A novel clustering approach: Artificial Bee Colony(ABC) algorithm, Applied Soft Computing, 11 (2011) 652-657.
[36] M.H. Kashan, A.H. Kashan, N. Nahavandi, A novel differential evolutionalgorithm for binary optimization, Comput Optim Appl, 55 (2013) 481-513.
[37] M.H. Kashan, N. Nahavandi, A.H. Kashan, DisABC: A new artificial bee colonyalgorithm for binary optimization, Appl. Soft Comput., 12 (2012) 342-352.
[38] H. Kellerer, U. Pferschy, D. Pisinger, Knapsack Problems, Springer, 2004.
[39] M.S. Kiran, M. Gunduz, XOR-based artificial bee colony algorithm for binaryoptimization, Turkish Journal of Electrical Engineering & Computer Sciences, 21(2013) 2307-2328.
[40] T. Kohonen, Self-Organizing Maps, Springer-Verlag, Springer Series inInformation Sciences, NewYork, USA, 1995.
[41] R. Krishnapuram, J.M. Keller, A possibilistic approach to clustering, IEEETransactions on Fuzzy Systems, 1 (1993) 98-110.
[42] R.J. Kuo, Y.J. Syu, Z.-Y. Chen, F.C. Tien, Integration of particle swarmoptimization and genetic algorithm for dynamic clustering, Information Sciences, 195(2012) 124-140.
[43] P. Lucic, D. Teodorovic, Bee system: modeling combinatorial optimizationtransportation engineering problems by swarm intelligence, in: Preprints of theTRISTAN IV Triennial Symposium on Transportation Analysis, Sao Miguel, AzoresIslands, Portugal, 2001, pp. 441–445.
[44] M. Ma, J. Liang, M. Guo, Y. Fan, Y. Yin, SAR image segmentation based onArtificial Bee Colony algorithm, Appl. Soft Comput., 11 (2011) 5205-5214.
[45] J. MacQueen, Some methods for classification and analysis of multivariateobservations, in: 5th Berkeley Symp. Math. Stat. Probability, 1967, pp. 281–297.
[46] J. Magalhaes-Mendes, A Comparative Study of Crossover Operators for GeneticAlgorithms to Solve the Job Shop Scheduling Problem WSEAS Transactions onComputers, 12 (2013) 164-173.
[47] M. Millonas, Swarms, phase transitions and collective intelligence, Addison-Wesley, Reading, MA, Reading, 1994.
[48] A. Mukhopadhyay, U. Maulik, S. Bandyopadhyay, Multiobjective GeneticAlgorithm-Based Fuzzy Clustering of Categorical Attributes, IEEE Transactions onEvolutionary Computation, 13 (2009) 991-1005.
[49] F. Neri, C. Cotta, Memetic algorithms and memetic computing optimization: Aliterature review, Swarm and Evolutionary Computation, 2 (2012) 1-14.
[50] M. Omran, Particle Swarm Optimization Methods for Pattern Recognition andImage Processing, Ph.D Thesis, University of Pretoria, Environment and InformationTechnology, 2004.
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
[51] M.G.H. Omran, A. Salman, A.P. Engelbrecht, Dynamic clustering using particleswarm optimization with application in image segmentation, Pattern Anal. Appl., 8(2006) 332-344.
[52] C. Ozturk, E. Hancer, D. Karaboga, Improved Clustering Criterion for ImageClustering with Artificial Bee Colony Algorithm, Pattern Anal. Appl., (2014).
[53] C. Ozturk, E. Hancer, D. Karaboga, Color Quantization: A Short Review and AnApplication with Artificial Bee Colony Algorithm, Informatica, 25 (2014) 485-503.
[54] C. Ozturk, D. Karaboga, B. Gorkemli, Probabilistic Dynamic Deployment ofWireless Sensor Networks by Artificial Bee Colony Algorithm, Sensors, 11 (2011)6056-6065.
[55] P. N. Suganthan, N. Hansen, J. J. Liang, K. Deb, Y.-P. Chen, A. Auger, S.Tiwari, Problem Definitions and Evaluation Criteria for the CEC 2005 SpecialSession on Real-Parameter Optimization, Nanyang Technological University,Singapore, IIT Kanpur, India, 2005
[56] G. Pampara, A.P. Engelbrecht, Binary artificial bee colony optimization, in:IEEE Symposium on Swarm Intelligence (SIS), 2011, pp. 1-8.
[57] D. Pham, A. Ghanbarzadeh, E. Koc, S. Otri, S. Rahim, M. Zaidi, The beesalgorithm, in: Manufacturing Engineering Centre, Cardiff University, UK, 2005.
[58] D.L. Pham, Spatial Models for Fuzzy Clustering, Computer Vision and ImageUnderstanding, 84 (2001) 285-297.
[59] J. Puzicha, T. Hofmann, J.M. Buhmann, Histogram Clustering for UnsupervisedImage Segmentation, in: CVPR'99, 1999.
[60] S.Raziuddin, S. Sattar, R. Lakshmi, M. Parvez, Differential artificial bee colonyfor dynamic environment, in: K.B. Meghanathan N, Nagamalai (Ed.) Advances inComputer Science and Information Technology, Communications in Computer andInformation science, Springer, Berlin, 2011, pp. 59–69.
[61] D. Teodorovic, M. Dell’orco, Bee colony optimization - a cooperative learningapproach to complex transportation problems, in: 16th mini-EURO Conference onAdvanced OR and AI Methods in Transportation, 2005, pp. 51–60.
[62] R.H. Turi, Clustering-Based Colour Image Segmentation, Ph.D. Thesis, MonashUniversity, Australia, 2001.
[63] C. Wallace, D. Dowe, Intrinsic Classification by MML – the snob program, in:Seventh Australian Joint Conference on Artificial Intelligence, UNE,Armidale, NSW,Australia, 1994, pp. 37-44.
[64] H. Wedde, M. Farooq, The wisdom of the hive applied to mobile ad-hocnetworks, in: Swarm Intelligence Symposium SIS 2005, 2005, pp. 341–348.
[65] H. Wedde, M. Farooq, Y. Zhang, BeeHive: An Efficient Fault-Tolerant RoutingAlgorithm Inspired by Honey Bee Behavior, in, 2004, pp. 83-94.
[66] B. Wu, S. Fan, Improved artificial bee colony algorithm with chaos, in: Y. Yu, Z.Yu, J. Zhao (Eds.) Computer science for Environmental Engineering andEcoinformatics, Communications in Computer and Information Science, Springer,Berlin, 2011, pp. 51–56.
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
[67] X. Xie, G. Beni, Validity measure for fuzzy clustering, IEEE Trans. PatternAnal. Machine Learning, 3 (1991) 841–846.
[68] X. Xu, X. Lei, Multiple sequence alignment based on ABC_SA, in: D.H. WangF, Gao Y, Lei J (Ed.) Lecture notes in computer science, Springer, Berlin, 2010, pp.98–105.
[69] Z. Yan, Z. Chun-Guang, W. Sheng-Sheng, H. Lan, A dynamic clustering basedon genetic algorithm, in: International Conference on Machine Learning andCybernetics, 2003, pp. 222-224 Vol.221.
[70] X. Yang, Engineering optimizations via nature-inspired virtual bee algorithms,in: J. Mira, J. lvarez (Eds.) Artificial intelligence and knowledge engineeringapplications: a bioinspired approach, Lecture notes in computer science, Springer,2005.
[71] O. Yew-Soon, L. Meng-Hiot, C. Xianshun, Memetic Computation-Past, Present& Future [Research Frontier], IEEE Computational Intelligence Magazine, 5 (2010)24-31.
[72] G. Zhu, S. Kwong, Gbest-Guided Artificial Bee Colony Algorithm for NumericalFunction Optimization, Appl Math Comput, (2010).
[73] E. Zitzler, M. Laumanns, Test Problems and Test Data for MultiobjectiveOptimizers, in: Systems Optimization, Test Problem Suite, ETH, Zurich,2014. http://www.tik.ee.ethz.ch/sop/download/supplementary/testProblemSuite/
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
A (Tahoe) C (MRI) E (JET)
B (Tahoe) D (MRI) F (JET)Fig. 8 The original input images and the clustered images by GB-ABC, respectively for tahoe (A,B), MRI (C,D) and jet (E,F) images.
A(Lena) C(mandril) E(pepper)
B(Lena) D(mandril) F(pepper)Fig. 9 The original input images and the clustered images by GB-ABC, respectively for Lena (A,B), mandril (C,D), and pepper (E,F) images.
* Corresponding Author: Celal OZTURK Tel: +90 352 207 66 66 #32581
Fig. 10 The convergence graph of the algorithms for Jet, Lena and Pepper images.