Professional Documents
Culture Documents
My KHElsevier
My KHElsevier
My KHElsevier
a r t i c l e i n f o a b s t r a c t
Article history: Krill herd algorithm is a stochastic nature-inspired algorithm for solving optimization problems. The
Received 28 December 2015 performance of krill herd algorithm is degraded by poor exploitation capability. In this study, we propose
Received in revised form 7 April 2016 an improved krill herd algorithm (IKH) by making the krill the global search capability. The enhancement
Accepted 20 April 2016
comprises of adding global search operator for exploration around the defined search region and thus the
Available online 10 May 2016
krill individuals move towards the global best solution. The elitism strategy is also applied to maintain
the best krill during the krill update steps. The proposed method is tested on a set of twenty six well-
Keywords:
known benchmark functions and is compared with thirteen popular optimization algorithms, including
Global optimization
High convergence
original KH algorithm. The experimental results show that the proposed method produced very accurate
Global exploration results than KH and other compared algorithms and is more robust. In addition, the proposed method
Metaheuristic has high convergence rate. The high performance of the proposed algorithm is then employed for data
Data clustering clustering problems and is tested using six real datasets available from UCI machine learning laboratory.
The experimental results thus show that the proposed algorithm is well suited for solving even data
clustering problems.
© 2016 Elsevier B.V. All rights reserved.
1. Introduction food foraging behavior of bird flocks or fish school. Ant Colony
Optimization (ACO) [12] was initially developed by Marco Dorigo,
Over the last few decades, many nature-inspired algorithms simulates the behavior of ants searching for a path between food
have been proposed for solving numerical optimization problems. source and colony. Artificial bee colony (ABC) [11] is another opti-
Nature-inspired algorithms [3–5] play a vital role in solving many mization algorithm developed by Karaboga in 2005 based on the
engineering optimization problems [6–10,23–25,32] owing to the foraging behavior of honey bee swarm. Biogeography-based opti-
global exploration and exploitation ability. These algorithms imi- mization (BBO) [15] is developed by Simon in 2008, is imitating
tate the behavior of living things in nature such as animals, birds, the migrating behavior of species between islands. StudGA (SGA)
fishes, etc. Several heuristic algorithms have been developed in the [20], a variant of GA, is an optimization algorithm in which fittest
literature. Genetic algorithm (GA) [17] was proposed by Goldberg individual is selected rather than stochastic selection and shares
in 1998 simulating the survival of fittest among individuals in the this information among others using genetic algorithm operators.
population over many generations. Evolutionary strategy (ES) [16] Cuckoo search (CS) [21] which is a recently proposed optimization
and differential evolution (DE) [13,14] algorithms belong to the algorithm by Yang et al. in 2009, inspired by the obligate brood
sub class of evolutionary algorithms which use selection, muta- parasitic behavior of cuckoo species. Firefly algorithm (FFA) [22]
tion and recombination operators. Population-based Incremental is developed by Yang in 2010 based on the flashing behaviour of
Learning (PBIL) [18] developed by Shumeet Baluja in 1994, is an fireflies. In Refs. [2,35] authors proposed a new PSO algorithm com-
optimization algorithm which combines the genetic algorithm with bined with levy flight for solving optimization problems. Krill herd
simple competitive learning. Particle Swarm Optimization (PSO) (KH) [1] is introduced by Gandomi and Alavi in recent times based
[19] was proposed by Eberhart and Kennedy in 1995, imitates the on the imitation of krill individuals behavior. Even though there
exists several optimization algorithms, research is still going on
in the development of optimization algorithm which will provide
∗ Corresponding author.
high convergence rate and global optimum solution.
E-mail address: r jensi@yahoo.co.in (R. Jensi).
http://dx.doi.org/10.1016/j.asoc.2016.04.026
1568-4946/© 2016 Elsevier B.V. All rights reserved.
R. Jensi, G.W. Jiji / Applied Soft Computing 46 (2016) 230–245 231
The remaining section of the paper is organized as follows. where Ni is the movement induced by other krill individuals, Fi is
Section 2 provides the variants of krill herd algorithm in solving the foraging action and Di is the random physical diffusion of the
function optimization problems. Section 3 briefly explains the orig- ith krill individuals.
inal krill herd algorithm, and Section 4 presents the proposed IKH The direction of motion induced, ∝i , depends on the three com-
approach. Section 5 provides the experimental results, Section 6 ponents, namely local swarm density, a target swarm density and
provides the experimental results of data clustering and Section 7 a repulsive swarm density. The movement of a krill individual Ni is
is concluded with future discussion. defined as:
n n
i
2
n
f03 Schwefel 1.2 f3 (x) = xj U 20 −100 ≤ xi ≤ 100 0
i=1
j=1
f04 Schwefel 2.21 f4 (x) = max |xi |, 1 ≤ i ≤ n U 20 −100 ≤ xi ≤ 100 0
i
n
2
f06 Step f6 (x) = (xi + 0.5) U 20 −100 ≤ xi ≤ 100 0
i=1
f07 Quartic with noise f7 (x) = ixi 4 + random (0, 1) U 20 −1.28 ≤ xi ≤ 1.28 0
i=1
n
f08 Schwefel 2.26 f8 (x) = 418.9828 × n − xi sin |xi | M 20 −512 ≤ xi ≤ 512 0
i=1
n
f09 Rastrigin f9 (x) = 10n + xi 2 − 10cos (2xi ) M 20 −5.12 ≤ xi ≤ 5.12 0
i=1
n
n
1 1
f10 Ackley f10 (x) = −20 exp −0.2 n
xi 2 − exp n
cos(2xi ) + (20 + e) M 20 −32 ≤ xi ≤ 32 0
i=1 i=1
n n
x
f11 Griewank f11 (x) = 1
4000
xi 2 − cos √i +1 M 20 −600 ≤ xi ≤ 600 0
i
i=1 i=1
2 n−1 2
2
n
f12 Penalized 1 f12 (x) =
n
(10sin (y1 ) + (yi − 1) 1 + 10sin (yi+1 ) + (yn − 1)2 + u (xi , 10, 100, 4) M 20 −50 ≤ xi ≤ 50 0
i=1 i=1
1
yi = 1 + 4 (xi + 1)
2 n−1 2
2
2
n
f13 Penalized 2 f13 (x) = 1
10
sin (3x1 ) + (xi − 1) 1 + sin (3xi+1 ) + (xn − 1)2 1 + sin (2xn ) + u (xi , 5, 100, 4) M 20 −50 ≤ xi ≤ 50 0
i=1 i=1
233
234
Table 1 (Continued)
n
2
f16 Dixon & Price f16 (x) = (x1 − 1)2 + i 2xi2 − xi−1 U 20 −10 ≤ xi ≤ 10 0
i=2
n n
2
f17 Trid10 f17 (x) = (xi − 1) − xi xi−1 M 10 −100 ≤ xi ≤ 100 −200
i=1 i=2
n−1
2
2
2
2
f19 Levy f19 (x) = (xi − 1) 1 + sin (3xi+1 ) + sin (3x1 ) + |xn − 1| 1 + sin (3xn ) M 20 −10 ≤ xi ≤ 10 0
i=1
n−1 2 2
sin2 100x +x −0.5
f20 Pathological function f20 (x) = 0.5 + i i+1
2 M 20 −100 ≤ xi ≤ 100 0
1+0.001 x2 −2xi xi+1 +x2
i=1 i i+1
2
n n
k xi k
f21 Perm f21 (x) = i +ˇ i
−1 M 4 −4 ≤ xi ≤ 4 0
k=1 i=1
n/4
2 2 4 4
f22 Powell f22 (x) = (x4i−3 + 10x4i−2 ) + 5(x4i−1 − x4i ) + (x4i−2 − x4i−1 ) + 10(x4i−3 − x4i ) U 20 −4 ≤ xi ≤ 5 0
i=1
n
n
2
f24 Power sum f24 (x) = xik − bk M 4 0 ≤ xi ≤ 4 0
k=1 i=1
where b = [8,18,44,114]
n n
2 n
4
1 1
f25 Zakharov f25 (x) = xi2 + 2
ixi + 2
ixi M 20 −5 ≤ xi ≤ 10 0
i=1 i=1 i=1
n
−x2
i
1
f26 Wavy1 f26 (x) = 1 − n
cos (kxi ) e 2 M 20 − ≤ xi ≤ 0
i=1
Table 2
Mean real optimization results in twenty six benchmark functions.
f01 1.24E + 03 8.00E + 03 1.75E + 01 2.04E + 03 1.01E + 03 2.43E + 03 1.60E + 04 4.76E + 03 2.51E + 04 8.86E + 03 7.63E + 02 1.64E + 01 1.86E + 02 2.61E − 35
f02 5.49E − 01 4.64E + 01 3.27E + 00 5.06E + 01 7.13E + 00 2.10E + 01 7.92E + 01 3.66E + 01 6.10E + 01 4.93E + 01 1.06E + 01 1.13E + 00 1.23E + 06 2.70E − 25
*The best value obtained by the algorithms for each benchmark function is shown in boldface.
235
236
Table 3
Best real optimization results in twenty six benchmark functions.
F01 1.88E + 00 4.69E + 03 5.71E − 01 8.91E + 02 3.05E + 02 1.12E + 03 1.04E + 04 8.13E + 02 1.46E + 04 5.56E + 03 1.35E + 02 2.17E + 00 8.65E + 00 3.82E − 48
F02 2.92E − 01 3.15E + 01 1.25E + 00 3.26E + 01 3.10E + 00 1.45E + 01 4.97E + 01 1.68E + 01 4.57E + 01 2.72E + 01 4.30E + 00 5.35E − 01 5.19E + 01 2.42E − 27
*The best value obtained by the algorithms for each benchmark function is shown in boldface.
Table 4
Statistical results of IKH for the benchmark functions f01–f11 using Wilcoxon rank sum test (number of runs = 100).
f01 p-value 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34
h-value 1 1 1 1 1 1 1 1 1 1 1 1 1
Zval −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157
f02 p-value 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34
h-value 1 1 1 1 1 1 1 1 1 1 1 1 1
Zval −12.2157 −12.2157 −12.2157 −12.2158 −12.2158 −12.2157 −12.2157 −12.2158 −12.2158 −12.2157 −12.2158 −12.2157 −12.2157
2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34
f04 p-value 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34
h-value 1 1 1 1 1 1 1 1 1 1 1 1 1
Zval −12.2157 −12.2157 −12.2157 −12.2158 −12.2158 −12.2157 −12.2157 −12.2158 −12.2158 −12.2157 −12.2158 −12.2157 −12.2157
f05 p-value 2.56E − 34 2.56E − 34 2.56E − 34 9.26E − 32 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34
h-value 1 1 1 1 1 1 1 1 1 1 1 1 1
Zval −12.2157 −12.2157 −12.2157 −11.727 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157
f06 p-value 5.64E − 39 5.64E − 39 2.1E − 38 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.5E − 39 5.63E − 39
h-value 1 1 1 1 1 1 1 1 1 1 1 1 1
Zval −13.0591 −13.0591 −12.9585 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591 −13.061 −13.0592
f07 p-value 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34
h-value 1 1 1 1 1 1 1 1 1 1 1 1 1
Zval −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157
f08 p-value 2.56E − 34 2.56E − 34 2.56E − 34 9.33E − 24 0.004937 2.56E − 34 2.56E − 34 6.49E − 10 2.56E − 34 2.56E − 34 0.208708 2.72E − 34 2.56E − 34
h-value 1 1 1 1 1 1 1 1 1 1 0 1 1
Zval −12.2157 −12.2157 −12.2157 −10.0484 2.811119 −12.2157 −12.2157 −6.17811 −12.2157 −12.2157 1.257127 −12.2108 −12.2157
f09 p-value 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39
h-value 1 1 1 1 1 1 1 1 1 1 1 1 1
Zval −13.0591 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591
f10 p-value 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39
h-value 1 1 1 1 1 1 1 1 1 1 1 1 1
Zval −13.0591 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591
f11 p-value 2.89E − 34 2.56E − 34 1.54E − 33 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 7.76E − 34 2.64E − 34
h-value 1 1 1 1 1 1 1 1 1 1 1 1 1
Zval −12.206 −12.2157 −12.0691 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2158 −12.1253 −12.2133
Total 11 11 11 11 11 11 11 11 11 11 10 11 11
237
238
Table 5
Statistical results of IKH for the benchmark functions F12–F22 using Wilcoxon rank sum test (number of runs = 100).
f12 p-value 1.21E − 29 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34
h-value 1 1 1 1 1 1 1 1 1 1 1 1 1
Zval −11.3068 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157
f13 p-value 5.41E − 29 2.56E − 34 0.000132 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 0.001549 4.15E − 26
h-value 1 1 1 1 1 1 1 1 1 1 1 1 1
Zval −11.1748 −12.2157 −3.82268 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −3.16541 −10.5689
2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34
f15 p-value 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34
h-value 1 1 1 1 1 1 1 1 1 1 1 1 1
Zval −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157
f16 p-value 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34
h-value 1 1 1 1 1 1 1 1 1 1 1 1 1
Zval −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157
f17 p-value 0.000763 2.56E − 34 1.32E − 30 5.42E − 34 2.91E − 32 2.64E − 34 2.56E − 34 1.95E − 33 2.56E − 34 2.56E − 34 9E − 34 1.71E − 05 5.72E − 30
h-value 1 1 1 1 1 1 1 1 1 1 1 1 1
Zval −3.36577 −12.2157 −11.4998 −12.1546 −11.8248 −12.2133 −12.2157 −12.0496 −12.2157 −12.2157 −12.1131 4.299143 −11.3728
f18 p-value 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34
h-value 1 1 1 1 1 1 1 1 1 1 1 1 1
Zval −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157
f19 p-value 1.03E − 22 2.56E − 34 1.87E − 08 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 1.08E − 09 2.64E − 34
h-value 1 1 1 1 1 1 1 1 1 1 1 1 1
Zval −9.80898 −12.2157 −5.62346 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 6.097477 −12.2133
f20 p-value 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.22E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 0.000131
h-value 1 1 1 1 1 1 1 1 1 1 1 1 1
Zval −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2272 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −3.82513
f21 p-value 0.944483 0.005206 9.53E − 26 1.5E − 12 7.34E − 15 3.32E − 05 1.41E − 28 1.91E − 15 4.78E − 19 9.53E − 12 1.03E − 18 5.75E − 05 6.97E − 27
h-value 0 1 1 1 1 1 1 1 1 1 1 1 1
Zval 0.069637 −2.79402 −10.4907 −7.07484 −7.77855 −4.1501 −11.0893 −7.94717 −8.91715 −6.81339 −8.83167 4.02304 −10.735
f22 p-value 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34
h-value 1 1 1 1 1 1 1 1 1 1 1 1 1
Zval −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157
Total 10 11 11 11 11 11 11 11 11 11 11 11 11
Table 6
f23 p-value 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34
h-value 1 1 1 1 1 1 1 1 1 1 1 1 1
Zval −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157
f24 p-value 0.022124 0.746128 1.54E − 07 7.36E − 12 1.79E − 20 1.19E − 16 4.21E − 33 1.85E − 23 1.25E − 33 2.51E − 32 9.29E − 27 6.17E − 33 1.36E − 15
h-value 1 0 1 1 1 1 1 1 1 1 1 1 1
Zval −2.28823 0.323749 −5.24718 −6.85056 −9.27405 −8.28431 −11.986 −9.98075 −12.0863 −11.837 −10.7085 −11.9543 −7.98866
f25 p-value 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34 2.56E − 34
h-value 1 1 1 1 1 1 1 1 1 1 1 1 1
Zval −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157 −12.2157
f26 p-value 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.63E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39 5.64E − 39
h-value 1 1 1 1 1 1 1 1 1 1 1 1 1
Zval −13.0591 −13.0591 −13.0591 −13.0591 −13.0592 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591 −13.0591
Total 4 3 4 4 4 4 4 4 4 4 4 4 4
239
240 R. Jensi, G.W. Jiji / Applied Soft Computing 46 (2016) 230–245
1: Initialize algorithm parameters Vf, Dmax , Nmax , NP, MI , and minimum and maximum bounds
2: Randomly generate krill individuals (solutions)
3: Evaluate the objective function value f and find the worst and best fitness values
4: iteration=1
5: while (stopping condition is not met) do
6: Store the pre-specified number of best krill
7: for each krill
8: Calculate Movement influenced by other krill individual
9: Calculate Foraging action
10: Calculate Physical diffusion
11: Implement crossover operator.
12: Update krill position using Eq. (7).
13: if rand < 0.5 then
14: Fine-tune the krill position using Eq. (9)
15: else
16: Fine-tune the krill position using Eq. (10)
17: end if
18: Evaluate the objective function value f and update the krill individual if necessary
19: end for
20: Replace the worst krill with the best krill stored before
21: end while
22: Output the global best solution.
Clustering is widely used in many fields of science and engi- where, cj is cluster center for a cluster j and is calculated as follows:
neering and it must often be solved as part of complicated tasks in
pattern recognition, data mining, information retrieval and image 1
cj = di (14)
analysis. The clustering algorithms are mainly classified into two nj
di ∈ cj
[47]: hierarchical and partitional. The most well known partitional
clustering algorithm is K-means which is the center-based clus- where nj is the total number of objects in cluster j.
tering algorithm. The advantage of K-means algorithm is simple
and efficient. But K-means suffers from initial cluster seed selection 6.2. K-means algorithm
since it is easily trapped in local minima. In this section, we intend
to apply the efficiency of proposed IKH algorithm to solve the data K-means [47] is the simplest partitional clustering algorithm
clustering problems. Several clustering approaches are introduced and it is widely used due its simplicity and efficiency. Given a set of
in the literature [26,27,48–57]. N data objects and the number of clusters k, the k-means algorithm
proceeds as follows:
Step 1: Randomly select ‘k’ cluster centers.
6.1. The problem statement
Step 2: Calculate the Euclidean distance between each data point
and cluster centers.
Clustering is the process of partitioning the set of N data objects
Step 3: Assign the data point to the cluster center whose distance
into K clusters or groups based on some distance (or similarity)
from the cluster center is minimum of all the cluster centers.
metric. Let D = {d1 , d2 ,. . .,dN } be a set of N data objects to be
Step 4: Update cluster center using Eq. (14).
partitioned and each data object di , i = 1,2,. . .,N is represented as
Step 5: If no data point was reassigned then stop, otherwise
di = {di1 ,di2 ,. . ..,dim } where dim represents mth dimension value of
repeat from step 2.
data object i.
The aim of clustering algorithm is to find a set of K partitions
C = {C1 ,C2 , . . ..,Ck |∀k: Ck =
/ ∅ and ∀l =/ k: Ck ∩Cl = ∅} in such a way 6.3. IKH for data clustering
that objects within the clusters are more similar and dissimilar
to objects in different clusters. These similarities are measured by To start with IKH for solving data clustering problems, two mod-
some optimization criterions, especially squared error function and ifications have to be done in the IKH algorithm. The first one is
it has been calculated as follows: solution representation and the other is fitness calculation. Given
the data set D and the number of clusters k as input to the algorithm,
k k the solution is represented as a row vector of size k × m, where k
f = min E di , cj (12) is the number of clusters and m is the number of features for the
j=1 i=1
clustering problem and it is shown in Fig. 4.
where cj represents a jth cluster center; E is a distance measure Then the population of solutions is represented as given below:
between a data object di and a cluster center cj . This optimization ⎡ ⎤
S1
criterion is used as the objective function value in our study. There
⎢ S2 ⎥
are many distance metric used in literature. In our study Euclidean ⎢ ⎥
⎢ ⎥
distance is used as distance metric which is defined as follows: P = ⎢ S3 ⎥ (15)
⎢ . ⎥
⎣ .. ⎦
M
E di , cj =
2 SN
dim − cjm (13)
m=1 Si = [C1, C2. . .., Ck] (16)
R. Jensi, G.W. Jiji / Applied Soft Computing 46 (2016) 230–245 241
No Yes
rand<0.5
j=j+1
Yes
j<NP
No
Replace the worst krill with the best krill and
find the current best krill in the swarm
Yes
iteration<
maxiter
No
Output Best Solution
Cj = [cj1, cj2, . . .cjm] , ∀j ∈ 1, 2, . . ., k (17) (12). The remaining steps of KH and IKH algorithms are executed
as described in Sections 3 and 4 respectively.
where k is the number of clusters, m is the set of features, N is the
number krill individuals. 6.4. Experimental results of IKH for data clustering
The population solutions are randomly initialized with the data
objects as cluster centroids. The second modification made to the To evaluate the performance of proposed algorithm for data
algorithm is fitness calculation. Fitness is calculated by using Eq. clustering, six datasets have been used. The datasets, namely,
242 R. Jensi, G.W. Jiji / Applied Soft Computing 46 (2016) 230–245
Fig. 3. Performance Comparison for (a) Rosenbrock (b) Schwefel2.26 (c) Step (d) Penalized1 (e) Alpine (f) Wavy functions.
c12 c13 … …
c11 c1m c21 c22 c23 c2m ck1 ck2 ck3 … ckm
C1 C2 …. Ck
Table 7
Test dataset descriptions.
Iris, Wine, Glass, Wisconsin Breast Cancer (WBC), Contraceptive of other methods. Our proposed algorithm achieves much better
Method Choice (CMC), Vowel and Liver Disorder (LD), are collected results for almost all datasets with small standard deviation. For
from Machine Learning Laboratory [33]. The datasets used in this iris dataset, KH and IKH algorithms converge to 96.6555 for each
study are described in Table 7. run.
In order to evaluate the performance and accuracy of the cluster- For wine dataset, IKH obtains better solution for worst, mean
ing result, total intra-cluster distances as defined in Eq. (12) criteria and standard deviation than K-means, K-means++, GA, SA, TS, ACO,
is used. The low value of the sum is, the higher the quality of the HBMO, PSO, and KH. KH obtains best value of 16292.19, but it fails
clustering is. to reach that value in all runs and thus the KH algorithm gets sixth
With the aim of compare the performance of our proposed algo- rank among all others. As for glass data set, KH and IKH achieve
rithm, several heuristic methods in the literature are used such as best optimum value of 210.242 and 210.252, respectively which
K-means, K-means++ [55], GA [49], SA [48], TS [52], ACO [51], HBMO needs 10,000 fitness function evaluations. The worst value of IKH
[54], PSO [53], whose results are directly taken from Refs. [56,57] is 215.9355, while the worst values over 100 runs for glass dataset
and is given in Table 8. Table 8 lists the best, worst, average and for the algorithms K-means, K-means++, GA, SA, TS, ACO, HBMO,
standard deviation of solutions and ranks the algorithms based on PSO, KH are 227.35, 223.71, 286.77, 287.18, 286.47, 280.08, 249.54,
the mean value for all datasets in Table 7. As compared algorithm 283.52, 251.2749 respectively. Thus IKH algorithm reaches near
results are directly taken from Ref. [56], KH and IKH algorithms best value in all runs.
are executed 100 times independently with same parameters as The best solution for Cancer dataset obtained by IKH and KH
described in section 5 except that the number of krill individuals algorithm is 2964.387. In some cases KH gets stuck in local min-
are set to 25 and maximum number of generations is 400 (Max. imum and thus it does not obtain optimum mean value when
NFE = 10,000). compared to IKH. For CMC dataset, MKH-H achieves best, worst
The experimental results given in Table 8 show that proposed and mean solutions of 5693.72, 5693.779 and 5693.735 with a
algorithm obtains near optimal solutions in compare to those standard deviation of 0.007975, while K-means, K-means++, GA,
Table 8
Comparison of objective function values for different datasets with other methods.
Iris Best 97.3259 97.3259 113.98650 97.4573 97.36597 97.10077 96.752047 96.8942 96.6555 96.6555
Worst 123.9695 122.2789 139.778272 102.0100 98.569485 97.808466 97.757625 97.8973 96.6555 96.6555
Mean 106.5766 98.5817 125.197025 99.9570 97.868008 97.171546 96.95316 97.2328 96.6555 96.6555
Std 12.938 5.578 14.563 2.01 0.53 0.367 0.531 0.347168 1.9E − 06 9.8E − 06
Rank 9 7 10 8 6 4 3 5 1 1
Glass Best 215.73 15.36 282.32 275.16 283.79 273.46 247.71 270.57 210.2421 210.252
Worst 227.35 223.71 286.77 287.18 286.47 280.08 249.54 283.52 251.2749 222.8008
Mean 218.70 217.56 278.37 282.19 279.87 269.72 245.73 275.71 215.7225 215.9355
Std 2.456 2.455 4.138712 4.238 4.192734 3.584829 2.438120 4.557134 5.44876 2.737919
Rank 4 3 8 10 9 6 5 7 1 2
Cancer Best 2988.43 2986.96 3249.46 2993.45 3251.37 3046.06 3112.42 2973.50 2964.387 2964.387
Worst 2999.19 2988.43 3427.43 3421.95 3434.16 3242.01 3210.78 3318.88 3580.312 2964.393
Mean 2988.99 2987.99 2999.32 3239.17 2982.84 2970.49 2989.94 3050.04 2971.977 2964.389
Std 2.469 0.689 229.734 230.192 232.217 90.50028 103.471 110.8013 62.26148 0.001258
Rank 6 5 8 10 4 2 7 9 3 1
CMC Best 5703.20 5703.20 5756.5984 5849.03 5993.594 5819.1347 5713.9800 5700.985 5693.72 5693.72
Worst 5704.57 5705.37 5812.6480 5966.94 5999.805 5912.4300 5725.3500 5923.249 6755.956 5693.779
Mean 5705.37 5704.19 5705,6301 5893.48 5885.062 5701.9230 5699.2670 5820.965 5737.234 5693.735
Std 1.033 0.955 50.3694 50.867 40.84568 45.63470 12.690000 46.95969 178.0245 0.007975
Rank 5 4 10 9 8 3 2 7 6 1
Vowel Best 149,398.66 149,394.56 159,153.498 149,370.47 162,108.5381 159,458.1438 161,431.0431 148,976.0152 148,967.246 148,967.247
Worst 162,455.69 161,845.54 165,991.6520 165,986.42 165,996.4280 165,939.8260 165,804.671 149,121.1834 158,503.045 158,600.525
Mean 151,987.98 151,445.29 149,513.735 161,566.28 149,468.268 149,395.602 149,201.632 148,999.8251 150,035.986 150,172.425
Std 3425.250 3119.751 3105.5445 2847.085 2846.23516 3485.3816 2746.0416 28.8134692 1707.84248 1732.45161
Rank 9 8 5 10 4 3 2 1 6 7
Mean rank 7 5.83 7.67 9.5 6.33 3. 67 3.5 5.33 3.83 2.167
Final rank 8 6 9 10 7 3 2 5 4 1
243
244 R. Jensi, G.W. Jiji / Applied Soft Computing 46 (2016) 230–245
SA, TS, ACO, HBMO, PSO, fail to obtain the best solutions. Similar [18] B. Shumeet, Population-Based Incremental Learning: A Method for
to Cancer dataset, KH algorithm achieves best solution of 5693.72 Integrating Genetic Search Based Function Optimization and Competitive
Learning, Carnegie Mellon University, Pittsburgh, 1994.
for CMC dataset, but fails to reach the value in all runs. For vowel [19] J. Kennedy, R. Eberhart, Particle swarm optimization, in: Proceeding of the
dataset, PSO performs well than all other algorithms including IKH. IEEE International Conference on Neural Networks, Perth, Australia, 1995, pp.
Nevertheless, KH and IKH achieve best solution of 148967.24. As 1942–1948.
[20] W. Khatib, P. Fleming, The stud GA: a mini revolution? in: A. Eiben, T. Back, M.
a conclusion, our proposed scheme reaches better optimal solu- Schoenauer, H. Schwefel (Eds.), Proceeding of the 5th International
tions with a small standard deviation in limited number iterations. Conference on Parallel Problem Solving from Nature, Springer-Verlag,
Based on the ranking analysis, IKH gets first rank among all others. London, 1998, pp. 683–691.
[21] X.S. Yang, S. Deb, Cuckoo search via Levy flights, in: World Congress on Nature
The time consumptions of the algorithms are nearly the same for
& Biologically Inspired Computing, IEEE Publication, USA, 2009, pp. 210–214.
solving each data clustering problems. [22] X.S. Yang, Firefly algorithm, Levy flights and global optimization, in: M.
Bramer, R. Ellis, M. Petridis (Eds.), Research and Development in Intelligent
Systems, XXVI, Springer, London, 2010, pp. 209–218.
7. Conclusion [23] L.-Y. Chuang, H.-W. Chang, C.-J. Tu, C.-H. Yang, Improved binary PSO for
feature selection using gene expression data Comput. Biol. Chem. 32 (2008)
29–38.
Krill herd (KH) is a new optimization method for solving many [24] Y. Zhang, D. Huang, M. Ji, F. Xie, Image segmentation using PSO and PCM with
complex global optimization problems. In this paper, we pre- Mahalanobis distance, Expert Syst. Appl. 38 (2011) 9036–9040.
[25] J.-P. Yang, C.-K. Kung, F.-T. Liu, Y.-J. Chen, C.-Y. Chang, R.-C. Hwang, Logic
sented an improved krill algorithm to solve function optimization circuit design by neural network and PSO algorithm, in: 2010 First
problems. The original krill herd saturated quickly and hence International Conference on Pervasive Computing Signal Processing and
trapped in local minimum. To alleviate the shortcomings of krill Applications (PCSPA), Harbin, China, 2010, pp. 456–459.
[26] R. Jensi, G. Wiselin Jiji, Hybrid data clustering approach using k-means and
herd, improved krill herd was proposed by introducing global
flower pollination algorithm, Adv. Comput. Intell. 2 (2) (2015).
exploration operator. Using these modifications, IKH algorithm [27] R. Jensi, G. Wiselin Jiji, MBA-LF: a new data clustering method using modified
converges to optimal solutions quickly. The simulation results show bat algorithm and Levy flight, ICTACT J. Soft Comput. 6 (1) (2015) 1093–1101.
[28] Gai-Ge Wang, Amir H. Gandomi, Amir H. Alavi, Stud krill herd algorithm,
that our proposed method is fast and efficient for solving function
Neurocomputing 128 (2014) 363–370.
optimization problems. The proposed algorithm IKH is then applied [29] Gai-Ge Wang, Amir H. Gandomi, Amir H. Alavi, An effective krill herd
to data clustering problem. The experimental results are compared algorithm with migration operator in biogeography-based optimization, Appl.
with other methods in the literature and indeed it reveals that the Math. Model. 38 (2014) 2454–2462.
[30] Gai-Ge Wang, Amir H. Gandomi, Amir H. Alavi, Guo-Sheng Hao, Hybrid krill
proposed algorithm is suitable for solving clustering problems. herd algorithm with differential evolution for global numerical optimization,
In the future study, we aim to do the following Neural Comput. Appl. 25 (2014) 297–308.
[31] Gaige Wang, Lihong Guo, Amir Hossein Gandomi, Lihua Cao, Amir Hossein
Alavi, Hong Duan, Jiang Li, Lévy-flight krill herd algorithm, Math. Prob. Eng.
(1) The proposed method is applied to solve other practical engi- 2013 (2013), Article ID 682073, http://dx.doi.org/10.1155/2013/682073, 14
pages.
neering problems, such as scheduling, path planning, text
[32] R. Jensi, G. Wiselin Jiji, A survey on optimization approaches to text document
documents clustering and constrained optimization. clustering, Int. J. Comput. Sci. Appl. 3 (6) (2013) 31–44.
(2) The performance of KH can be improved by hybridizing KH with [33] C.L. Blake, C.J. Merz, University of California at Irvine Repository of Machine
other optimization strategies. Learning Databases, 1998 http://www.ics.uci.edu/mlearn/MLRepository.html.
[34] http://www.mathworks.com/matlabcentral/fileexchange/55486-krill-herd-
algorithm.
[35] R. Jensi, G. Wiselin Jiji, An enhanced particle swarm optimization with levy
References flight for global optimization, Appl. Soft Comput. 43 (c) (2016), http://dx.doi.
org/10.1016/j.asoc.2016.02.018.
[1] A.H. Gandomi, A.H. Alavi, Krill herd: a new bio-inspired optimization [36] X. Yao, Y. Liu, G. Lin, Evolutionary programming made faster, IEEE Trans.
algorithm, Commun. Nonlinear Sci. Numer. Simul. 17 (12) (2012) 4831–4845, Evolut. Comput. 3 (1999) 82–102.
http://dx.doi.org/10.1016/j.cnsns.2012.05.010. [37] G. Wang, L. Guo, H. Wang, H. Duan, L. Liu, J. Li, Incorporating mutation scheme
[2] Hüseyin Haklı, Harun Uguz, A novel particle swarm optimization algorithm into krill herd algorithm for global numerical optimization, Neural Comput.
with Levy flight, Appl. Soft Comput. 23 (2014) 333–345. Appl. 24 (3) (2014) 853–871, http://dx.doi.org/10.1007/s00521-012-1304-8.
[3] R.C. Eberhart, Y. Shi, J. Kennedy, Swarm Intelligence, Morgan Kaufmann, 2001. [38] Lihong Guo, Gai-Ge Wang, Amir H. Gandomi, Amir H. Alavi, Hong Duan, A
[4] X.-S. Yang, Engineering Optimization an Introduction with Metaheuristic new improved krill herd algorithm for global numerical optimization,
Applications, first ed., John Wiley & Sons, New Jersey, 2010. Neurocomputing 138 (2014) 392–402.
[5] X.S. Yang, Nature-Inspired Metaheuristic Algorithms, second ed., Luniver [39] Gai-Ge Wang, Amir H. Gandomi, Amir H. Alavi, Yong-Quan Dong, A hybrid
Press, 2010. meta-heuristic method based on firefly algorithm and krill herd, in: A Hybrid
[6] D.M. Van, A.P. Engelbrecht, Data clustering using particle swarm Meta-Heuristic Method Based on Firefly Algorithm and Krill Herd, IGI Global,
optimization, in: Proceedings IEEE Congress on Evolutionary Computation, 2015, pp. 521–540, http://dx.doi.org/10.4018/978-1-4666-9479-8.ch019.
Canbella, Australia, 2003, pp. 215–220. [40] Dervis Karaboga, Bahriye Akay, A comparative study of artificial bee colony
[7] X. Cui, T. Potok, P. Palathingal, Document clustering using particle swarm algorithm, Appl. Math. Comput. 214 (2009) 108–132.
optimization, Proc. of IEEE Swarm Intelligence Symposium, IEEE Press (2005). [41] Momin Jamil, Xin-She Yang, A literature survey of benchmark functions for
[8] A. Sarangi, R.K. Mahapatra, S.P. Panigrahi, DEPSO and PSO-QI in digital filter global optimization problems, Int. J. Math. Model. Numer. Optim. 4 (2) (2013)
design, Expert Syst. Appl. 38 (2011) 10966–10973. 150–194.
[9] Y.-P. Chang, C.-N. Ko, A PSO method with nonlinear time-varying evolution [42] Ali R. Alroomi, The Farm of Unconstrained Benchmark Functions, University
based on neural network for design of optimal harmonic filters, Expert Syst. of Bahrain, Electrical and Electronics Department, Bahrain (October) 2013,
Appl. 36 (2009) 6809–6816. [Online] Available: http://www.al-roomi.org/cv/publications.
[10] H.-C. Yang, S.-B. Zhang, K.-Z. Deng, P.-J. Du, Research into a feature selection [43] Junpeng Li, Yinggan Tang, Changchun Hua, Xinping Guan, An improved krill
method for hyper spectral imagery using PSO and SVM, J. China Univ. Min. herd algorithm: krill herd with linear decreasing step, Appl. Math. Comput.
Technol. 17 (2007) 473–478. 234 (2014) 356–367.
[11] D. Karaboga, B. Basturk, A powerful and efficient algorithm for numerical [44] Liangliang Li, Yongquan Zhou, Jian Xie, A free search krill herd algorithm for
function optimization: artificial bee colony (ABC) algorithm, J. Glob. Optim. 39 functions optimization, Math. Prob. Eng. 2014 (2014), http://dx.doi.org/10.
(3) (2007) 459–471, http://dx.doi.org/10.1007/s10898-007-9149-x. 1155/2014/936374.
[12] M. Dorigo, T. Stutzle, Ant Colony Optimization, MIT Press, Cambridge, 2004. [45] G.-G. Wang, A.H. Gandomi, X.-S. Yang, A.H. Alavi, A new hybrid method based
[13] R. Storn, K. Price, Differential evolution—a simple and efficient heuristic for on krill herd and cuckoo search for global optimisation tasks, Int. J. Bio
global optimization over continuous spaces, J. Glob. Optim. 11 (4) (1997) Inspired Comput. (in press).
341–359. [46] Gai-Ge Wang, Lihong Guo, Amir Hossein Gandomi, Amir Hossein Alavi, Hong
[14] X. Li, M. Yin, Application of differential evolution algorithm on self-potential Duan, Simulated Annealing-Based Krill Herd Algorithm for Global
data, PLoS One 7 (12) (2012) http://dx.doi.org/10.1371/journal.pone.0051199. Optimization, 2013 (2013), http://dx.doi.org/10.1155/2013/213853.
[15] D. Simon, Biogeography-based optimization, IEEE Trans. Evol. Comput. 12 (6) [47] Jiawei han, Michelin Kamber, Data Mining Concepts and Techniques, Elsevier,
(2008) 702–713. 2010.
[16] H. Beyer, The Theory of Evolution Strategies, Springer, New York, 2001. [48] S.Z. Selim, K.S. Al-Sultan, A simulated annealing algorithm for the clustering
[17] D.E. Goldberg, Genetic Algorithms in Search. Optimization and Machine problem, Pattern Recognit. 24 (10) (1991) 1003–1008.
Learning, Addison-Wesley, New York, 1998.
R. Jensi, G.W. Jiji / Applied Soft Computing 46 (2016) 230–245 245
[49] Ujjwal Maulik, Sanghamitra Bandyopadhyay, Genetic algorithm-based [55] D. Arthur, S. Vassilvitskii, K-means++: the advantages of careful seeding, in:
clustering technique, Pattern Recognit. 33 (2000) 1455–1465. Proceedings of the eighteenth annual ACM-SIAM symposium on discrete
[50] C. Sung, H. Jin, A tabu-search-based heuristic for clustering, Pattern Recognit. algorithms SODA ’07, Philadelphia, PA, USA: Society for Industrial and Applied
33 (2000) 849–858. Mathematics, 2007, pp. 1027–1035.
[51] P.S. Shelokar, V.K. Jayaraman, B.D. Kulkarni, An ant colony approach for [56] Taher Niknam, Babak Amiri, An efficient hybrid approach based on PSO, ACO
clustering, Anal. Chim. Acta 509 (2) (2004) 187–195. and k-means for cluster analysis, Appl. Soft Comput. 10 (2010) 183–197.
[52] Y. Liu, Z. Yi, H. Wu, M. Ye, K. Chen, A tabu search approach for the minimum [57] Ganesh Krishnasamy, Anand J. Kulkarni, Raveendran Paramesran, A hybrid
sum-of-squares clustering problem, Inf. Sci. 178 (2008) 2680–2704. approach for data clustering based on modified cohort intelligence and
[53] Yi-Tung Kao, Erwie Zahara, I-Wei Kao, A hybridized approach to data K-means, Expert Syst. Appl. 41 (2014) 6009–6016.
clustering, Expert Syst. Appl. 34 (3) (2008) 1754–1762.
[54] M. Fathian, B. Amiri, A honey-bee mating approach on clustering, Int. J. Adv.
Manuf. Technol. 38 (2008) 809–821.