Expert Systems with Applications 42 (2015) 6686–6698

Contents lists available at ScienceDirect

Expert Systems with Applications

journal homepage: www.elsevier.com/locate/eswa

TSA: Tree-seed algorithm for continuous optimization

Mustafa Servet Kiran ⇑
Department of Computer Engineering, Faculty of Engineering, University of Selcuk, 42075 Konya, Turkey

a r t i c l e i n f o a b s t r a c t

Article history: This paper presents a new intelligent optimizer based on the relation between trees and their seeds for
Available online 4 May 2015 continuous optimization. The new method is in the field of heuristic and population-based search. The
location of trees and seeds on n-dimensional search space corresponds with the possible solution of an
Keywords: optimization problem. One or more seeds are produced from the trees and the better seed locations
Heuristic search are replaced with the locations of trees. While the new locations for seeds are produced, either the best
Tree and seed solution or another tree location is considered with the tree location. This consideration is performed by
Numeric optimization
using a control parameter named as search tendency (ST), and this process is executed for a pre-defined
Multilevel thresholding
number of iterations. These mechanisms provide to balance exploitation and exploration capabilities of
the proposed approach. In the experimental studies, the effects of control parameters on the performance
of the method are firstly examined on 5 well-known basic numeric functions. The performance of the
proposed method is also investigated on the 24 benchmark functions with 2, 3, 4, 5 dimensions and mul-
tilevel thresholding problems. The obtained results are also compared with the results of state-of-art
methods such as artificial bee colony (ABC) algorithm, particle swarm optimization (PSO), harmony
search (HS) algorithm, firefly algorithm (FA) and the bat algorithm (BA). Experimental results show that
the proposed method named as TSA is better than the state-of-art methods in most cases on numeric
function optimization and is an alternative optimization method for solving multilevel thresholding
problem.
Ó 2015 Elsevier Ltd. All rights reserved.

1. Introduction non-differentiable optimization problems). Over the last decades,

a growing popularity of population-based intelligent search meth-
Optimization is a task performed to find the best solution ods has been occurred due to their successes in solving the opti-
amongst the possible solutions for an optimization problem with mization problems, which are easy adaptation and implem
regard to some criteria. Optimization deals with optimization entation for the optimization problem and low computation costs.
problems, and can contain maximization or minimization pro- These techniques use biological or physical phenomena in local
cesses. If S is a search space, F; F # S, is a set of acceptable solutions interactions amongst the agents while solving the optimization
of S and f is an objective function, minimization or maximization is problem. Genetic algorithm (GA) (Goldberg, 1989;Holland, 1992)
to find ~ x 2 F in Eqs. (1) and (2), those define minimization and has been developed by considering Darwinian evolutionary theory.
maximization processes, respectively. In GA, crossover and mutation mechanisms are used to produce a
new generation. Ant colony optimization (ACO) (Dorigo, Maniezzo,
f ð~
xÞ 6 f ð~
yÞ 8~
y2F ð1Þ & Colorni, 1996) mimics the behavior of real ants between the nest
and food source. Basically, ACO is proposed to solve discrete opti-
f ð~ yÞ 8~
xÞ P f ð~ y2F ð2Þ
mization problems and the main concept is based on pheromone
If the function to be optimized is non-continuous or mechanism to construct the solutions. Particle swarm optimization
non-differentiable, or optimizing the function is computationally (PSO) (Kennedy & Eberhart, 1995) has been inspired by cognitive
expensive for high dimensional search space, heuristic search and social behaviors of fish or birds. The solution update rule in
methods can be applied to optimize it. PSO uses subtraction operation between the current solution and
The expert or intelligent systems aim to present a solution the best (the best of population and the best solution obtained so
methodology to solve this type of problems (nonlinear or far by the particle) solutions to produce a new possible solution.
Harmony search (HS) (Geem, Kim, & Loganathan, 2001) algorithm
⇑ Tel.: +90 332 223 1992; fax: +90 332 223 2106. simulates the improvisation process of musicians. Bat algorithm
E-mail addresses: [email protected], [email protected] (BA) (Yang, 2010b) imitates the echolocation behavior of bats.

http://dx.doi.org/10.1016/j.eswa.2015.04.055
0957-4174/Ó 2015 Elsevier Ltd. All rights reserved.
 M.S. Kiran / Expert Systems with Applications 42 (2015) 6686–6698 6687

Firefly algorithm (FA) (Yang, 2010a) has been inspired by the flash- problem. To obtain a location of a seed that will be produced from
ing behavior of fireflies. Artificial bee colony algorithm (ABC) a tree is important for the optimization problem because this pro-
(Karaboga & Basturk, 2007) simulates foraging and information cess constitutes the core of search. We propose two search equa-
sharing behaviors of bees in a hive. In ABC, the same tions for this process. The first equation (Eq. (1)) considers the
subtraction-based update rule is used in both employed bee phase tree location that the seed will be produced for this tree and the
and onlooker bee phase of the algorithm, and the best solution best location of the tree population. This search equation also
obtained so far is not used in this algorithm. As well as organisms’ improves the local search or intensification capability of the pro-
behaviors, some methods modeled physical laws for solving opti- posed algorithm. The second update rule (Eq. (2)) uses two differ-
mization problems. Simulated Annealing (SA) (Kirkpatrick, Gelatt, ent tree locations for producing a new seed for the tree.
& Vecchi, 1983) has been inspired by thermodynamic effects, and
Si;j ¼ T i;j þ ai;j  ðBj  T r;j Þ ð3Þ
gravitational search algorithm (GSA) (Rashedi, Nezamabadi-Pour,
& Saryazdi, 2009) is based on the Newtonian laws of gravity. In Si;j ¼ T i;j þ ai;j  ðT i;j  T r;j Þ ð4Þ
GSA, all the solutions are used to produce a new solution by using where, Si;j is jth dimension of ith seed that will be produced ith tree,
the Newtonian laws of gravity.
T i;j is the jth dimension of ith tree, Bj is the jth dimension of best tree
In the present study, it aims to establish a new optimizer named
location obtained so far, T r;j is the jth dimension of rth tree ran-
as tree–seed algorithm (TSA) to solve continuous optimization
domly selected from the population, a is the scaling factor ran-
problems. In the new proposed method, two aspects, exploration
domly produced in range of [1, 1] and i and r are different indices.
and exploitation, are considered to overcome the characteristics
The most important point is which equation will be selected to
of the optimization problems. For exploration, the method is
produce a new seed location. This selection is controlled by a con-
designed as population-based, the search starts from
trol parameter of the method named as search tendency (ST) in
multiple-points on the search space of the problem, and random-
range of [0, 1]. The higher value of ST provides a powerful local
ness is included in local interactions amongst the agents. These
search and speed convergence, the lower value of ST causes slow
agents in the proposed method are called as trees. In order to com-
convergence but powerful global search. In other words, the explo-
pensate exploitation, seeds are used. The number of seeds produced
ration and exploitation capabilities of the TSA are controlled by ST
from a tree is randomly obtained. The locations of trees and seeds
parameter.
on n-dimensional search space correspond to the possible solu-
In the beginning of search with TSA, the initial tree locations
tions of the optimization problem, and while a new seed is pro-
which are possible solutions for the optimization problem are pro-
duced from a tree, either another tree location or the best tree
duced by using Eq. (5).
(the best solution obtained so far) location is considered with its
own location. The search tendency to the best or randomly selected T i;j ¼ Lj;min þ r i;j ðHj;max  Lj;min Þ ð5Þ
tree is handled with a control parameter named as search tendency
where, Lj;min is the lower bound of the search space, Hj;max is the
(ST) and this control in the TSA provides a local intensification and
higher bound of the search space and r i;j is a random number pro-
convergence to the optimum or near optimum for the problem.
duced for each dimension and location, in range of [0, 1].
The paper is organized as follows: the study is introduced in
For minimization, the best solution is selected from the popula-
Section 1, and the main contribution of the study is presented in
tion using Eq. (6).
this section. TSA method is detailed in Section 2, experimental
studies and comparisons are conducted in Section 3, the obtained B ¼ minff ð~
T i Þgi ¼ 1; 2; . . . ; N ð6Þ
results are reported and discussed in Section 4 and finally, the
study is concluded and future directions are given in Section 5. where, N is the number of trees in the population.
While the new seed locations are generated for a tree, the num-
ber of seeds can be more than one and this number depends on the
1.1. Main contribution population size. In the analysis of effects of control parameters to
the performance of TSA, 10% of population size is the minimum
In this study, a novel population-based iterative search algo- number of seeds produced for a tree and 25% of the population size
rithm is developed for solving continuous optimization problems. is the maximum number of seeds produced from a tree. The num-
In the present study, two mechanisms are merged to balance ber of seed production is completely random in TSA.
exploration and exploitation capabilities of the method. First The detailed algorithmic framework of the TSA is given in Fig. 1.
mechanism aims to improve exploration capability of the method In the algorithm, it is seen how to use the control parameter ST in
by using search tendency parameter. The new solutions are pro- TSA. If randomly produced number (rand in Fig. 1) in range of [0, 1]
duced by considering the current solution and the best or ran- is less than ST, Eq. (3) is used for updating the dimension, other-
domly selected solution with this parameter. The second wise Eq. (4) is used.
mechanism focuses on improving the exploitation of the method.
In this mechanism, more than one solution is created around a
3. Experimental studies
solution. Therefore, the search around the found solutions is
improved by using the second search mechanism. Based on the lit-
3.1. Performance analysis and comparisons on the benchmark
erature review of the study, these mechanisms show the novelty
functions
and differences from the approaches exist in the literature.

In the experiments, the effects of control parameters to the per-

2. TSA: Tree–seed algorithm formance of TSA are analyzed and then the performance of TSA
with the selected control parameters is compared with the results
The natural phenomena in TSA is the relationship between trees of PSO, ABC, HS, FA, BA methods. In the analyses, TSA is applied to
and their seeds. In nature, trees spread to the surface through their solve five basic benchmark functions (F1, F10, F11, F13, F15 in the
seeds. These seeds grow over time and new trees become from Table 1) to obtain the best control parameters. The methods are
these seeds. If we assume that the surface of these trees as a search run on the 2,3,4,5 dimensional 24 numeric functions given in
space for the optimization problem, the location of trees and seeds Table 1 to compare with the other methods. For all experiments,
can be considered as possible solutions for the optimization the termination condition is selected as the maximum number of
6688 M.S. Kiran / Expert Systems with Applications 42 (2015) 6686–6698

Fig. 1. The algorithmic framework of the proposed algorithm.

function evaluations (Max_FEs) and it is set by using the dimen- and ST parameter is selected as 0.1, 0.2, 0.3, 0.4 and 0.5. In these
sionality of the function given in Eq. (7) as given in (Suganthan premises, TSA is run 30 times for solving 2, 3, 4 and 5 dimensional
et al., 2005) functions, mean and standard deviation of the results are reported
in Tables 2–21.
Max FEs ¼ D  10; 000 ð7Þ The analysis results show that a small size of the population is
The numeric benchmark functions used in the experiments enough to obtain well-quality and robust results using TSA because
have some characteristics. If a function has more than one local the search in TSA is performed with the seeds. Higher values of ST
optimum, this function is called as multimodal (M) and these types parameter have caused the stagnation behavior for TSA because
of functions test the global search ability of the algorithm. more dimensions of seed location are affected by the best solution
Unimodal functions (U) have only one local optimum, and it is glo- in population obtained so far. Therefore, the small values for pop-
bal optimum. The exploitation ability of the algorithm is examined ulation size and ST parameter are required for obtaining good qual-
on these types of problems. If a function with n-variable can be ity results with TSA. In the analysis, it is also shown that when the
written as the sum of the n functions of one variable, then this dimension of the optimization problem is increased, the perfor-
function is called as a separable (S) function. Non-separable func- mance of the method is decreased. This is because the dimension-
tions cannot be written in this form because there is interrelation ality has caused the growth of search space exponentially. TSA is
amongst variables of these functions. Some functions have flatness more successful in solving unimodal numeric functions than the
on the surface and solving such functions is difficult for algorithms multimodal numeric functions because the intensification capabil-
because these functions do not give any direction information to ity of TSA (producing more than one seeds for a tree and effecting
the methods. If the global optimum of the function is in the narrow the best solution) is more effective than the global search capabil-
curving valley Rosenbrock’s Banana function, the methods should ity. But results obtained with TSA for all numeric functions are in
keep up with the direction changes in the functions(Kiran & an acceptable and comparable level.
Gunduz, 2013). The dimensionality of the function is another issue
for the method because the search space of the problem increases
3.1.2. Comparison of TSA with the state-of-art methods on benchmark
exponentially. Therefore, these benchmark functions are selected
functions
to investigate the performance of the proposed method.
According to parameter analysis of TSA, ST is set to 0.1 and the
population size is taken as 10 in the comparative studies. In these
3.1.1. The analysis of control parameters of TSA premises, TSA is applied to solve the rest of the functions and
The control parameters of TSA are accepted as population size obtained results are compared with the results of BA, FA, HS, ABC
and ST. The population size is selected as 10, 20, 30, 40 and 50, and PSO algorithms. The specific control parameters of BA are
 M.S. Kiran / Expert Systems with Applications 42 (2015) 6686–6698 6689

Table 1
Benchmark Functions used in Experiments.

No of Name Search C Function

Func. Range
F1 Sphere [100,100]D US Pn
f 1 ð~
XÞ ¼ 2
i¼1 xi
F2 Elliptic [100,100]D UN P ði1Þ=ðn1Þ
XÞ ¼ ni¼1 ð106 Þ
f 2 ð~ x2i
F3 SumSquares [10,10]D US ~ Pn 2
f 3 ðXÞ ¼ i¼1 ixi
F4 SumPower [10,10]D MS P
XÞ ¼ ni¼1 jxi jðiþ1Þ
f 4 ð~
F5 Schwefel2.22 [10,10]D UN P Q
f ð~
5 XÞ ¼ n jxi j þ n jxi j
i¼1 i¼1
F6 Schwefel2.21 [100,100]D UN f 6 ð~
XÞ ¼ maxi fjxi j; 1 6 i 6 ng
F7 Step [100,100] D
US P
XÞ ¼ ni¼1 ðbxi þ 0:5cÞ2
f 7 ð~
F8 Quartic [1.28,1.28]D US P
f 8 ðXÞ ¼ ni¼1 ix4i
~
F9 QuarticWN [1.28,1.28]D US P
XÞ ¼ ni¼1 ix4i þ random½0; 1Þ
f 9 ð~
F10 Rosenbrock [10,10]D UN P h  
2 2
i
f 10 ð~XÞ ¼ n1 i¼1 100 xiþ1  xi þ ðxi  1Þ2
P  
F11 Rastrigin [5.12,5.12]D MS f 11 ð~XÞ ¼ ni¼1 x2i  10 cosð2pxi Þ þ 10
F12 Non-Continuous [5.12,5.12]D MS Pn  2 
f 12 ð~
XÞ
( ¼ i¼1 yi  10 cosð2 ) pyi Þ þ 10
Rastrigin xi 1
jxi j < 2
yi ¼ roundð2xi Þ
2 jxi j P 12
P Q  
F13 Griewank [600,600]D MN f 13 ð~XÞ ¼ 4000 i¼1 x2i  ni¼1 cos pxiffii þ 1
1 n

D P pffiffiffiffiffiffiffi
F14 Schwefel2.26 [500,500] UN f 14 ð~XÞ ¼ 418:98  n  ni¼1 xi sin jxi j
n qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P ffio P
F15 Ackley [32,32]D MN f 15 ð~XÞ ¼ 20 exp 0:2 1n i¼1 x2i  exp 1n ni¼1 cosð2pxi Þ þ 20 þ e
n

F16 Penalized1 [50,50] D

MN P Pn
XÞ ¼ pn f10 sin2 ðpy1 Þ þ 8
f 16 ð~ n1 2 2 2
i¼1 ðyi  1Þ ½1 þ 10 sin ðpyiþ1 Þ þ ðyn  1Þ g þ i¼1 uðxi ; 10; 100; 4Þ
m
< kðxi  aÞ xi > a
yi ¼ 1 þ 14 ðxi þ 1Þ uxi ;a;k;m ¼ 0 a 6 xi 6 a
:
kðxi  aÞm xi < a
n P h i h io P
F17 Penalized2 [50,50]D MN f ð~ XÞ ¼ 1 sin2 ðpx1 Þ þ n1 ðxi  1Þ2 1 þ sin2 ð3pxiþ1 Þ þ ðxn  1Þ2 1 þ sin2 ð2pxiþ1 Þ þ n uðxi ; 5; 100; 4Þ
17 10 i¼1 i¼1

F18 Alpine [10,10]D MS P

XÞ ¼ ni¼1 jxi  sinðxi Þ þ 0:1  xi j
f 18 ð~
F19 Levy [10,10]D MN P 2 2 2
XÞ ¼ n1
f 19 ð~ 2
i¼1 ðxi  1Þ ½1 þ sin ð3pxiþ1 Þ þ sin ð3px1 Þ þ jxn  1j½1 þ sin ð3pxn Þ
F20 Weierstrass [0.5,0.5]D MN ~ PD Pkmax h k k
i Pkmax h k k
i
f 20 ðXÞ ¼ i¼1 k¼0 a cosð2pb ðxi þ 0:5ÞÞ  D k¼0 a cosð2pb 0:5Þ a ¼ 0:5; b ¼ 3; kmax ¼ 20
D  pffiffiffiffiffiffiffiffiffiffiffiffiffi
Pn 2 
F21 Schaffer [100,100] MN sin2 x 0:5
f 21 ð~
XÞ ¼ 0:5 þ  P i¼1 i
n
2
1þ0:001 x2
i¼1 i

F22 Himmelblau [5,5]D MS P

XÞ ¼ 1n ni¼1 ðx4i  16x2i þ 5xi Þ
f 22 ð~
P  2
F23 Michalewicz [0,p]D MS f 23 ðXÞ ¼  ni¼1 sinðxi Þ sin20 pi
~ ix

F24 Dixon&Price [10,10] D

UN P 2
XÞ ¼ ðx1  1Þ2 þ ni¼2 ið2x2i  xi1 Þ
f 23 ð~

Table 2
The analysis results of 2-dimensional functions for ST = 0.1.

Func. Pop = 10 Pop = 20 Pop = 30 Pop = 40 Pop = 50

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 3.99E232 0.00E+00 2.87E102 1.57E101 1.56E65 4.19E65 1.70E43 4.50E43 2.42E32 4.14E32
F10 3.67E16 1.42E15 2.12E09 5.36E09 2.42E06 9.37E06 9.63E06 1.84E05 1.10E04 4.61E04
F11 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00
F13 1.23E04 2.80E03 3.28E04 1.39E03 1.93E04 7.51E04 3.22E04 1.08E03 5.98E05 1.61E04
F15 0.00E+00 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00

Table 3
The analysis results of 3-dimensional functions for ST = 0.1.

Func. Pop = 10 Pop = 20 Pop = 30 Pop = 40 Pop = 50

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 1.01E238 0.00E+00 8.01E110 3.62E109 1.17E66 5.11E66 1.96E44 2.72E44 9.65E33 1.34E32
F10 3.30E03 1.05E02 9.74E03 1.37E02 2.03E02 4.92E02 2.25E02 3.58E02 1.75E02 2.62E02
F11 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 3.55E16 1.95E15
F13 3.06E03 3.88E03 2.96E03 3.99E03 3.97E03 4.07E03 6.11E03 4.32E03 7.88E03 4.54E03
F15 0.0E+00 0.0E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00

loudness (A) and rate of pulse emission (r). They are set to 1.5 and FA control parameters, respectively (Yang, 2011). There are five
0.5, respectively. In addition, a and b constants for BA are set to 0.9 control parameters named as harmony memory size (HMS), har-
according to (Yang & Gandomi, 2012) and the population size for mony memory consideration rate (HSMCR), pitch adjusting rate
BA is taken as 20 for the experiments based on (Yang, 2012). The (PAR) and arbitrary distance bandwidth (BW) in HS algorithm.
minimum value for attractiveness ðbÞ, absorption coefficient ðcÞ, These control parameters are set to 5, 0.9, 0.3, 0.01 according to
randomization ðaÞ and population size are set to 0.2, 0.5, 1, 20 as (Wang & Yan, 2013), respectively. ABC algorithm has only two
6690 M.S. Kiran / Expert Systems with Applications 42 (2015) 6686–6698

Table 4
The analysis results of 4-dimensional functions for ST = 0.1.

Func. Pop = 10 Pop = 20 Pop = 30 Pop = 40 Pop = 50

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 4.64E240 0.00E+00 2.08E111 9.27E111 2.32E68 5.78E68 1.76E44 7.62E44 2.19E33 2.68E33
F10 1.70E01 3.67E01 2.20E01 4.14E01 1.30E01 2.07E01 1.10E01 1.30E01 1.14E01 1.13E01
F11 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 8.29E15 4.41E14 2.30E09 1.05E08
F13 7.95E03 6.69E03 1.21E02 8.15E03 2.19E02 1.15E02 2.22E02 1.06E02 2.80E02 1.29E02
F15 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 1.01E15 6.49E16

Table 5
The analysis results of 5-dimensional functions for ST = 0.1.

Func. Pop = 10 Pop = 20 Pop = 30 Pop = 40 Pop = 50

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 7.64E243 0.00E+00 1.17E113 4.16E113 2.41E69 5.94E69 9.01E46 1.35E45 7.35E34 7.79E34
F10 3.15E01 1.05E+00 4.50E01 6.62E01 3.35E01 2.99E01 4.04E01 3.00E01 4.49E01 3.09E01
F11 0.00E+00 0.00E+00 0.00E+00 0.00E+00 8.05E15 4.41E14 2.85E07 1.52E06 1.29E03 4.81E03
F13 2.02E02 1.57E02 2.53E02 1.53E02 4.02E02 2.23E02 4.66E02 1.74E02 5.88E02 2.20E02
F15 7.11E16 1.45E15 1.01E15 6.49E16 8.88E16 0.00E+00 1.01E15 6.49E16 1.24E15 1.08E15

Table 6
The analysis results of 2-dimensional functions for ST = 0.2.

Func. Pop = 10 Pop = 20 Pop = 30 Pop = 40 Pop = 50

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 7.12E233 0.00E+00 1.18E100 6.35E100 6.22E62 1.76E61 3.98E41 6.79E41 8.48E30 4.25E29
F10 1.81E12 5.25E12 6.82E07 2.49E06 6.81E06 2.96E05 3.54E05 4.30E05 6.53E05 1.09E04
F11 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00
F13 5.08E04 1.80E03 7.39E05 4.05E04 1.54E05 7.49E05 2.50E04 1.35E03 9.95E05 3.22E04
F15 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 1.60E15 1.45E15

Table 7
The analysis results of 3-dimensional functions for ST = 0.2.

Func. Pop = 10 Pop = 20 Pop = 30 Pop = 40 Pop = 50

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 3.40E242 0.00E+00 1.83E105 8.30E105 6.05E64 2.08E63 1.67E42 2.39E42 2.47E31 3.46E31
F10 4.13E02 1.62E01 1.52E02 3.48E02 1.47E02 1.78E02 2.06E02 2.78E02 2.35E02 3.29E02
F11 6.63E02 2.52E01 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 1.67E14 8.81E14
F13 2.60E03 3.99E03 2.33E03 3.40E03 4.12E03 4.18E03 5.44E03 4.51E03 9.68E03 4.72E03
F15 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 1.48E15 1.35E15

Table 8
The analysis results of 4-dimensional functions for ST = 0.2.

Func. Pop = 10 Pop = 20 Pop = 30 Pop = 40 Pop = 50

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 1.27E247 0.00E+00 1.05E108 3.48E108 8.65E66 2.44E65 1.71E43 3.25E43 8.10E32 1.05E31
F10 3.54E01 1.05E+00 7.86E02 1.03E01 1.08E01 1.10E01 1.09E01 1.15E01 1.46E01 1.30E01
F11 6.63E02 2.52E01 0.00E+00 0.00E+00 0.00E+00 0.00E+00 1.09E14 5.58E14 1.37E09 6.82E09
F13 7.49E03 7.13E03 1.05E02 8.51E03 2.05E02 1.09E02 2.65E02 1.18E02 3.09E02 1.20E02
F15 1.13E15 9.01E16 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 1.01E15 6.49E16

Table 9
The analysis results of 5-dimensional functions for ST = 0.2.

Func. Pop = 10 Pop = 20 Pop = 30 Pop = 40 Pop = 50

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 1.79E248 0.00E+00 1.33E110 2.44E110 4.71E67 9.43E67 2.50E44 5.87E44 1.53E32 1.59E32
F10 4.45E01 9.48E01 3.63E01 4.24E01 2.51E01 2.19E01 2.72E01 2.72E01 4.75E01 3.05E01
F11 1.33E01 3.44E01 0.00E+00 0.00E+00 7.11E16 3.89E15 2.01E03 9.71E03 5.11E02 1.91E01
F13 1.90E02 1.25E02 2.54E02 1.57E02 3.93E02 2.04E02 6.08E02 2.72E02 5.81E02 1.87E02
F15 1.24E15 1.08E15 8.88E16 0.00E+00 1.01E15 6.49E16 1.01E15 6.49E16 2.43E15 1.79E15
 M.S. Kiran / Expert Systems with Applications 42 (2015) 6686–6698 6691

Table 10
The analysis results of 2-dimensional functions for ST = 0.3.

Func. Pop = 10 Pop = 20 Pop = 30 Pop = 40 Pop = 50

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 1.68E215 0.00E+00 4.59E95 1.49E94 2.72E58 9.01E58 3.06E39 6.96E39 3.38E29 8.00E29
F10 8.19E11 4.41E10 1.27E06 3.85E06 1.61E05 2.70E05 1.25E04 3.54E04 2.79E04 9.83E04
F11 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00
F13 9.87E04 2.56E03 1.07E04 5.55E04 1.54E05 6.13E05 2.61E04 1.22E03 1.31E04 4.59E04
F15 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 6.34E15 3.82E15

Table 11
The analysis results of 3-dimensional functions for ST = 0.3.

Func. Pop = 10 Pop = 20 Pop = 30 Pop = 40 Pop = 50

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 1.28E234 0.00E+00 1.04E100 3.56E100 2.40E61 3.65E61 4.29E40 1.01E39 9.70E30 1.95E29
F10 4.52E01 1.08E+00 2.11E02 4.73E02 2.69E02 5.03E02 2.78E02 3.66E02 3.90E02 7.96E02
F11 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 3.19E13 1.68E12
F13 4.14E03 4.09E03 2.66E03 3.33E03 5.77E03 3.89E03 6.23E03 5.48E03 8.47E03 5.03E03
F15 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 4.20E15 1.85E15

Table 12
The analysis results of 4-dimensional functions for ST = 0.3.

Func. Pop = 10 Pop = 20 Pop = 30 Pop = 40 Pop = 50

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 1.17E239 0.00E+00 8.28E105 3.73E104 3.19E63 4.41E63 1.65E41 2.45E41 2.92E30 4.46E30
F10 2.90E01 3.95E01 1.36E01 1.66E01 1.02E01 1.14E01 1.29E01 1.14E01 1.33E01 1.36E01
F11 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 5.76E13 3.15E12 3.35E06 1.82E05
F13 7.33E03 5.93E03 1.13E02 7.72E03 2.23E02 9.21E03 2.51E02 1.36E02 3.26E02 1.30E02
F15 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 3.61E15 1.53E15

Table 13
The analysis results of 5-dimensional functions for ST = 0.3.

Func. Pop = 10 Pop = 20 Pop = 30 Pop = 40 Pop = 50

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 8.10E243 0.00E+00 2.14E106 7.42E106 2.27E64 4.57E64 3.72E42 6.06E42 5.54E31 5.81E31
F10 3.63E01 7.51E01 2.66E01 2.05E01 2.95E01 2.31E01 3.66E01 2.72E01 3.75E01 3.33E01
F11 3.65E01 8.05E01 0.00E+00 0.00E+00 1.76E12 7.78E12 1.16E03 6.33E03 2.06E02 1.01E01
F13 1.74E02 9.87E03 2.88E02 1.98E02 3.84E02 1.83E02 6.25E02 2.45E02 6.43E02 2.14E02
F15 1.36E15 1.23E15 1.01E15 6.49E16 1.01E15 6.49E16 1.13E15 9.01E16 3.02E15 1.77E15

Table 14
The analysis results of 2-dimensional functions for ST = 0.4.

Func. Pop = 10 Pop = 20 Pop = 30 Pop = 40 Pop = 50

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 4.15E208 0.00E+00 1.42E89 5.03E89 6.61E55 1.54E54 1.33E36 2.81E36 2.65E27 5.67E27
F10 1.77E09 8.54E09 1.30E06 3.08E06 1.79E05 2.42E05 4.03E05 3.96E05 1.25E04 2.58E04
F11 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00
F13 9.94E06 5.10E05 9.38E05 3.94E04 2.54E05 7.31E05 1.81E04 4.41E04 3.82E04 6.21E04
F15 1.01E15 6.49E16 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 3.24E14 2.48E14

Table 15
The analysis results of 3-dimensional functions for ST = 0.4.

Func. Pop = 10 Pop = 20 Pop = 30 Pop = 40 Pop = 50

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 4.71E220 0.00E+00 2.24E95 1.01E94 1.36E57 3.04E57 2.76E38 3.15E38 6.54E28 7.03E28
F10 2.41E02 5.81E02 1.62E02 2.28E02 2.51E02 3.06E02 2.31E02 1.89E02 2.12E02 3.68E02
F11 1.33E01 7.27E01 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 8.39E11 3.67E10
F13 2.84E03 4.08E03 2.54E03 3.22E03 5.57E03 4.28E03 7.71E03 5.92E03 9.36E03 5.51E03
F15 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 1.75E14 1.10E14
6692 M.S. Kiran / Expert Systems with Applications 42 (2015) 6686–6698

Table 16
The analysis results of 4-dimensional functions for ST = 0.4.

Func. Pop = 10 Pop = 20 Pop = 30 Pop = 40 Pop = 50

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 1.50E230 0.00E+00 5.58E99 1.80E98 1.06E59 1.54E59 3.12E39 6.41E39 1.69E28 3.49E28
F10 5.75E01 1.11E+00 1.67E01 1.95E01 1.30E01 1.11E01 1.24E01 9.27E02 1.75E01 1.48E01
F11 3.32E02 1.82E01 0.00E+00 0.00E+00 0.00E+00 0.00E+00 7.47E11 3.72E10 2.49E04 1.18E03
F13 6.43E03 5.90E03 1.09E02 6.17E03 2.03E02 7.64E03 2.66E02 1.43E02 3.11E02 1.17E02
F15 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 9.30E15 3.90E15

Table 17
The analysis results of 5-dimensional functions for ST = 0.4.

Func. Pop = 10 Pop = 20 Pop = 30 Pop = 40 Pop = 50

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 5.52E235 0.00E+00 2.86E101 1.34E100 3.34E61 6.24E61 3.45E40 6.04E40 3.27E29 4.25E29
F10 5.52E01 1.09E+00 2.30E01 1.94E01 3.47E01 2.77E01 3.40E01 2.53E01 3.69E01 3.21E01
F11 9.95E02 3.04E01 0.00E+00 0.00E+00 8.72E13 4.77E12 4.59E04 2.46E03 2.67E02 7.49E02
F13 1.61E02 9.46E03 3.00E02 1.50E02 4.57E02 1.82E02 5.48E02 2.14E02 8.08E02 2.27E02
F15 1.60E15 1.45E15 8.88E16 0.00E+00 8.88E16 0.00E+00 1.13E15 9.01E16 5.15E15 1.96E15

Table 18
The analysis results of 2-dimensional functions for ST = 0.5.

Func. Pop = 10 Pop = 20 Pop = 30 Pop = 40 Pop = 50

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 4.76E187 0.00E+00 6.65E83 1.81E82 1.15E50 3.36E50 2.01E34 4.30E34 1.58E25 2.39E25
F10 7.53E10 1.82E09 2.03E06 3.73E06 7.84E05 1.77E04 1.83E04 5.00E04 3.54E04 5.11E04
F11 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 1.78E15 6.90E15
F13 5.15E04 1.87E03 1.04E04 3.90E04 2.11E04 9.06E04 2.88E04 1.35E03 3.45E04 7.45E04
F15 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 2.47E13 2.29E13

Table 19
The analysis results of 3-dimensional functions for ST = 0.5.

Func. Pop = 10 Pop = 20 Pop = 30 Pop = 40 Pop = 50

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 1.25E208 0.00E+00 9.66E90 3.12E89 1.16E53 2.79E53 1.28E35 1.34E35 2.28E26 2.13E26
F10 1.68E02 4.20E02 3.22E02 5.77E02 4.25E02 7.89E02 3.99E02 4.69E02 3.86E02 4.98E02
F11 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 1.78E15 7.15E15 1.24E08 6.03E08
F13 2.20E03 3.63E03 3.83E03 3.66E03 6.04E03 4.30E03 7.08E03 4.81E03 1.12E02 6.35E03
F15 1.01E15 6.49E16 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 1.86E13 1.51E13

Table 20
The analysis results of 4-dimensional functions for ST = 0.5.

Func. Pop = 10 Pop = 20 Pop = 30 Pop = 40 Pop = 50

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 7.85E218 0.00E+00 1.48E93 3.57E93 9.36E56 1.95E55 5.95E37 6.24E37 1.02E26 9.94E27
F10 3.33E01 7.14E01 1.19E01 1.69E01 1.37E01 1.57E01 1.60E01 1.33E01 1.14E01 1.04E01
F11 6.63E02 2.52E01 0.00E+00 0.00E+00 0.00E+00 0.00E+00 4.16E06 2.28E05 1.27E04 5.06E04
F13 6.18E03 5.74E03 1.31E02 8.99E03 2.16E02 1.16E02 3.07E02 1.50E02 3.52E02 1.39E02
F15 1.13E15 9.01E16 8.88E16 0.00E+00 8.88E16 0.00E+00 8.88E16 0.00E+00 6.38E14 3.50E14

Table 21
The analysis results of 5-dimensional functions for ST = 0.5.

Func. Pop = 10 Pop = 20 Pop = 30 Pop = 40 Pop = 50

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 8.96E226 0.00E+00 1.31E96 2.61E96 7.26E58 1.81E57 1.22E37 2.31E37 1.76E27 1.47E27
F10 3.54E01 7.77E01 2.66E01 2.61E01 3.39E01 2.97E01 3.11E01 2.05E01 4.22E01 3.12E01
F11 9.95E02 3.04E01 0.00E+00 0.00E+00 1.10E11 4.17E11 1.79E02 6.57E02 1.10E01 1.71E01
F13 1.23E02 8.09E03 3.23E02 1.35E02 4.91E02 1.95E02 6.07E02 2.55E02 7.97E02 2.70E02
F15 5.49E02 3.01E01 1.13E15 9.01E16 8.88E16 0.00E+00 1.84E15 1.60E15 2.35E14 1.00E14
 M.S. Kiran / Expert Systems with Applications 42 (2015) 6686–6698 6693

Table 22
The comparison results of BA, FA, HS, ABC, PSO and TSA on the 2-dimensional benchmark functions.

FNO BA FA HS ABC PSO TSA

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 4.63E+01 1.32E+02 9.42E08 7.30E08 3.85E03 1.13E02 1.64E18 1.64E18 1.79E05 6.86E05 3.99E232 0.00E+00
F2 6.47E+04 1.98E+05 1.68E+02 4.82E+02 3.76E+03 1.99E+04 1.76E18 2.03E18 5.21E+03 4.12E+03 2.08E228 0.00E+00
F3 4.29E10 4.55E10 1.43E09 1.62E09 1.53E09 1.87E09 1.27E18 1.23E18 4.06E07 1.50E06 1.67E230 0.00E+00
F4 3.43E02 1.88E01 1.97E11 2.31E11 7.82E10 1.26E09 6.33E19 6.53E19 1.41E07 7.72E07 5.45E270 0.00E+00
F5 1.75E01 5.18E01 3.47E05 2.04E05 4.35E05 2.68E05 1.78E17 8.74E18 4.55E03 1.85E02 4.30E125 2.35E124
F6 7.17E+00 6.80E+00 2.80E04 1.19E04 2.07E01 2.73E01 3.64E17 2.04E17 2.54E02 6.88E02 3.46E99 1.47E98
F7 1.76E+02 2.05E+02 6.67E02 2.54E01 4.33E01 6.26E01 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00
F8 1.37E19 2.26E19 8.99E22 1.69E21 4.28E18 1.41E17 3.06E20 7.32E20 1.50E17 7.66E17 0.00E+00 0.00E+00
F9 4.38E02 6.80E02 5.29E02 4.46E02 4.10E03 5.35E03 4.30E04 2.95E04 2.10E03 3.45E03 2.81E04 2.14E04
F10 2.07E+00 5.93E+00 1.24E08 1.42E08 1.92E+00 2.89E+00 8.62E03 1.19E02 1.91E+00 2.26E+00 3.67E16 1.42E15
F11 3.32E+00 2.97E+00 5.77E08 5.48E08 3.32E02 1.82E01 0.00E+00 0.00E+00 6.19E02 2.12E01 0.00E+00 0.00E+00
F12 3.83E+00 3.47E+00 6.89E08 6.12E08 6.18E07 1.44E06 0.00E+00 0.00E+00 6.47E02 2.09E01 0.00E+00 0.00E+00
F13 1.85E+00 2.11E+00 2.67E03 3.95E03 4.41E02 3.23E02 0.00E+00 0.00E+00 1.41E02 8.50E03 1.23E03 2.80E03
F14 6.55E+04 0.00E+00 5.53E+01 6.77E+01 6.61E01 1.49E+00 0.00E+00 0.00E+00 1.38E+02 1.10E+02 3.95E+00 2.16E+01
F15 6.98E+00 3.66E+00 2.61E04 1.18E04 1.30E04 9.29E05 0.00E+00 0.00E+00 4.32E02 1.59E01 0.00E+00 0.00E+00
F16 7.88E+00 2.02E+01 3.40E08 3.42E08 5.18E02 2.84E01 1.20E18 9.84E19 1.23E03 6.61E03 2.36E31 4.45E47
F17 2.51E+02 9.44E+02 3.05E08 2.88E08 1.34E02 1.79E02 1.72E18 1.73E18 7.03E06 2.71E05 1.35E32 5.57E48
F18 4.04E03 1.49E02 4.00E06 2.02E06 5.71E05 1.23E04 2.13E17 9.57E18 8.63E03 1.90E02 2.04E17 1.11E16
F19 2.54E+00 2.73E+00 5.11E06 3.09E06 2.20E02 4.47E02 1.03E19 1.48E19 7.91E04 1.75E03 1.35E31 6.68E47
F20 4.77E02 1.10E01 1.75E03 5.82E04 1.02E02 4.14E03 0.00E+00 0.00E+00 5.17E03 1.03E02 0.00E+00 0.00E+00
F21 2.55E01 1.35E01 7.14E03 3.08E03 3.03E02 2.09E02 1.63E05 4.34E05 6.71E03 3.81E03 1.15E03 2.93E03
F22 7.41E+01 6.59E+00 7.83E+01 7.76E09 7.83E+01 2.34E08 7.83E+01 1.45E14 7.83E+01 2.33E01 7.83E+01 1.45E14
F23 1.92E+00 6.06E02 1.80E+00 6.14E10 1.80E+00 2.29E08 1.80E+00 9.03E16 1.16E+00 2.64E01 1.80E+00 9.03E16
F24 1.28E3 6.84E2 4.41E09 5.24E09 4.66E07 6.61E07 8.13E16 1.83E15 6.11E3 3.06E2 3.74E32 2.25E33

control parameters named as population size and limit specific to 5-dimensional functions. The results of the method that produce
the method. The number of bees in ABC hive is set to 40 and limit the best result are marked with bold font type.
is set to 100 according to Matlab Code of ABC available in
(Karaboga, 2008). The PSO control parameters are set according
to PSO tutorials available in (Hu, 2006). The number of particles 3.2. Performance assessment of TSA on multilevel thresholding
in PSO is set to 30, c1 (coefficient of cognitive component) and
c2 (coefficient of social component) control parameters are set to 2. Image thresholding is one of the popular and simple methods
Using these control parameters, each method is run 30 times used for image segmentation (Pal & Pal, 1993). Because image
with random initialization to solve for each 2, 3, 4, and 5 dimen- thresholding uses only the brightness of the image for image seg-
sional functions. The obtained results are reported in Table 22 for mentation, this thresholding process is fast and it is preferred for
2-dimensional functions, Table 23 for 3-dimensional functions, real-world applications. The performance of image thresholding
Table 24 for 4-dimensional functions, and Table 25 for directly affects the success of the other image processing stages.
Therefore, it is important that image thresholding should be

Table 23
The comparison results of BA, FA, HS, ABC, PSO and TSA on the 3-dimensional benchmark functions.

FNO BA FA HS ABC PSO TSA

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 3.68E+02 4.90E+02 1.43E06 8.26E07 1.05E03 5.73E03 6.27E18 4.59E18 6.36E04 1.69E03 1.01E238 0.00E+00
F2 5.23E+06 8.52E+06 1.23E+03 1.46E+03 7.46E01 3.07E+00 5.21E18 3.37E18 2.03E+06 3.56E+06 2.96E233 0.00E+00
F3 3.46E03 1.90E02 2.12E08 1.31E08 3.33E09 3.58E09 5.57E18 3.09E18 4.52E02 2.12E01 4.93E239 0.00E+00
F4 1.39E02 7.59E02 6.58E11 9.98E11 1.57E10 1.97E10 1.17E18 1.23E18 1.07E01 2.67E01 0.00E+00 0.00E+00
F5 1.87E+00 3.13E+00 1.42E04 5.07E05 5.44E05 3.41E05 3.48E17 1.22E17 9.87E02 1.91E01 3.57E135 1.90E134
F6 1.46E+01 1.17E+01 8.50E04 2.81E04 1.66E01 2.30E01 1.63E16 1.02E16 3.52E02 6.64E02 1.48E86 7.96E86
F7 6.18E+02 5.76E+02 6.67E02 2.54E01 1.67E01 4.61E01 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00
F8 2.42E17 3.04E17 5.90E20 7.00E20 3.21E17 1.13E16 1.92E19 1.84E19 5.16E05 2.80E04 0.00E+00 0.00E+00
F9 7.55E02 7.56E02 4.60E02 4.95E02 3.03E03 3.49E03 9.74E04 6.49E04 5.87E03 5.04E03 2.70E04 1.81E04
F10 1.35E+01 4.09E+01 4.45E06 7.69E06 1.38E+00 1.73E+00 2.35E02 2.57E02 6.21E+00 4.65E+00 3.30E03 1.05E02
F11 8.52E+00 7.14E+00 3.32E01 4.77E01 5.02E07 7.40E07 0.00E+00 0.00E+00 2.03E+00 1.31E+00 0.00E+00 0.00E+00
F12 9.38E+00 7.08E+00 4.00E01 4.98E01 2.14E06 6.26E06 0.00E+00 0.00E+00 2.33E+00 4.32E+00 6.67E02 2.54E01
F13 5.75E+00 5.99E+00 1.29E02 8.00E03 6.80E02 6.09E02 1.19E12 6.49E12 1.07E01 5.32E02 3.06E03 3.88E03
F14 6.55E+04 0.00E+00 8.29E+01 7.06E+01 3.03E01 6.49E01 7.58E15 4.15E14 3.39E+02 1.16E+02 3.95E+00 2.16E+01
F15 1.25E+01 3.50E+00 8.05E04 2.40E04 1.22E04 6.60E05 3.55E16 1.08E15 1.78E+00 2.59E+00 0.00E+00 0.00E+00
F16 5.09E+04 2.79E+05 1.04E07 8.66E08 3.46E02 1.89E01 4.89E18 3.64E18 2.79E02 1.10E01 1.57E31 6.68E47
F17 2.33E+05 1.25E+06 1.88E07 1.14E07 6.16E03 9.76E03 5.08E18 3.84E18 1.63E02 4.77E02 1.35E32 5.57E48
F18 1.89E01 3.82E01 1.42E05 6.07E06 6.16E05 6.56E05 4.81E17 1.79E17 3.40E01 4.65E01 6.74E12 3.69E11
F19 5.03E+00 3.28E+00 8.23E06 5.07E06 1.10E02 3.35E02 1.94E18 1.73E18 3.23E01 1.64E+00 1.35E31 6.68E47
F20 5.57E01 3.76E01 4.53E03 1.02E03 1.67E02 6.33E03 0.00E+00 0.00E+00 9.41E02 1.08E01 0.00E+00 0.00E+00
F21 3.53E01 1.27E01 9.68E03 2.03E04 2.99E02 2.32E02 4.20E03 4.63E03 1.34E02 1.33E02 9.50E03 5.79E03
F22 6.99E+01 6.24E+00 7.83E+01 2.74E08 7.83E+01 1.91E08 7.83E+01 1.45E14 7.65E+01 3.42E+00 7.83E+01 1.45E14
F23 2.58E+00 2.32E01 2.74E+00 4.97E02 2.76E+00 8.74E08 2.76E+00 6.85E16 1.48E+00 3.23E01 2.76E+00 5.09E16
F24 4.32E+00 1.22E+2 7.89E08 4.65E08 2.09E06 6.33E06 3.1E14 5.69E14 5.1E+00 1.44E+2 3.21E26 1.76E25
6694 M.S. Kiran / Expert Systems with Applications 42 (2015) 6686–6698

Table 24
The comparison results of BA, FA, HS, ABC, PSO and TSA on the 4-dimensional benchmark functions.

FNO BA FA HS ABC PSO TSA

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 8.95E+02 9.55E+02 5.03E06 3.17E06 1.19E08 2.03E08 1.32E17 7.80E18 2.22E+00 5.94E+00 4.64E240 0.00E+00
F2 1.83E+07 3.42E+07 2.08E+03 2.20E+03 9.21E02 2.07E01 1.36E17 8.12E18 4.78E+06 1.68E+07 4.49E236 0.00E+00
F3 5.91E08 4.90E08 8.18E08 3.34E08 1.04E08 1.15E08 1.42E17 8.04E18 1.52E+00 2.24E+00 3.78E241 0.00E+00
F4 1.18E+00 3.43E+00 1.16E10 2.35E10 3.03E10 8.17E10 2.31E18 1.81E18 1.58E+01 2.69E+01 0.00E+00 0.00E+00
F5 2.36E+00 3.27E+00 3.18E04 9.52E05 8.35E05 3.85E05 5.17E17 1.85E17 2.36E+00 3.29E+00 1.93E142 1.02E141
F6 2.40E+01 9.69E+00 1.44E03 3.73E04 2.40E01 2.70E01 4.84E16 3.44E16 3.87E+00 6.69E+00 8.29E75 3.09E74
F7 2.00E+03 1.56E+03 6.67E02 2.54E01 3.00E01 5.96E01 0.00E+00 0.00E+00 1.07E+00 2.35E+00 0.00E+00 0.00E+00
F8 4.08E16 4.86E16 6.30E19 5.42E19 1.13E16 3.01E16 8.51E19 8.21E19 4.98E03 1.53E02 0.00E+00 0.00E+00
F9 1.93E01 1.62E01 4.16E02 4.21E02 3.88E03 4.27E03 1.64E03 1.04E03 2.64E02 3.25E02 3.85E04 2.25E04
F10 3.82E+01 8.15E+01 9.03E03 7.05E03 1.19E+00 1.48E+00 2.61E02 2.81E02 7.07E+02 2.53E+03 1.70E01 3.67E01
F11 1.59E+01 8.14E+00 1.13E+00 1.10E+00 8.40E07 6.98E07 0.00E+00 0.00E+00 1.01E+01 6.78E+00 0.00E+00 0.00E+00
F12 1.70E+01 8.71E+00 1.03E+00 6.69E01 6.22E07 5.83E07 0.00E+00 0.00E+00 9.42E+00 9.03E+00 3.67E01 4.90E01
F13 1.52E+01 1.13E+01 2.79E02 2.09E02 7.99E02 5.07E02 2.79E10 1.39E09 3.74E01 2.24E01 7.95E03 6.79E03
F14 6.55E+04 0.00E+00 1.84E+02 1.15E+02 3.84E01 6.88E01 7.58E15 4.15E14 5.85E+02 1.79E+02 2.37E+01 4.82E+01
F15 1.42E+01 2.58E+00 1.25E03 2.85E04 2.11E04 2.06E04 2.13E15 1.77E15 1.05E+01 5.40E+00 8.88E16 0.00E+00
F16 7.09E+04 2.98E+05 1.68E07 1.03E07 7.25E09 1.13E08 1.25E17 6.52E18 7.58E01 1.67E+00 1.18E31 2.23E47
F17 2.12E+05 4.06E+05 3.44E07 1.74E07 6.85E03 1.04E02 1.19E17 6.79E18 4.32E+00 2.28E+01 1.35E32 5.57E48
F18 3.14E01 6.00E01 3.51E05 7.42E06 1.32E04 1.22E04 8.16E17 4.32E17 1.46E+00 7.95E01 4.08E17 1.55E16
F19 9.14E+00 7.04E+00 1.03E05 6.26E06 1.47E02 3.80E02 7.01E18 4.94E18 3.51E+00 4.90E+00 1.35E31 6.68E47
F20 1.36E+00 5.44E01 8.97E03 1.71E03 2.15E02 7.33E03 0.00E+00 0.00E+00 8.25E01 7.14E01 0.00E+00 0.00E+00
F21 4.17E01 1.07E01 9.72E03 2.77E07 4.79E02 2.85E02 9.09E03 2.38E03 1.19E01 1.35E01 5.91E03 5.65E10
F22 7.03E+01 6.36E+00 7.83E+01 4.29E08 7.83E+01 2.37E08 7.83E+01 1.71E14 7.00E+01 6.78E+00 7.83E+01 2.19E14
F23 3.14E+00 4.01E01 3.65E+00 5.11E02 3.70E+00 1.93E06 3.70E+00 1.49E15 1.72E+00 3.36E01 3.70E+00 1.70E02
F24 5.04E+00 1.58E+01 4.1E07 2.41E07 1.56E1 2.85E1 2.3E13 5.86E13 8.14E+1 1.07E+2 1.59E26 8.7E26

Table 25
The comparison results of BA, FA, HS, ABC, PSO and TSA on the 5-dimensional benchmark functions.

FNO BA FA HS ABC PSO TSA

Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
F1 1.41E+03 1.11E+03 1.07E05 4.91E06 2.39E08 2.81E08 2.25E17 8.00E18 2.98E+01 5.14E+01 7.64E243 0.00E+00
F2 3.10E+07 4.40E+07 6.32E+03 8.83E+03 2.13E01 8.02E01 1.82E17 9.31E18 1.17E+07 1.92E+07 3.85E240 0.00E+00
F3 8.81E03 3.34E02 2.24E07 1.10E07 3.74E08 5.12E08 2.19E17 8.59E18 3.39E+01 3.94E+01 5.44E239 0.00E+00
F4 1.29E+00 4.15E+00 2.54E10 2.38E10 1.99E10 3.23E10 3.20E18 2.15E18 2.45E+02 2.35E+02 0.00E+00 0.00E+00
F5 7.03E+00 7.38E+00 5.29E04 1.13E04 1.56E04 8.49E05 7.76E17 2.06E17 8.65E+00 6.67E+00 3.49E147 1.75E146
F6 2.80E+01 9.88E+00 2.12E03 4.44E04 3.79E01 3.35E01 2.08E14 4.56E14 1.69E+01 1.37E+01 9.47E63 5.15E62
F7 2.49E+03 1.23E+03 0.00E+00 0.00E+00 2.00E01 4.84E01 0.00E+00 0.00E+00 2.60E+02 1.07E+03 3.33E02 1.83E01
F8 3.48E15 2.43E15 1.91E18 1.68E18 7.49E17 9.64E17 2.29E18 1.90E18 8.29E02 1.91E01 0.00E+00 0.00E+00
F9 3.45E01 2.82E01 4.80E02 3.64E02 5.77E03 5.29E03 2.39E03 1.22E03 2.94E01 2.64E01 5.16E04 3.05E04
F10 7.95E+00 2.42E+01 2.15E01 8.80E01 1.52E+00 1.68E+00 3.42E02 3.57E02 1.84E+03 3.81E+03 3.15E01 1.05E+00
F11 2.07E+01 9.29E+00 2.35E+00 1.39E+00 1.69E06 1.36E06 0.00E+00 0.00E+00 2.47E+01 9.34E+00 0.00E+00 0.00E+00
F12 2.65E+01 1.17E+01 2.40E+00 7.24E01 3.33E02 1.83E01 0.00E+00 0.00E+00 2.30E+01 1.09E+01 4.67E01 5.71E01
F13 2.64E+01 1.57E+01 3.64E02 2.62E02 7.62E02 4.84E02 1.23E03 2.80E03 3.22E+00 1.06E+01 2.02E02 1.57E02
F14 6.55E+04 0.00E+00 2.44E+02 1.38E+02 1.74E01 3.62E01 0.00E+00 0.00E+00 9.02E+02 1.96E+02 1.18E+01 3.61E+01
F15 1.53E+01 2.78E+00 1.60E03 4.02E04 8.83E03 4.72E02 3.20E15 1.08E15 1.58E+01 3.30E+00 7.11E16 1.45E15
F16 1.12E+05 2.82E+05 2.57E07 1.12E07 1.72E08 2.25E08 2.07E17 8.82E18 1.48E+03 7.90E+03 9.42E32 3.34E47
F17 1.92E+06 4.35E+06 5.83E07 2.76E07 6.03E03 1.04E02 1.94E17 7.85E18 1.59E+05 5.83E+05 3.66E04 2.01E03
F18 5.57E01 7.05E01 5.46E05 2.63E05 1.48E04 1.20E04 1.27E16 7.56E17 3.42E+00 1.52E+00 1.30E09 7.14E09
F19 1.58E+01 1.22E+01 1.49E05 7.31E06 1.47E02 3.80E02 1.14E17 4.66E18 7.68E+00 4.57E+00 7.32E03 2.79E02
F20 2.26E+00 5.13E01 1.29E02 1.81E03 3.29E02 8.45E03 0.00E+00 0.00E+00 2.68E+00 1.38E+00 0.00E+00 0.00E+00
F21 4.60E01 4.59E02 9.72E03 4.45E11 6.97E02 3.63E02 9.44E03 1.07E03 2.99E01 1.57E01 7.16E03 9.38E10
F22 6.95E+01 5.08E+00 7.80E+01 1.43E+00 7.83E+01 4.71E08 7.83E+01 1.58E14 6.24E+01 7.77E+00 7.83E+01 1.77E14
F23 3.39E+00 4.98E01 4.57E+00 7.56E02 4.69E+00 1.11E06 4.69E+00 1.54E15 1.76E+00 2.71E01 4.69E+00 6.70E02
F24 2.07E+01 8.89E+01 8.66E07 3.66E07 1.56E01 2.87E01 3.31E13 6.46E13 1.17E+03 1.57E+03 1.82E26 8.99E26

performed to reveal as much as possible information in the images 3.2.1. Problem Formulation
(Pekdemir, 2012). Bi-level image thresholding focuses on dividing the image into
There are many non-parametric methods and their applications two parts as background and foreground and can be used for
for image thresholding are presented in the literature (Akay, 2013; retrieval of information and image segmentation. Let L be gray
Bhandari, Kumar, & Singh, 2015; Bhandari, Singh, Kumar, & Singh, levels in a given image, t threshold value between 0 and L  1, I a
2014; Horng, 2011; Kapur, Sahoo, & Wong, 1985; Otsu, 1979; Pal & given image, bi-level thresholding can be defined as:
Pal, 1993; Pun, 1980; Sezgin & Tasaltin, 2000; Yin & Chen, 1997). A
well-known method is Otsu’s method based on histogram data of PF ¼ fMðx; yÞ 2 Ij0 6 Mðx; yÞ 6 t  1g
ð8Þ
image. Therefore, we used maximization of between class variance PB ¼ fMðx; yÞ 2 Ijt 6 Mðx; yÞ 6 L  1g
proposed by Otsu (1979) for obtaining thresholds given an image.
 M.S. Kiran / Expert Systems with Applications 42 (2015) 6686–6698 6695

Bi-level thresholding can be extended to multilevel threshold-

ing by increasing number of parts for the image and threshold
values.
P0 ¼ fMðx; yÞ 2 Ij0 6 Mðx; yÞ 6 t 0  1g
P1 ¼ fMðx; yÞ 2 Ijt 0 6 Mðx; yÞ 6 t 1  1g
ð9Þ
...
Pn ¼ fMðx; yÞ 2 Ijt n1 6 Mðx; yÞ 6 L  1g
For bi-level thresholding, t value can be easily obtained but
multilevel thresholding can require more computational efforts.
Therefore the heuristic search algorithm is more appropriate for
finding the optimal t values in multilevel thresholding of an image.
For bilevel thresholding, Otsu (1979) defined and verified
between-class variance for obtaining t threshold value. When we
maximize the sum of the sigma functions calculated for each class,
we can obtain optimal t value for bi-level thresholding. The math-
ematical definition of objective function is given as follows:
t  ¼ arg max½f b ðtÞ ð10Þ
f b ðtÞ ¼ r0 þ r1 ð11Þ
r0 ¼ x0 ðl0  lT Þ2 and r1 ¼ x1 ðl1  lT Þ2 ð12Þ
1 X
t1
1 X
L1
l0 ¼ i  pi and l1 ¼ i  pi ð13Þ
x0 i¼0
x1 i¼t
X
t1 X
L1
x0 ¼ pi and x1 ¼ pi ð14Þ
i¼0 i¼t

pi ¼ xi =X ð15Þ
where, xi is the number of pixels at level i, X is the total number of
pixels at each level and pi is the normal value of ith gray levels in Eq.
(15). x0 and x1 are the estimated probability of 0th and 1st classes
occurrence in Eq. (14), l0 is the mean intensity of 0th class, l1 is the
mean intensity of 1st class Eq. (13) and lT is the mean intensity of
the original image in Eq. (12). r0 is the variance of 0th class and r1
is the variance of 1st class in Eq. (11).
According to basic formulation of between-class variance, we
extend between-class variance-based image thresholding to multi-
level thresholding given as follows:
t  ¼ arg max½f m ðtÞ ð16Þ
X
n Fig. 2. The images used for testing the performance of ABC and TSA on multilevel
f b ðtÞ ¼ ri ð17Þ thresholding.
i¼1

r0 ¼ x0 ðl0  lT Þ2 ; r1 ¼ x1 ðl1  lT Þ2 ; . . . ; rn ¼ xn ðln  lT Þ2

applied to maximize between-class variance (Eq. (17)) in multi-
ð18Þ level thresholding for images given in Fig. 2. Lena and
1 X 1 X 1 X
t 0 1 t 1 1 L1 Cameraman images are frequently used in image processing appli-
l0 ¼ i  pi ; l1 ¼ i  pi ; . . . ; ln ¼ i  pi ð19Þ cations as test images, and other pictures (Horses, Ostrich, Soldier,
x0 i¼0
x1 i¼t0 xn i¼tn
Starfish, Wherry, Zebra) are taken from (Martin, Fowlkes, Tal, &
tX
0 1 tX
1 1 X
L1
Malik, 2001). For TSA, the location of seeds or tree represents
x0 ¼ pi ; x1 ¼ pi ; . . . ; xn ¼ pi ð20Þ
threshold value in the range of [0, 256] on 2, 3, 4 or 5 dimensional
i¼0 i¼t 0 i¼t n
search space. The range is between 0 and 255 because images with
8 bit gray level are used in the experiments.
3.2.2. Application of TSA and ABC to the problem For all images, we can try to obtain optimal 2, 3, 4 and 5 thresh-
Finding optimal threshold values in multilevel thresholding is a old values by using ABC and TSA. For the adaptation of the methods
crucial task for image segmentation applications. If we divide an to the problem, before the objective function is evaluated by a seed
image into 5 parts with different gray level, we need 4 threshold of TSA or an agent of ABC, the possible solution (threshold values)
values. When an image has 256 gray levels, Eq. (17) should be cal- is sorted and rounded. The TSA and ABC methods are run 10 times
culated 172061505 times for obtaining the optimal t values for each image and for each number of threshold value, and
(Pekdemir, 2012). Therefore, this process can require more compu- obtained the best results given in Tables 26–29 are compared with
tational efforts and this exhaustive search cannot be suitable for each other. Control parameters for the methods are set according
real-time applications. to experiment 1 (solving the benchmark functions).
To overcome the high computation cost, heuristic search meth- According to comparison tables, TSA and ABC show similar per-
ods can be applied to solve the multilevel thresholding problem. formance in solving the multilevel threshold problem but the run-
Due to the fact that we obtain more quality results by using TSA ning time of TSA is slightly better than the ABC algorithm as given
and ABC on the benchmark functions, TSA and ABC methods are in Fig. 3.
6696 M.S. Kiran / Expert Systems with Applications 42 (2015) 6686–6698

Table 26
The results of maximizing between-class variance for 2 threshold values by ABC and TSA.

Image ABC TSA

BCV Elapsed time Thresholds BCV Elapsed time Thresholds
Cameraman 3929.462 1.76 65,142 3929.462 1.53 65,142
Horses 2458.408 2.03 91,169 2458.408 1.78 91,168
Lena 2634.77 1.78 77,145 2634.77 1.57 77,145
Ostrich 1086.159 2.17 75,135 1086.159 1.79 75,135
Soldier 1515.494 2.13 105,167 1515.494 2.03 105,167
Starfish 2590.543 2.47 83,156 2590.543 1.81 83,156
Wherry 3482.043 2.28 105,188 3482.043 1.97 105,188

Table 27
The results of maximizing between-class variance for 3 threshold values by ABC and TSA.

Image ABC TSA

BCV Elapsed time Thresholds BCV Elapsed time Thresholds
Cameraman 4012.297 3.02 55,119,156 4012.297 2.72 55,119,156
Horses 2624.742 3.66 83,140,187 2624.742 3.27 83,140,187
Lena 2859.198 2.96 65,119,170 2859.198 2.65 65,119,170
Ostrich 1151.347 3.79 69,101,149 1151.347 3.39 69,101,149
Soldier 1667.793 3.88 92,133,184 1667.793 3.30 92,133,184
Starfish 2831.242 3.91 67,118,177 2831.242 3.84 67,118,177
Wherry 3721.888 3.96 98,158,217 3721.888 3.66 98,158,217

Table 28
The results of maximizing between-class variance for 4 threshold values by ABC and TSA.

Image ABC TSA

BCV Elapsed time Thresholds BCV Elapsed time Thresholds
Cameraman 4069.157 4.78 37,92,138,169 4069.157 4.39 37,92,138,169
Horses 2711.834 5.37 68,114,150,194 2711.834 4.83 68,114,150,194
Lena 2943.803 4.38 57,101,138,179 2943.803 4.00 57,101,138,179
Ostrich 1192.025 5.60 65,93,128,178 1192.025 5.14 65,93,128,178
Soldier 1741.774 5.46 82,117,150,197 1741.774 5.26 82,117,150,197
Starfish 2917.5 5.78 59,100,137,187 2917.5 5.30 59,100,137,187
Wherry 3780.383 5.82 78,117,161,217 3780.383 5.45 78,117,161,217

Table 29
The results of maximizing between-class variance for 5 threshold values by ABC and TSA.

Image ABC TSA

BCV Elapsed time Thresholds BCV Elapsed time Thresholds
Cameraman 4103.198 7.33 30,80,122,150,174 4103.198 6.49 30,80,122,150,174
Horses 2756.931 7.65 59,102,132,162,201 2756.931 6.99 59,102,132,162,201
Lena 2979.169 6.06 56,97,129,157,189 2979.169 5.60 56,97,129,157,189
Ostrich 1217.057 7.75 54,76,99,130,179 1217.057 7.05 54,76,99,130,179
Soldier 1784.428 7.96 74,106,132,163,207 1784.428 7.09 74,106,132,163,207
Starfish 2964.623 8.15 50,84,115,148,192 2964.623 7.31 50,84,115,148,192
Wherry 3818.731 8.09 75,113,150,183,224 3818.731 7.11 75,113,150,183,224

4. Results and discussion For the rest of 2-dimensional functions, TSA is better than the other
methods.
Based on the comparison tables, it is shown that the proposed For the results of 3-dimensional functions in Table 23, ABC algo-
method is better than the other methods in most cases for solving rithm is better than the other methods on F12, F13, F14 and F21
the numeric benchmark functions. functions, and FA is better than the other methods in solving F10
In solving 2-dimensional functions in Table 22, ABC algorithm is function. FA, HS, ABC and TSA are better than the other methods
better than the other methods on F13, F14 and F21 functions, BA is on F22 function and HS, ABC and TSA are better than the other
better than the other methods on F23 function, and all the algo- methods on F23 function. ABC, PSO and TSA have equal perfor-
rithms have equal performance on F22 function, except BA. For mance in solving F7 function, and ABC and TSA show equal perfor-
F7 functions ABC, PSO and TSA are better than the other methods mance on solving F11 and F20 functions. In solving the rest of
but they have equal performance. TSA and ABC show equal perfor- 3-dimensional functions, TSA shows better performance than the
mance in solving 2-dimensional F11, F12, F15 and F20 functions. other methods. Based on Table 24, FA is better than the other
 M.S. Kiran / Expert Systems with Applications 42 (2015) 6686–6698 6697

003 005
004
003
004
002 003
Thr=2 003 Thr=3
002
002
001 ABC 002 ABC
001
001 TSA TSA
001
000 000

007 009
006 008
007
005 006
004 Thr=4 005 Thr=5
003 004
ABC 003 ABC
002
002
001 TSA TSA
001
000 000

Fig. 3. The time comparison of ABC and TSA on the multilevel thresholding problem for different number of thresholding.

methods in solving F10 function, and ABC is better than the other of trees with seeds. In the algorithm, the location of trees and seeds
methods on F12, F13, F14 and F15 functions. ABC and TSA are bet- n-dimensional space represents the possible solution for the opti-
ter than the other methods on solving F7, F11 and F20 functions mization problem. TSA is an iterative algorithm, one or more seeds
but have equal performance. For F22 function, FA, HS, ABC and are produced for each tree, the best seed location is selected, and
TSA have equal performance, and HS, ABC and TSA have equal per- this seed substitutes for the parent tree. This process is carried
formance on F23 function. TSA is better than the other methods in out until a termination condition is met. The maximum number
solving F1, F2, F3, F4, F5, F6, F8, F9, F16, F17, F18, F19, F21 and F23. of function evaluations is used for terminating TSA method in this
Generally, TSA has higher or equal performance on 19 study. Based on the results tables, the performance of the proposed
4-dimensional benchmark functions, but lower performance on 5 method is better than the other intelligent optimization methods
4-dimensional functions. At the same time, it is shown that TSA in terms of solution quality and robustness in most cases. In solv-
has higher or equal performance on 16 5-dimensional functions ing multilevel thresholding, TSA has lower computation time than
but slightly lower performance on 8 5-dimensional functions from ABC algorithm according to comparison table.
Table 25. As seen from the results, the proposed approach is better than
Based on the image thresholding experiment, TSA method is the other methods in most cases because the search tendency of
successful in solving the multilevel thresholding problem as well TSA is controlled and one or more solutions can be obtained
as ABC algorithm, and has lower computation time than the ABC around a solution to improve intensification of the population.
according to Fig. 3. In addition, this application shows that the The decision or design parameters of the problem are updated by
TSA can be used in real-time image segmentation in terms of solu- using the solution (current tree location) with the best solution
tion quality and computation time. (best tree location) or randomly selected solution (neighbor tree
According to function optimization and image thresholding location). The seeds (candidate solutions) obtained by this opera-
application, it is shown that TSA is an alternative and competitive tion are compared with the current tree location and the best can-
solver for optimizing continuous optimization problems. The main didate solution is replaced with the current tree. More seeds
reason is that TSA uses both random solution strategy (Eq. (4)) and provide to improve the exploitation ability of the method and
the solution strategy with best tendency (Eq. (3)), and one or more affecting randomly selected tree provide to improve the explo-
seed locations are produced for each tree location at the each cycle ration ability of the method. This search mechanism is different
of TSA. Based on standard deviations in the results tables, the TSA from the search mechanisms of the other population-based search
is a robust algorithm against unimodality, multimodality and ini- algorithms.
tialization conditions. It is mentioned that when the dimension The proposed method in this paper cannot be applied to solve
of the problem is increased, the performance of the methods discrete or binary optimization problem as is. Moreover, the per-
decreases because the characteristics of the problem and search formance of the proposed method should be compared with the
space range, which exponentially increases by depending on other methods on solving huge dimensional optimization
dimensionality of the problem, have caused an issue in solving problems.
the problem for the methods. Future works include performance analysis of TSA on con-
strained optimization in short-term and we still pursue to develop
5. Conclusion and future works a discrete version of TSA for solving discrete optimization prob-
lems. Two major topics for TSA in the future are discretization
The present study introduces a new population-based heuristic and adaptation to solve constrained optimization problems. In
search algorithm. The algorithm simulates reproduction behavior addition, the proposed method can be applied to neural network
6698 M.S. Kiran / Expert Systems with Applications 42 (2015) 6686–6698

training, data clustering and optimization of high dimensional Kennedy, J., & Eberhart, R., (1995). Particle swarm optimization. In 1995 IEEE
international conference on neural networks proceedings, Vols. 1–6 (pp. 1942–
problems.
1948).
Kiran, M. S., & Gunduz, M. (2013). A recombination-based hybridization of particle
Acknowledgments swarm optimization and artificial bee colony algorithm for continuous
optimization problems. Applied Soft Computing, 13(4), 2188–2203. http://
dx.doi.org/10.1016/j.asoc.2012.12.007.
The author thanks Professor Binshan Lin and specialist referees Kirkpatrick, S., Gelatt, C. D., & Vecchi, M. P. (1983). Optimization by simulated
for the constructive comments which are presented to improve annealing. Science, 220(4598), 671–680. http://dx.doi.org/
both the content and the presentation of the study. The author also 10.1126/science.220.4598.671.
Martin, D., Fowlkes, C., Tal, D., & Malik, J., (2001). A database of human segmented
would like to thank ‘‘The Scientific Projects Coordinatorship of Selcuk natural images and its application to evaluating segmentation algorithms and
University’’ and ‘‘The Scientific and Technological Research Council of measuring ecological statistics. In Eighth IEEE international conference on
Turkey’’ for their institutional supports. computer vision, Vol II, Proceedings (pp. 416–423).
Otsu, N. (1979). A threshold selection method from gray-level histograms. IEEE
Transactions on Systems Man and Cybernetics, 9(1), 62–66.
References Pal, N. R., & Pal, S. K. (1993). A review on image segmentation techniques. Pattern
Recognition, 26(9), 1277–1294. http://dx.doi.org/10.1016/0031-3203(93)90135-
Akay, B. (2013). A study on particle swarm optimization and artificial bee colony J.
algorithms for multilevel thresholding. Applied Soft Computing, 13(6), Pekdemir, G., (2012). A performance analysis of metaheuristic algorithms on multilevel
3066–3091. http://dx.doi.org/10.1016/j.asoc.2012.03.072. thresholding problems (Master). Konya, Turkey: Selçuk University.
Bhandari, A. K., Kumar, A., & Singh, G. K. (2015). Modified artificial bee colony based Pun, T. (1980). A new method for grey-level picture thresholding using the entropy
computationally efficient multilevel thresholding for satellite image of the histogram. Signal Processing, 2(3), 223–237. http://dx.doi.org/10.1016/
segmentation using Kapur’s, Otsu and Tsallis functions. Expert Systems with 0165-1684(80)90020-1.
Applications, 42(3), 1573–1601. http://dx.doi.org/10.1016/j.eswa.2014.09.049. Rashedi, E., Nezamabadi-Pour, H., & Saryazdi, S. (2009). GSA: A gravitational search
Bhandari, A. K., Singh, V. K., Kumar, A., & Singh, G. K. (2014). Cuckoo search algorithm. Information Sciences, 179(13), 2232–2248. http://dx.doi.org/10.1016/
algorithm and wind driven optimization based study of satellite image j.ins.2009.03.004.
segmentation for multilevel thresholding using Kapur’s entropy. Expert Sezgin, M., & Tasaltin, R. (2000). A new dichotomization technique to multilevel
Systems with Applications, 41(7), 3538–3560. http://dx.doi.org/10.1016/ thresholding devoted to inspection applications. Pattern Recognition Letters,
j.eswa.2013.10.059. 21(2), 151–161. http://dx.doi.org/10.1016/S0167-8655(99)00142-7.
Dorigo, M., Maniezzo, V., & Colorni, A. (1996). Ant system: Optimization by a colony Suganthan, P.N., Hansen, N., Liang, J.J., Deb, K., Chen, Y.-P., Auger, A., & Tiwari, S.,
of cooperating agents. IEEE Transactions on Systems Man and Cybernetics Part B- (2005). Problem definitions and evaluation criteria for the CEC 2005 special session
Cybernetics, 26(1), 29–41. http://dx.doi.org/10.1109/3477.484436. on real-parameter optimization. Singapore: Nanyang Technological University,
Geem, Z. W., Kim, J. H., & Loganathan, G. V. (2001). A new heuristic optimization IIT Kanpur: Kanpur Genetic Algorithms Laboratory.
algorithm: Harmony search. Simulation, 76(2), 60–68. Wang, X. H., & Yan, X. F. (2013). Global best harmony search algorithm with control
Goldberg, D. E. (1989). Genetic algorithms in search, optimization, and machine parameters co-evolution based on PSO and its application to constrained
learning. Reading: Addison Wesley. optimal problems. Applied Mathematics and Computation, 219(19),
Holland, J. H. (1992). Adaptation in natural and artificial systems. Second Edition. 10059–10072. http://dx.doi.org/10.1016/j.amc.2013.03.111.
Cambridge: MIT Press. Yang, X. S. (2010a). Firefly algorithm, stochastic test functions and design
Horng, M. H. (2011). Multilevel thresholding selection based on the artificial bee optimisation. International Journal of Bio-Inspired Computation, 2(2), 78–84.
colony algorithm for image segmentation. Expert Systems with Applications, Yang, X. S. (2010b). A new metaheuristic bat-inspired algorithm. Nicso 2010: Nature
38(11), 13785–13791. http://dx.doi.org/10.1016/j.eswa.2011.04.180. Inspired Cooperative Strategies for Optimization, 284, 65–74.
Hu, X., (2006). Particle swarm optimization tutorial. Retrieved 29.11.2014. Yang, X.S., (2011). Firefly algorithm matlab code. Retrieved 10.12.2014.
Kapur, J. N., Sahoo, P. K., & Wong, A. K. C. (1985). A new method for gray-level Yang, X.S., (2012). Bat algorithm matlab code. Retrieved 18.12.2014.
picture thresholding using the entropy of the histogram. Computer Vision Yang, X. S., & Gandomi, A. H. (2012). Bat algorithm: A novel approach for global
Graphics and Image Processing, 29(3), 273–285. http://dx.doi.org/10.1016/0734- engineering optimization. Engineering Computations, 29(5-6), 464–483. http://
189x(85)90125-2. dx.doi.org/10.1106/02644401211235834.
Karaboga, D., (2008). Artificial bee colony algorithm matlab code. Retrieved Yin, P. Y., & Chen, L. H. (1997). A fast iterative scheme for multilevel thresholding
06.12.2014. methods. Signal Processing, 60(3), 305–313. http://dx.doi.org/10.1016/S0165-
Karaboga, D., & Basturk, B. (2007). A powerful and efficient algorithm for numerical 1684(97)00080-7.
function optimization: artificial bee colony (ABC) algorithm. Journal of Global
Optimization, 39(3), 459–471. http://dx.doi.org/10.1007/s10898-007-9149-x.