levy flight distribution for scout bee in artificial bee …lnse.org/papers/55-ca034.pdfartificial...
TRANSCRIPT
Abstract—There are many population based optimization
methods used for numeric functions and engineering problems.
The biggest problem of these methods is setting the balance
between exploration and exploitation. The artificial bee colony
algorithm proposed by Karaboga gives better results compared
to other known nature-inspired methods. Yet, while the ABC
algorithm is better in the exploration part, which is known as
exploring new places, it is not well enough in the exploitation
part, which is explained as exploiting the results found. To
overcome this problem, instead of random distribution of the
scout bees in the search space in ABC algorithm, this paper
proposed the Levy Flight ABC (LFABC) algorithm performing
the distribution using Levy Flight method. By this way, it was
ensured for the ABC algorithm to improve the exploitation. The
two methods were tested on 10 benchmark functions, and the
proposed method was seen to perform the results better.
Index Terms—Artificial bee colony, levy flight, levy
distribution, optimization.
I. INTRODUCTION
In the last 10-20 years, many population based algorithms
were proposed to solve numerical benchmark functions and
engineering problems, such as particle swarm optimization
[1], ant colony optimization [2], genetic algorithms [3],
artificial bee colony [4] and so on. These algorithms which
are also referred to as biological-inspired were generally
proposed through inspiration from social and foraging
behaviors of particular animals. Being inspired from social
and foraging behaviors of bees, ABC was first proposed in
2005 by Karaboga [4].
Showing a better performance compared to other
optimization techniques, the ABC algorithm [5], considering
its simplicity and efficiency, was used in the fields such as
function optimization [6], [7], vehicle routing problem [8],
data clustering [9], image processing and segmentation [10],
[11] , electric load forecasting [12], engineering design [13]
and so on.
There are two important points for biological-inspired
algorithms: exploration and exploitation. The exploration
part is concerned the ability of autonomously seeking for the
global optimum, whereas the exploitation part is related to
the ability of applying the existing knowledge to look for
better solutions [14]. Although the ABC algorithm is more
effective in many numerical benchmark functions compared
to other algorithms, the ABC algorithm has some deficiencies.
While ABC is good in exploration, it is not good enough in
Manuscript received February 9, 2013; revised April 14, 2013. This work
was supported by Scientific Research Project of Selcuk University.
Hüseyin Hakli and Harun Uğuz are with the Computer Engineering
Department, Selcuk University, Konya, Turkey (e-mail: hhakli@
selcuk.edu.tr, harun_uguz@ selcuk.edu.tr).
exploitation, which refers to ability of applying the existing
knowledge to look for better solutions. In other words, while
the ABC algorithm can perform global search better, it is
weak in local search. Many ABC variants were proposed to
overcome this problem. The common goal of the new
algorithms proposed was to strengthen exploitation by
ensuring the ABC algorithms perform local search better.
Inspired from the PSO algorithm, Zhu ve Kwong [15]
proposed gbest-guided ABC (GABC) algorithm by
incorporating the information of global best (gbest) solution
into the solution search equation to improve the exploitation.
And inspired from the DE algorithm, Gao et al. [16]
improved exploitation by ensuring the ABC algorithm
perform search only around the best solution. Xiang ve An [6]
proposed the ERABC algorithm, a combinatorial solution
search equation that is introduced to accelerate the search
process instead of ABC solution search equation, and a
chaotic search technique that is employed on scout bee phase.
Li et al. [14] used 3 different solution search equations in
parallel, and chose the one with better results, and developed
the ABC algorithm by performing this phase for both
employed and onlooker bees. Alatas [17] used chaotic maps
for parameter adaption, initial the artificial colony and scout
of employed bee to prevent the ABC to get stuck on local
solutions. In this paper, if improvement as much as the
predetermined limit value number was not attained in the
ABC algorithm, that food source was abandoned, and instead
of selecting a random food source within the search space, it
was ensured a new food source search be performed
according to Levy Flight distribution. GlobalMin (the food
source giving the best result at that time) information were
also used while searching for a new food source with this
distribution. Thus, by performing the search around
GlobalMin, it was that the ABC algorithm performs local
search more effectively, and that the scout bees that were not
useful enough become more useful. According to the
experimental results, the proposed method was seen to give
better results compared to the ABC algorithm and to prevent
being stuck on local minimum in several functions.
The rest of the paper is divided as follows. In Section II,
original ABC algorithm and Levy flight method are
presented. The proposed approach is detailed in Section III.
In Section IV, the experimental results and comparison of the
methods are presented. As a final, the paper is concluded with
the future works.
II. ABC AND LEVY FLIGHTS
A. Original Artificial Bee Colony
Artificial bee colony optimization algorithm was
developed in 2005 by being motivated from bee colonies by
Levy Flight Distribution for Scout Bee in Artificial Bee
Colony Algorithm
Hüseyin Hakli and Harun Uğuz
Lecture Notes on Software Engineering, Vol. 1, No. 3, August 2013
254DOI: 10.7763/LNSE.2013.V1.55
Karaboga [4]. ABC algorithm has three various bees:
employed, onlooker and scout. Colony is equal to both of
employed and onlooker bee, so half of the colony is
employed bees and another half is onlooker bees. In ABC
algorithm each food source stand for possible solution and
the amount of nectar of a food source represent the quality of
the possible solution. The number of food sources equals the
half of the colony size and equal to employed bees.
First, each employed bee randomly selects a food source
within the determined search space. This distribution is
performed according to the (1) below.
min max min
, (0, 1) ( )i j j j jx x rand x x (1)
where iX is D-dimensional vector, i=1, 2, 3…. SN (number
of food source), j=1, 2, 3, …, D, min
jx and max
jx are the
determined minimum and maximum limits.
After positions of employed bees are determined as such,
each employed bee selects different neighbor itself and
updates a randomly selected dimension according to the
following equation.
, , , , , ( )i j i j i j i j k jv x x x
(2)
where , i jv represents new value of jth dimension of ith particle,
, i jx represents jth dimension of ith particle, ,i j represents
the random value determined in the interval [-1,1], and
,k jx represents jth dimension of kth particle. If fitness value of
the new food source determined is better, the old food source
is forgotten by applying greedy search mechanism, and the
employed bee is updated with the new one. In the opposite
case, the old food source remains in the memory. After the
employed bees phase is completed, the probabilities are
determined for sharing quality of the solutions found with the
onlooker bees.
1
ii SN
jj
fitp
fit
(3)
where ifit denotes the fitness value of solution iX . The
fitness value ifit is defined as follows:
1
, ( ) 01 ( )
1 ( ) ( ) 0
i
ii
i i
if f Xf Xfit
f X if f X
(4)
Onlooker bees receive the information on quality of the
solutions (nectar amount of food source) from the employed
bees, and select the solution with better result. In Equation (3),
a pi probability is determined for each solution, the better the
quality of a solution, the higher its probability of being
selected. Then, the onlooker bees search for a new food
source using (2) around the food source they have selected,
and if the food source they find is better, they replace the old
food source with the newly found one.
Finally, if any improvement cannot be made as much as
the predetermined limit value number for a selected food
source, this food source is forgone, and the employed bee on
this source becomes scout bee. Scout bee randomly searches
for a food source within the determined search space.
B. Levy Flights
Levy flight [18] is a class of non-Gaussian random
processes whose random walks are drawn from Levy stable
distribution. This distribution is a simple power-law formula
L(s) ~ |s|-1-β where 0 < ß < 2 is an index. Mathematically
speaking, a simple version of Levy distribution can be
defined as [19], [20]:
3 2
1exp 0 ,
( , , ) 2 2( ) ( )
0 0
if sL s s s
if s
(5)
where γ>0 parameter is scale (controls the scale of
distribution) parameter, µ parameter is location or shift
parameter.
In general, Levy distribution should be defined in terms of
Fourier transform
,20],exp[)(
kkF (6)
where α is a parameter within [-1,1] interval and known as
skewness or scale factor. An index of o stability β є (0, 2] is
also referred to as Levy index.
In particular, for β = 1, the integral can be carried out
analytically and is known as the Cauchy probability
distribution. Another special case when β= 2, the distribution
correspond to Gaussian distribution.
β and α parameters take a major part in determination of
the distribution. The parameter β controls the shape of the
probability distribution in such a way that one can obtain
different shapes of probability distribution, especially in the
tail region depending on the parameter β. Thus, the smaller β
parameter causes the distribution to make longer jumps since
there will be longer tail [21]. It makes longer jumps for
smaller values whereas it makes shorter jumps for bigger
values.
III. THE PROPOSED ALGORITHM LFABC
The ABC algorithm gives successful results for most
benchmark functions. Although it is better in setting the
balance between exploration and exploitation compared to
other algorithms, it has several deficiencies. While it can
perform better the global search by performing random
searches around each food source and cover the search space
optimally, it encounters several problems in the exploitation
part, and gets stuck on local minimums particularly in
complex multimodal functions. If the original ABC
algorithm fails to make improvement for a particular food
source as much as a predetermined limit value number, then
the employed bee having that source become scout bee. Then
this scout bee is randomly distributed within the search space,
boundaries of which are determined. In this case, scout bees
Lecture Notes on Software Engineering, Vol. 1, No. 3, August 2013
255
select a zone within the search space completely randomly
without using the results found until that moment as a result
of the iterations. It is a very low probability for this newly
selected food source to be better than the GlobalMin (best
value in current iteration) found until that moment. In this
paper, in order to utilize the scout bees better, they are made
find a new food source using Levy Flight distribution with
also the effect of GlobalMin instead of selecting a random
food source. By this way, it is sought to attain a better
solution using the best food source until that moment instead
of the food source that could not be improved. By performing
enough random searches, exploration of the ABC algorithm,
which is already good in exploration, is improved.
Let us review in some more detail how distribution is
performed using Levy flight. By Levy flight, new state of the
particle is calculated as [19]:
)(1 LevyXX tt (7)
α is the step size which should be related to the scales of
the problem of interest. In the proposed method α is random
number for all dimensions of particles.
)())((1 LevyDsizerandomXX tt (8)
The product means entry-wise multiplications.
A non-trivial scheme of generating step size s samples are
discussed in detail in[19], [20], it can be summarized as
)(01.0~)())((1
tt
j gbestxv
uLevyDsizerandoms
(9)
where u and v are drawn from normal distributions. That is
2 2~ (0, ) ~ (0, )u vu N v N (10)
with
1,2]2/)1[(
)2/sin()1(1
2)1(
vu
(11)
Here is standard Gamma
function.
One of the important points to be considered while
performing distribution by Levy flights is the value taken by
the β parameter. As stated before, Yang and Deb mentioned
that the β parameter gave different results at different values
in the trials conducted in the study named Multiobjective
cuckoo search for design optimization [19]. Accordingly, it
can be concluded that a different β
parameter for each
benchmark function gave a more effective result. Moreover,
in the Evolutionary Algorithms with Adaptive Levy
Mutations study of Chang-Yong Lee, different constant
values were set of the β parameter, procedures were
conducted by calculating the distribution for each of these
values and selecting the best of the offspring produced [21].
As seen in these two examples, β parameter substantially
affects distribution.
IV. EXPERIMENTS AND RESULTS
A. Benchmark Functions and Parameter Settings
Experiments are implemented with the population size of
50(namely, SN=25) and the control parameter limit equals
100 for all the test functions. The algorithms are examined on
benchmark functions with dimension (D) of 50. For all
benchmark function, number of iteration is determined as
10000.The bees are initialized in the search range [Xmin,
Xmax], where Xmax and Xmin are the maximum and
minimum value in search range. The experimental results in
Table II are the average of minimum values of the benchmark
functions for 30 runs. The stopping criterion is set number of
maximum iteration. The performance of LFABC is compared
with the original ABC. For LFABC, β parameter is set 1.5 in
Table II.
The determined benchmark functions are numbered f1 to f10
given in Table I. All the functions are to be minimized. Table
1 shows characteristic(C), search range (Range), and
formulation of the functions. The functions f1-f2-f3-f7-f8 are
unimodal, f4-f5-f6-f9 are multimodal and f10 is shifted.
TABLE I: BENCHMARK FUNCTIONS(C: CHARACTERISTIC, U: UNIMODAL, M:
MULTIMODAL, S: SHIFTED)
Range C Function Formulation
[-100,
100] U Sphere
n
i
ixf1
2
1
[-100,
100] U Elliptic
n
i
i
ni xf1
2)1()1(6
2 )10(
[-10,
10] U
Rosenbro
ck
1
1
222
13 1100n
i
iii xxxf
[-5.12,
5.12] M Rastrigin
n
i
ii xxf1
2
4 10)2cos(10
[-600,
600] M
Griewan
k 1cos
4000
1
11
2
5
n
i
in
i
ii
xxf
[-32,
32] M Ackley
2
6
1
1
120exp 0.2
1exp cos(2 ) 20
D
i
i
D
i
i
f xD
x eD
[-10,
10] U
SumSqua
re
2
7
1
n
i
i
f ix
[-10,
10] U
Schwefel
2.22
n
i
n
i
ii xxf1 1
8
[-10,
10] M Alpine
n
i
iii xxxf1
9 1.0)sin(
[-100,
100] S
Shifted
Sphere oxzzf
n
i
i 1
2
10
B. Experimental Results
Table II shows the mean result of 30 runs with 10,000
iterations of ten benchmark functions (f1 to f10) with
dimension 50. The information about the mean optimum
solution, standard deviation, minimum and maximum value
and Wilcoxon test result by ABC and LFABC in 30 runs over
benchmark functions are given in Table I. The best mean
result and the best standard deviations of the benchmark
Lecture Notes on Software Engineering, Vol. 1, No. 3, August 2013
256
functions obtained by the algorithms are shown in bold.
According to Table II, the mean result of LFABC algorithm
performs better compared to ABC algorithm for the all
benchmark functions except f3. While ABC algorithm is
being trapped in local minimum, the LFABC algorithm
attained the optimum value for function f4 and f5. We can
conduct robustness comparisons of the algorithms by
checking the standard deviations. The fact inconsistency
between its results is high shows that it is not robust. LFABC
algorithm is a robust algorithm compared to ABC algorithm
for all benchmark functions except f3.
TABLE II: RESULTS FOR BENCHMARK FUNCTIONS (OPT: OPTIMUM STD:
STANDARD DEVIATION, MIN: MINIMUM, MAX: MAXIMUM, SIGN:
STATISTICAL TEST SIGN)
Func. Opt. ABC LFABC Sign
f1 0 Mean 1,44E-15 6,22E-16 +
STD 1,87E-16 7,4E-17
Min 8,91E-16 5,16E-16
Max 1,85E-15 7,18E-16
f2 0 Mean 1,44E-15 5,53E-16 +
STD 2,59E-16 5,83E-17
Min 9,75E-16 4,64E-16
Max 2,03E-15 6,93E-16
f3 0 Mean 0,023337 0,027578 -
STD 0,035322 0,076159
Min 0,000754 2,11E-05
Max 0,177254 0,420718
f4 0 Mean 5,24E-14 0 +
STD 1,09E-13 0
Min 0 0
Max 5,19E-13 0
f5 0 Mean 7,8E-15 0 +
STD 1,03E-14 0
Min 1,11E-16 0
Max 4,27E-14 0
f6 0 Mean 1,17E-13 3,74E-14 +
STD 1,66E-14 3,58E-15
Min 9,24E-14 3,2E-14
Max 1,63E-13 4,26E-14
f7 0 Mean 1,45E-15 6,04E-16 +
STD 1,47E-16 7,55E-17
Min 1,17E-15 4,93E-16
Max 1,65E-15 7,18E-16
f8 0 Mean 3,24E-15 1,58E-15 +
STD 3,5E-16 8,79E-17
Min 2,54E-15 1,41E-15
Max 3,79E-15 1,81E-15
f9 0 Mean 3,37E-08 7,3E-10 +
STD 1,01E-07 2,02E-09
Min 8,79E-12 1,65E-15
Max 4,7E-07 1,06E-08
f10 0 Mean 1,46E-15 5,97E-16 +
STD 2,21E-16 7,55E-17
Min 9,98E-16 4,95E-16
Max 1,88E-15 7,35E-16
When examined Table II in general, it cannot be concluded
that the proposed method makes much more improvement
compared to the ABC algorithm. But the proposed method
found the optimum values by not being stuck on local
minimums for f4 and f5 functions. It also improved benchmark
functions (except f3) compared to the ABC algorithm.
Non-parametric statistical test has been implemented for
statistically significant of between LFABC and the ABC. We
used Wilcoxon’s test with significance level 0.05, and results
are shown in Table II. In the charts, in case Wilcoxon test was
at 0.05 significance level, + mark was used, otherwise, –
mark was used; no mark was used in case all results showed
the sale value or optimum value. From the results in Table II,
it can be observed that the LFABC algorithm’s performance
is significantly better compared to ABC algorithm for all
benchmark functions (except f3) in Table I.
V. CONCLUSION AND FUTURE WORKS
In order to overcome insufficiency of the ABC algorithm
in exploitation, this paper proposed the LFABC method
ensuring the scout bees in the ABC algorithm find a new food
source within the search space with also the effect of
GlobalMin. The original ABC algorithm was tested in 10
benchmark functions by the proposed method. While the
proposed method gave the optimum value by ensuring the
ABC algorithm escape from local minimums for several
functions, it was seen to make improvement in the other
functions.
As future work, a better improvement can be attained by
using the LFABC method together with other proposed
algorithms, and it can be used on different optimization
problems together with these methods. Furthermore, effect of
the β value, an important parameter for Levy Flight
distribution, on other functions can be examined and detailed.
VI. EDITORIAL POLICY
The submitting author is responsible for obtaining
agreement of all coauthors and any consent required from
sponsors before submitting a paper. It is the obligation of the
authors to cite relevant prior work.
Authors of rejected papers may revise and resubmit them
to the journal again.
ACKNOWLEDGMENT
This study has been supported by Scientific Research
Project of Selcuk University.
REFERENCES
[1] J. Kennedy and R. Eberhart, “Particle Swarm Optimization,” in Proc.
Sixth International Symposium on Micro Machine and Human Science,
Nagoya, Japan, 1995, pp. 39-43.
[2] M. Dorigo and G. D. Caro, “Ant colony optimization: a new
meta-heuristic,” in Proc. the 1999 Congress on Evolutionary
Computation, Washington, DC, 1999, pp. 1470-1477.
[3] K. S. Tang, K. F. Man, S. Kwong, and Q. He, “Genetic algorithms and
their applications,” IEEE Signal Processing Magazine, vol. 13, pp.
22-37, 1996.
[4] D. Karaboga, “An Idea Based On Honey Bee Swarm For Numerical
Optimization,” Technical Report-TR06, Erciyes University,
Engineering Faculty, Computer Engineering Department, 2005.
[5] D. Karaboga and B. Akay, “A comparative study of Artificial Bee
Colony algorithm,” Applied Mathematics and Computation, vol. 214,
pp. 108-132, 2009.
Lecture Notes on Software Engineering, Vol. 1, No. 3, August 2013
257
[6] W. Xiang and M. An, “An efficient and robust artificial bee colony
algorithm for numerical optimization,” Computers & Operations
Research, In Press, 2012.
[7] F. Kang, J. Li, and H. Li, “Artificial bee colony algorithm and pattern
search hybridized for global optimization,” Applied Soft Computing, In
Press, 2013.
[8] W. Y. Szeto, Y. Wu, and S. C. Ho, “An artificial bee colony algorithm
for the capacitated vehicle routing problem,” European Journal of
Operational Research, vol. 215, pp. 126-135, 2011.
[9] X. Yan, Y. Zhu, W. Zou, and L. Wang, “A new approach for data
clustering using hybrid artificial bee colony algorithm,”
Neurocomputing, vol. 97, pp. 241-250, 2012.
[10] J. Yu and H. Duan, “Artificial Bee Colony approach to information
granulation-based fuzzy radial basis function neural networks for
image fusion,” Optik, In Press, 2012.
[11] M.-H. Horng, “Multilevel thresholding selection based on the artificial
bee colony algorithm for image segmentation,” Expert Systems with
Applications, vol. 38, pp. 13785-13791, 2011.
[12] W.-C. Hong, “Electric load forecasting by seasonal recurrent SVR
(support vector regression) with chaotic artificial bee colony
algorithm,” Energy, vol. 36, pp. 5568-5578, 2011.
[13] W.-D. Chang, “Nonlinear CSTR control system design using an
artificial bee colony algorithm,” Simulation Modelling Practice and
Theory, vol. 31, pp. 1-9, 2013.
[14] G. Li, P. Niu, and X. Xiao, “Development and investigation of efficient
artificial bee colony algorithm for numerical function optimization,”
Applied Soft Computing, vol. 12, pp. 320-332, 2012.
[15] G. Zhu and S. Kwong, “Gbest-guided artificial bee colony algorithm
for numerical function optimization,” Applied Mathematics and
Computation, vol. 217, pp. 3166-3173, 2010.
[16] W. Gao, S. Liu, and L. Huang, “A global best artificial bee colony
algorithm for global optimization,” Journal of Computational and
Applied Mathematics, vol. 236, pp. 2741-2753, 2012.
[17] B. Alatas, “Chaotic bee colony algorithms for global numerical
optimization,” Expert Systems with Applications, vol. 37, pp.
5682-5687, 2010.
[18] A. V. Chechkin, R. Metzler, J. Klafter, and V. Y. Gonchar,
“Introduction To The Theory Of Lévy Flights,” in R. Klages, G.
Radons, and I. M. Sokolov (Eds.), Anomalous Transport: Foundations
and Applications, John Wiley & Sons, pp. 129-162, 2008.
[20] X.-S. Yang, Engineering Optimization An Introduction with
Metaheuristic Applications, first ed., New Jersey: John Wiley and Sons,
2010.
[21] C.-Y. Lee and X. Yao, “Evolutionary Algorithms with Adaptive Levy
Mutations,” in Proc. the 2001 Congress on Evolutionary Computation,
Seoul, South Korea, 2001, pp. 568-575.
H. Hakli was born in Turkey on September 3, 1989. He
graduated valedictorian from the Department of
Computer Engineering at Selcuk University in 2011. He
is receiving his MS degree (thesis phase) from the same
department and same university. He has been a research
assistant in Department of Computer Engineering since
2011. His interest areas include optimization,
nature-inspired algorithms, machine learning and
artificial intelligence systems.
H. Uğuz was born in Turkey in 1977. He graduated from
the Department of Computer Engineering at Selcuk
University in 1999. He received his MS degree from the
same department and same university in 2002. His PhD
degree was received from the Department of Electrical
and Electronics Engineering at Selcuk University in
2007. He has been an associate professor in the
Department of Computer Engineering since 2007. His
interest areas include hidden Markov models, machine learning and artificial
intelligence systems.
Lecture Notes on Software Engineering, Vol. 1, No. 3, August 2013
258
[19] X.-S.Yang and S. Deb, “Multiobjective cuckoo search for design
optimization,” Computers & Operations Research.