Arabian Journal for Science and Engineering (2018) 43:7499–7520

https://doi.org/10.1007/s13369-018-3155-9

RESEARCH ARTICLE - COMPUTER ENGINEERING AND COMPUTER SCIENCE

A Comparative Analysis of Metaheuristic Approaches for

Multidimensional Two-Way Number Partitioning Problem
Mehmet Hacibeyoglu1 · Kemal Alaykiran2 · Ayse Merve Acilar1 · Vahit Tongur1 · Erkan Ulker3

Received: 26 June 2017 / Accepted: 12 February 2018 / Published online: 23 February 2018
© King Fahd University of Petroleum & Minerals 2018

Abstract
In this study, a novel usage of four metaheuristic approaches Genetic algorithm (GA), Simulated annealing (SA), migrating
bird optimization algorithm (MBO) and clonal selection algorithm (CSA) are applied to multidimensional two-way number
partitioning problem (MDTWNPP). MDTWNPP is a classical combinatorial NP-hard optimization problem where a set of
vectors have more than one coordinate is partitioned into two subsets. The main objective function of MDTWNPP is to
minimize the maximum absolute difference between the sums per coordinate of elements. In order to solve this problem,
GA is applied with greedy crossover and mutation operators. SA is improved with dual local search mechanism. MBO is
specialized as multiple flock migrating birds optimization algorithms. CSA is applied with problem specific hyper mutation
process. Furthermore, all instances are solved using an integer linear programming model which was previously presented
in the literature. In the experiments, four metaheuristic approaches and integer linear programming model are used to solve
126 datasets with different sizes and coordinates. As a brief result, the GA and SA approaches designed for this problem
outperformed all other heuristics and the integer programming model. Both the performance of GA and SA approaches are in
a competitive manner where GA and SA yielded the best solution for 56 and 65 out of 125 datasets, respectively.

Keywords Multidimensional two-way number partitioning problem · Integer linear programming · Genetic algorithms ·
Migrating birds optimization algorithm · Clonal selection algorithm · Simulated annealing

1 Introduction as “The easiest hard problem” in the literature [1]. In

computer science and number theory, Two-way Number Par-
Number partitioning problem (NPP) is a classical, com- titioning Problem (TWNPP) is the task of partitioning a given
pelling and attractive optimization problem which is defined finite set of elements N into two subsets S1 and S2 where the
sum of coordinates in S1 equals the sum of coordinates in S2
B Mehmet Hacibeyoglu [2,3]. The MDTWNPP is a combinatorial NP-hard optimiza-
[email protected] tion problem which is a special form of TWNPP [4]. The
Kemal Alaykiran main difference between MDTWNPP and TWNPP is that
[email protected] in MDTWNPP multiple coordinates are considered which
Ayse Merve Acilar makes the problem harder. MDTWNPP has importance both
[email protected] from a theoretical and practical point of view. MDTWNPP is
Vahit Tongur used practically in many important real-life application like
[email protected] multiprocessors scheduling, public key encryption, database
Erkan Ulker processing and task scheduling, minimization of VSLI circuit
[email protected] size, choosing fair teams [1,5]. In multiprocessors schedul-
ing, there are number of jobs with different run times and
1 Department of Computer Engineering, Necmettin Erbakan number of parallel processors. In order to solve this prob-
University, Konya, Turkey
lem, these jobs are assigned to a processor to minimize the
2 Department of Industrial Engineering, Necmettin Erbakan completion time of the whole tasks [6,7]. In public k encryp-
University, Konya, Turkey
tion, Merkle-Hellmann system is based on a similar version
3 Department of Computer Engineering, Selcuk University, of TWNPP [8]. In choosing fair teams, a set of persons with
Konya, Turkey

123
7500 Arabian Journal for Science and Engineering (2018) 43:7499–7520

different competencies are considered and teams with equal [14], memetic algorithm [15], simulated annealing [16], inte-
powers are intended to be assigned [9]. This definition of ger linear programming model (ILP) [17] and Grasp [18].
the problem is quite the same as MDTWNPP since, every Recently, [19] developed three metaheuristic algorithms
competency may be thought as a coordinate of MDTWNPP Greedy algorithm, GA and hybrid GA and Variable Neigh-
and for a fair selection the difference between these compe- borhood Search (VNS) for solving bi-dimensional two-way
tencies must be minimized which is the same objective as number partitioning problem. VNS is applied inside GA with
MDTWNPP. shaking and local search operations in order to avoid falling
In this study, a set of problem instances with different into local optimum and finding encouraging results. The
sizes and different coordinates are created with 3 decimal hybrid GA-VNS algorithm obtained more favorable results
digits to have more challenging problems. First the math- in comparison with the Greedy algorithm and GA. Refer-
ematical model proposed in [4] is coded using General ence [20] presented an efficient memetic algorithm to solve
Algebraic Modeling System (GAMS) and the problems are the multidimensional multiway number partitioning problem
solved using CPLEX solver with a time limit of 1800 s which obtained by combining a GA with a powerful local search
is determined to be a reasonable time limit for the solution procedure. Memetic algorithm is hybridized with GA using
of these problems after empirical trials. Then, the problem generation of the initial population, crossover and muta-
instances are solved using four metaheuristic approaches: (1) tion operations. Initial population generated partially random
Greedy Genetic algorithm (GGA) which is a novel use of GA. and partially based on the MDTWNPP. In crossover opera-
MDTWNPP has been previously solved using GA but in this tion, single point crossover is selected and 85% of the new
study, a contribution is made in mutation and crossover oper- generation is formed with crossover operation and 15% is
ations in GA. (2) Multiple flock migrating bird optimization copied directly from the previous generation. Additionally,
algorithm (MFMBO) which is a problem specific modifi- in mutation operation new offsprings are generated by flip-
cation of MBO. MBO is a new but popular metaheuristic ping values with 10% probability. The results of the proposed
that attracts interest of researchers in the literature. To the memetic algorithm compared with the results obtained from
best of the authors’ knowledge this metaheuristic has not the ILP model proposed in [3]. In [21], MDTWNPP was
been used for MDTWNPP. Furthermore, after experimen- solved using two metaheuristic approaches, a VNS and an
tal studies a contribution to the definition of MBO is made Electromagnetism-like metaheuristic (EM). The VNS starts
and MFMBO is proposed as a new use of existing algorithm. with a random solution and 1-swap shaking is performed
(3) Clonal selection algorithm (CSA) which is more often in the local search procedure. In the EM approach, local
used for solving continuous optimization problems. As a search, scaling and moving stages are used for evolving the
contribution, this metaheuristic is tuned to solve a discrete solutions. In the experiment stage, the results of VNS and
optimization problem MDTWNPP. Furthermore, this is the EM are compared to the results of memetic algorithm in
first study using this algorithm for MDTWNPP. (4) Simulated [20] and ILP model in [3]. [4] presented the combination
Annealing (SA) with dual local search which is a novel used of GRASP with Exterior Path Relinking with a novel local
of SA for increasing the performance of local search. All search. GRASP is a multistart process and that performs con-
well-known metaheuristics used in this study are designed struction and solution improvements. GRASP is hybridized
and improved for efficiently solving MDTWNPP which is a with Path Relinking with elite solutions in order to make ben-
major contribution of this study. efit from the memory function that Path Relinking uses about
The paper is organized as follows: The next chapter gives the search space. In the local search process, swap method
the literature for the problem. At Sect. 3, the problem is is used until a better neighbor than the current solution is
defined in detail. At Sect. 4, the ILP model proposed in found. This process is named as first interchange local search
[3] and used in this study is given. The used metaheuristic method. The empirically compared experimental results of
approaches GGA, MFMBO and CSA are explained in detail at this study shows, proposed GRASP + ePR allows get better
Sects. 5, 6, 7 and 8, respectively. At Sect. 9 the computational results than VNS and ILP. Since the problem is NP-hard, the
results are given and the paper is finalized with conclusions. used metaheuristics in the literature performed better results
than exact methods in reasonable solution time.

2 Literature Review 3 Problem Definition

Number partitioning problem is solved with different In MDTWNPP, there is a set of vectors N where each element
approaches in the literature such as exact solutions [10], n i=1... p ∈ N is defined as k coordinates where k > 1. N=
greedy heuristic methods [11,12], a hybridization of tree n 1 (c1 , c2 , .., ck ) , n 2 (c1 , c2 , .., ck ) , .., n p (c1 , c2 , .., ck ) is
search, breadth first search and beam search [13], tabu search an example problem instance where p = |N | and kis the

123
Arabian Journal for Science and Engineering (2018) 43:7499–7520 7501

number of coordinates. The main objective of the problem is 4 Integer Linear Programming Model
to split this problem instance into two subsets (S1 , S2 ) where
S1 ∪ S2 = N , S1 ∩ S2 = ∅ and differences of the sums of In this study, the ILP model proposed in [17] is used.
each coordinate in the subsets S1 and S2 will be minimized.
The total number of the feasible solutions for each problem • The parameters of the ILP model are as follows:
instance is 2 p while every unique solution has a pair with
the same objective function value so that the total number of N The set of vectors,
unique solutions is 2 p /2. An example problem instance with i Number of Elements,
p = 8 and k = 3 and step by step example solution is given k Number of Coordinates,
below: S1 First subset,
S2 Second subset,
wik the vector at kth coordinate of ith element in set
Step 1 Determining an example problem instance with
N,
p = 8 and k = 3:
• The decision variable of ILP which shows whether the
N1 = {(1 4 2), (7 3 1), (6 5 9), (4 8 6), element is a member of S1 or S2 .

(2 2 5), (9 1 3), (2 4 5), (8 1 6)} 

1, if i ∈ S1
xi =
0, if i ∈ S2
Step 2 Splitting the problem instance into two example sub-
sets S1 and S2 where S1 has three elements and S2 • The objective of this problem is to minimize the maxi-
has five elements: mum difference between the vectors of S1 and S2 for all
coordinates. The objective function value (t) is calculated
S1 = {(1 4 2), (7 3 1), (6 5 9)} as follows in order to overcome nonlinearity :
S2 = {(4 8 6), (2 2 5), (9 1 3), (2 4 5), (8 1 6)}  
  
 

t = max  wik − wik 
Step 3 Calculating the sum of three coordinates for subsets i∈S1 i∈S2 
S1 and S2
• The objective function and the constraints of ILP are as
Sum of S1 = {(1 + 7 + 6 / 4 + 3 follows:
+5/2 + 1 + 9)} = {14/12/12}
min (t) (1)
Sum of S2 = {(4 + 2 + 9 + 2 + 8 / 8 + 2 + 1 + 4
+1/6 + 5 + 3 + 5 + 6)} = {25/16/25} Subject to:
n
Step 4 Calculating the absolute differences of all coordi- −0.5 ∗ t + wik ∗ xi ≤ 0.5 ∗ Nk ∀k (2)
i=1
nates separately: n
0.5 ∗ t + wik ∗ xi ≥ 0.5 ∗ Nk ∀k (3)
i=1
The differences are: {|14 − 25| / |12 − 16| / xi ∈ {0, 1} (4)
|12 − 25|} = {11 / 4 / 13}
The objective function (1) is to minimize the absolute dif-
ferences of all coordinates. Constraints (2) and (3) are to
Step 5 Calculating the objective function with selecting the
ensure that t has the smallest value close to 0.
maximum difference between the coordinates:

O (N1 ) = max (11, 4, 13) = 13 5 Greedy Genetic Algorithm (GGA) for

MDTWNPP
As seen at above example, the objective function is cal-
culated in five steps. The main goal of the problem is GA is a robust metaheuristic algorithm for solving optimiza-
determining the subsets S1 and S2 for minimizing the objec- tion problems which was introduced by John Holland in
tive function such that the sum of numbers in S1 and S2 as the early 1970s [22]. GA starts working with the set of ini-
nearly equal as possible. tial chromosomes called population and uses the survival of

123
7502 Arabian Journal for Science and Engineering (2018) 43:7499–7520

Table 1 GGA representation

Problem set S Chromosome Cutoff point
with nine example
chromosomes  
(1 4 2) , (6 5 9) , (2 4 5) , (7 3 1) ,
S= Ch1 = 2573|1468 4
(4 8 6) , (2 2 5) , (9 1 3) , (8 1 6)
Ch2 = 12563|847 5
Ch3 = 273|14568 3
Ch4 = 13|578642 2
Ch5 = 8517|4263 4
Ch6 = 382415|76 6
Ch7 = 41672|835 5
Ch8 = 6475|1823 4
Ch9 = 764|25138 3

Table 2 The initial population

Chromosome Subsets
with nine example
chromosomes Ch1 Ch1 S1 = {(6 5 9),(4 8 6),(9 1 3),(2 4 5)}
Ch1 S2 = {(1 4 2),(7 3 1),(2 2 5),(8 1 6)}
Ch2 Ch2 S1 = {(1 4 2),(6 5 9),(4 8 6),(2 2 5),(2 4 5) }
Ch2 S2 = {(8 1 6),(7 3 1),(9 1 3)}
Ch3 Ch3 S1 = {(6 5 9),(9 1 3),(2 4 5)}
Ch3 S2 = {(1 4 2),(7 3 1),(4 8 6),(2 2 5),(8 1 6)}
Ch4 Ch4 S1 = {(1 4 2),(2 4 5)}
Ch4 S2 = {(4 8 6),(9 1 3),(8 1 6),(2 2 5),(7 3 1),(6 5 9)}
Ch5 Ch5 S1 = {(8 1 6),(4 8 6),(1 4 2),(9 1 3)}
Ch5 S2 = {(7 3 1),(6 5 9),(2 2 5),(2 4 5)}
Ch6 Ch6 S1 = {(2 4 5),(8 1 6),(6 5 9),(7 3 1),(1 4 2),(4 8 6)}
Ch6 S2 = {(9 1 3),(2 2 5)}
Ch7 Ch7 S1 = {(7 3 1),(1 4 2),(2 2 5),(9 1 3),(6 5 9)}
Ch7 S2 = {(8 1 6),(2 4 5),(4 8 6)}
Ch8 Ch8 S1 = {(2 2 5),(7 3 1),(9 1 3),(4 8 6)}
Ch8 S2 = {(1 4 2),(8 1 6),(6 5 9),(2 4 5)}
Ch9 Ch9 S1 = {(9 1 3),(2 2 5),(7 3 1)}
Ch9 S2 = {(6 5 9),(4 8 6),(1 4 2),(2 4 5),(8 1 6)}

the best chromosomes strategy. In this strategy, while creat- the sequence of elements in the problem set N . Also this rep-
ing a new generation, the better chromosomes are selected resentation must provide that the set of elements belonging
and worse ones are eliminated. Then crossover and mutation to N is partitioned into two subsets S1 and S2 . Cutoff point is
operators are performed to increase the diversity of popula- used for this partitioning process. While the elements which
tion and avoid local optimization [23]. GA continues to work are on the left side of the cutoff point are included to the sub-
until an optimal solution or approximately optimal solution set S1 , on the right-side elements are included to the subset
is obtained. There are four steps in GGA. These are repre- S2 . An example of GGA representation with nine examples
sentation of problem instance, generating initial population, is shown in Table 1.
calculating the fitness value and implementing the selection
and applying crossover and mutation genetic operators.
5.2 Generating Initial Population of GGA

5.1 Representation of the Problem Instance of GGA The initial population of the GGA is obtained with random
generation. The population consists of chromosomes which
In order to solve MDTWNPP problem, we represented a are the set of genes. For this problem, the gene is the sequence
chromosome with using permutation coding where every number of the elements in the problem set N . For creating
chromosome is a fixed-size (p elements) randomly ordered a chromosome, genes are selected randomly from the range

123
Arabian Journal for Science and Engineering (2018) 43:7499–7520 7503

Table 3 Calculation of fitness value of nine chromosomes Table 4 Tournament selection with seven candidates
Chromosome Sum of coordinates Fitness value Seven candidates Champion chromosome

Ch1 {(21,18,23),(18,10,14)} max(3,8,9) = 9 Ch1 , Ch2 , Ch6 , Ch5 , Ch4 , Ch3 , Ch7 Ch5
Ch2 {(15,23,27),(24,5,10)} max(9,18,17) = 18 Ch2 , Ch3 , Ch7 , Ch8 , Ch1 , Ch9 , Ch6 Ch8
Ch3 {(17,10,17),(22,18,20)} max(5,8,3) = 8 Ch9 , Ch3 , Ch6 , Ch4 , Ch7 , Ch2 , Ch1 Ch3
Ch4 {(3,8,7),(36,20,30)} max(33,12,23) = 33 Ch1 , Ch2 , Ch6 , Ch5 , Ch4 , Ch3 , Ch7 Ch5
Ch5 {(22,14,17),(17,14,20)} max(5,0,3) = 5 Ch3 , Ch5 , Ch2 , Ch1 , Ch4 , Ch6 , Ch8 Ch5
Ch6 {(28,25,29),(11,3,8)} max(17,22,21) = 22 Ch8 , Ch7 , Ch1 , Ch3 , Ch4 , Ch2 , Ch9 Ch8
Ch7 {(25,15,20),(14,13,17)} max(11,2,3) = 11 Ch5 , Ch1 , Ch4 , Ch2 , Ch9 , Ch7 , Ch3 Ch5
Ch8 {(22,14,15),(17,14,22)} max(5,0,7) = 7 Ch4 , Ch2 , Ch9 , Ch6 , Ch3 , Ch1 , Ch7 Ch3
Ch9 {(18,6,9),(21,22,28)} max(3,16,19) = 19 Ch6 , Ch9 , Ch3 , Ch7 , Ch2 , Ch1 , Ch8 Ch8

Table 5 New generation

1 and p. As a result, we obtained a chromosome, which is
Chromosome Fitness value
a random ranked sequence of numbers from 1 to p. Lastly,
for obtaining the subset S1 and S2 of the chromosome, the Ch1 = 8517|4263 5
cutoff point is selected randomly from the range 2 and p − 1. Ch2 = 6475|1823 7
These processes are repeated up to the population number. Ch3 = 273|14568 8
An example of this initial population generation process is Ch4 = 8517|4263 5
given in Table 2. Ch5 = 8517|4263 5
Ch6 = 6475|1823 7
5.3 Calculating Fitness Value and Implementing Ch7 = 8517|4263 5
Selection of GGA Ch8 = 273|14568 8
Ch9 = 6475|1823 7
GGA needs a fitness value which defines the importance of
the chromosome in the population. The fitness value is calcu-
lated with the objective function which is described at Sect. 2. the GGA. We chose to make greedy crossover and mutation
Also, MDTWNPP is a minimization optimization problem operations for MDTWNPP problem because full randomness
so the power of the chromosome is inversely proportional often cannot find the optimal or approximately optimal solu-
to the fitness value of the chromosome. If the fitness value tion. Primarily in the crossover operation, a random vector
of the chromosome increases the power of the chromosome is selected from subset S1 . After that, the selected vector is
decreases and vice versa. The selection process is used to made swap one by one with the vectors in the subset S2 . The
select powerful chromosomes with lower fitness value for fitness value is calculated for each swap. Consequently, the
creating the new generation. We have implemented tourna- swap with the lowest fitness value is selected for a crossover
ment selection method with seven candidates for selection operation. The crossover operations are performed up to the
process [24]. Tournament selection process is repeated up to product of population number and crossover rate. After the
the population number. An example of calculation of fitness greedy crossover operation, a greedy mutation operation will
values for nine chromosomes and tournament selection with be performed. Two examples of greedy crossover operation
seven candidates are given in Tables 3 and 4, respectively. are shown in Table 6.
The new generation created as a result of these processes is Insertion method is selected for the mutation operation to
given in Table 5. obtain the optimum size of the subsets which is the deletion of
a vector in the selected subset and insertion to the other sub-
5.4 Applying Genetic Operators of GGA: Crossover set. To choose a subset for deletion, all elements are added in
and Mutation Operations all coordinates for the first and second subset, respectively. If
the results of any set is greater than the results of the other sub-
When new generation is created with selection process, a set for all the coordinates, this subset is selected for deletion.
proportion of the new generation is formed from the same In other cases, the subset selection process is done randomly.
chromosomes. So, the crossover and mutation processes are For finding the best mutation operation, one by one all the
performed to increase the diversity of chromosomes. In other elements in the selected subset are deleted and inserted to the
words, the crossover and mutation ratios are considered to other subset. The fitness value is calculated for each deletion
be as much as possible to avoid premature convergence of and insertion operations. Eventually, the lowest fitness value

123
7504 Arabian Journal for Science and Engineering (2018) 43:7499–7520

Table 6 Two greedy crossover

Chromosome/ Random vector Crossovers New chromosome/new
examples
fitness value (swaps)/fitness values fitness value

Ch1 = 8517|4263/5 5 8174|2635/13 Ch1 = 8173|4265/8

8172|4635/9
8176|4235/12
8173|4265/8
Ch6 = 6475|1823/7 7 6451|8237/11 Ch6 = 6458|1237/3
Bold value refers to the best individual solution

is selected for the mutation operation. The mutation opera- In addition MFMBO for MDTWNPP uses swap and insert
tions are performed up to the product of population number methods together for increasing the diversity of solutions.
and mutation rate. After the crossover and mutation processes The MFMBO consists of four steps. These are: generat-
the current fitness value of the chromosome may increase or ing initial population, generating neighborhoods, applying
decrease. This change is not important for the progress of the neighborhood sharing and implementing leader replacing.
algorithm. Here, it is important to ensure the diversity of the
population. The examples of the greedy mutation operation 6.1 Generating Initial Population of MFMBO
are shown in Table 7.
All operations of GGA are done according to the iteration MFMBO algorithm starts with random initial solutions and
number. As we will see in the experimental results section, tries to improve these solutions. Each bird in the population
our proposed GGA is effective in producing good solutions. represents a solution. One of these solutions is assigned as the
leader bird and the remaining are placed on the left and the
right side of the leader solution. Thereby, the flock is created
6 Multiple Flock Migrating Birds as it is demonstrated at Fig. 1a. MFMBO algorithm has a lot
Optimization Algorithm (MFMBO) for of flocks in the search space. Flocks are independent from
MDTWNPP each other to scan the search space. These flocks are shown
in Fig. 1b.
MBO is inspired by the energy saving procedure of the birds An example representation of a flock for MDTWNPP is
during migration which has a V formation. The algorithm given in Fig. 2. The problem set and representation of a solu-
firstly designed for discrete optimization problems and tested tion is like Table 1.
on quadratic assignment problem (QAP) arised from real-life
problems [25]. Also, this algorithm is applied on credit card 6.2 Generating Neighborhood (Local Search) of
fraud detection and scheduling problems [26–28]. MFMBO
It is observed from the experimental results that if basic
MBO is applied to MDTWNPP, the solution easily falls down MFMBO algorithm has a neighborhood approach which may
to the local optimum. In order to overcome this burden be used in local search. The number of the neighbors of the
MFMBO is proposed. The main difference between MBO leader l must be an odd number and l ≥ 3. The number
and MFMBO is that MFMBO uses multiple bird flocks at the of shared neighborhood m must be 1  m  (l − 1) /2.
same time. So that the solution space can be searched better. The number of neighbors of other solutions n is equal to

Table 7 Two greedy mutation examples

Chromosome/fitness value Sum of coordinates Selected subset Mutations/fitness values New chromo-
some/new fitness
value

Ch3 = 273|14568/8 (17, 10, 17)/(22, 18, 20) S2 2731|4568/3 Ch3 = 2731|4568/3
2734|1568/9
2735|1468/9
2736|1458/7
2738|1456/ 11
Ch9 = 6475|1823/7 (22, 14, 15)/(17, 14, 22) S1 475|18236/17 Ch9 = 675|18234/9
Bold value refers to the best individual solution

123
Arabian Journal for Science and Engineering (2018) 43:7499–7520 7505

Fig. 1 Initial population for MFMBO. a Single flock, b multiple flock

B0 = 2573|1468
B1 = 12563|847 B2 = 273|14568
B3 = 13|578642 B4 = 8517|4263
B5 = 382415|76 B6 = 41672|835
B7 = 6475|1823 B8 = 764|25138

Fig. 2 An example of flock representation for MDTWNPP

n = l − m. For applying local search to generate neighbor- 6.4 Implementing Leader Replacing of MFMBO
hood one-point swap mechanism and insertion are used. The
number of generated neighbor solutions for the leader has to The number of flaps parameter denotes to how long the flock
be at least three as a requirement of the algorithm. These solu- fly in the same order. Afterward, leader bird flies to the end
tions are ordered according to the objective function from the of the flock. One of the birds following the leader replaces its
best to the worst and the current solution is replaced with the position. Primarily the leader goes to left side of the flock and
best neighbor solution in order to improve the current solu- flock continues to fly with the new order. This procedure is
tion. Otherwise, best neighbor solution is discarded and the the same for the right side of the flock. This cycle continues
procedure continues with the existing solutions. This proce- till the algorithm is terminated. Leader replacing procedure
dure is the same for the remaining solutions. is showed in Fig. 3.
In the MFMBO, generating neighborhood, applying neigh-
6.3 Applying Neighborhood Sharing of MFMBO borhood sharing and implementing leader replacing opera-
tions are repeated up to the iteration number (Table 8).
The most important feature of MFMBO that is distinctive
from other metaheuristic algorithms is the benefit mecha-
nism. This mechanism may be described as neighborhood
sharing of birds in the flock. The leader bird generates three 7 Clonal Selection Principle and CLONALG
or more neighbors. It uses the best of these neighbors in order Algorithm
to improve the current solution and transfers m neighbors to
the bird on the left side and to the right side. For the birds Clonal selection algorithm is inspired by clonal selection the-
to the right and left to the leader bird, the neighbor solutions ory that explains the basic features of an immune system in
are reevaluated due to the objective function and reordered response to antigens. In this theory, cells that recognize the
in order to improve their current solutions. Afterward, they antigens are selected to proliferate. For improving the affin-
transfer their m neighbors to the following birds. This pro- ity of the selected cells against to the antigens, off springs
cess continues until the last bird in the flock has its neighbor are generated by mutation and reproduction processes. The
solution. An example neighbor generation and sharing for higher number of offspring means higher affinity [29]. In
the first three birds of the flock is given in Table. 8. [30] a clonal selection algorithm named CLONALG based

123
7506 Arabian Journal for Science and Engineering (2018) 43:7499–7520

Table 8 An example neighbor

Bird (B)/fitness Selected posi- Neighbor New bird/new fit-
generation and sharing of the
value tion/coordinates (N)/fitness value ness value
flock for the first three birds
B0 = 2573|1468/9 2/3 & (Swap) N11 = 2673|1458/7 B0 = 2673|1458/7
3/3 & (Insertion) N12 = 253|14768/15
5/2 & (Swap) N13 = 25734|168/17
B1 = 12563|847/18 2/1 & (Swap) N21 = 1253|8476/14 B1 = 18563|247/11
4/4 & (Insertion) N22 = 18563|247/11
Shared N12 N23 = 253|14768/15
B2 = 273|14568/8 1/3 & (Insertion) N31 = 5273|1468/9 B2 = 273|14568/8
4/4 & (Insertion)
Shared N12
Shared N13 N32 = 25734|168/17
2/4 & (Swap) N33 = 263|14578/19
Bold values refer to the best individual solution

Fig. 3 Leader replacing of MFMBO

on clonal selection principle is proposed. This algorithm is 7.2 Cloning of CSA

a population-based stochastic method, with binary represen-
tation of the variables [31]. CLONALG is applied to perform In the CSA, the selected Abs will be cloned for generating
machine-learning, pattern recognition tasks and optimization a repertory C of clones. The number of the clones placed in
problems [30,32]. the repertory C is calculated according to Eq. (5):
In this study, CLONALG algorithm has been adapted to 
A
MDTWNPP and called as CSA. There are three steps in NCc = round (β ∗ A) (5)
CSA. These are antibody (Ab) encoding and affinity func- 1
tion, cloning and hyper mutation function.
where NCc is the total amount of clones generated for each
7.1 Antibody Encoding and Affinity Function Ab, β is a multiplying factor, A is the total amount of Abs
(Objective Function) of CSA and round is the operator that rounds its argument toward
the closest integer. Each term of this sum corresponds to the
In this study, an Ab is encoded by a string with binary values clone size of each selected Ab [30]. For example, we take β
with the length of |N |. N is divided into two subsets S1 and as 0.5 and total amount of Abs in our initial population is 5.
S2 . The initial Ab population and its sum of coordinates and As stated by Eq. 5, 3 clones are generated for each Ab and
affinity values are given in Tables 9 and 10, respectively. N Cc is equal to 15 as a result. Multiplying factor β could be
Table 9 expresses the initial Ab population. Column 2 and changed according to population size. For the high number
3 are the S1 and S2 subsets. Bits represented by 0 in Abi Abs, a small β value could be selected to avoid complexity.
formed S1 and bits represented by 1 formed S2 . The affinity
function which is calculated according to that described in 7.3 Hyper Mutation Process of CSA
Sect. 2 as the objective function is given in column 3 of
Table 10. Random initialization is used to generate the initial In the original CLONALG algorithm for real-valued attributes,
population which consists of a predefined number of Abs. the hyper mutation rate to be applied to every immune cell is

123
Arabian Journal for Science and Engineering (2018) 43:7499–7520 7507

Table 9 Initial Ab population

Antibodies S1 {represented by 0 in Ab} S2 {represented by 1 in Ab}

Ab1 = 01011101 {(1 4 2),(6 5 9),(2 4 5)} {(7 3 1),(4 8 6),(2 2 5),(9 1 3),(8 1 6)}
Ab2 = 11001001 {(6 5 9),(4 8 6),(9 1 3),(2 4 5)} {(1 4 2),(7 3 1),(2 2 5),(8 1 6)}
Ab3 = 01000101 {(1 4 2),(6 5 9),(4 8 6),(2 2 5),(2 4 5)} {(7 3 1),(9 1 3),(8 1 6)}
Ab4 = 11011001 {(6 5 9),(9 1 3),(2 4 5)} {(1 4 2),(7 3 1),(4 8 6),(2 2 5),(8 1 6)}
Ab5 = 01011000 {(1 4 2),(6 5 9),(9 1 3),(2 4 5),(8 1 6)} {(7 3 1),(4 8 6),(2 2 5)}

Table 10 Sum of coordinates

Antibodies Sum of coordinates Affinity
and affinity values of initial Ab
population Ab1 = 01011101 {(9,13,16),(30,15,21)} max(21,2,5) = 21
Ab2 = 11001001 {(21,18,23),(18,10,14)} max(3,8,9) = 9
Ab3 = 01000101 {(15,23,27),(24,5,10)} max(9,18,17) = 18
Ab4 = 11011001 {(17,10,17),(22,18,21)} max(5,8,4) = 8
Ab5 = 01011000 {(26,15,25),(13,13,12)} max(13,2,13) = 13

inversely proportional to its affinity. The mutation operator normalized in the interval [0…1]. As c represents a candi-
can be described as Eq. (6). date solution, it must be within the range of the functions
specified domain [30]. However, TWNPP is a discrete so we
c = c + ∗ N (0, 1) (6) proposed new hyper mutation scheme accordance with this
problem in CSA.
where, c is the clone, c is the mutated clone, α is the amount In this scheme, firstly the f ∗ is calculated. According to
of mutation (step size was defined by user) and N (0, 1) is a Table 9, our f ∗ value is 1 because Ab1 is the individual which
Gaussian random vector of zero mean and unitary variance. have worst affinity value. Length of the Ab1 is 8. Mutation
The amount of mutation α has been calculated according to rate ρ is taken as 0.3 for this example. Now, the NumberOfBit
Eq. (7). value could be calculated for Ab1 . Ceil function rounds its
parameter to the nearest integers greater than or equal to its

α = (1/ρ) exp − f ∗ (7) parameter (NumberOfBit = ceil(8 × 0.3 × 1) = ceil(2,4) = 3).
Now three bits (for insertion) or there bit pairs (for swap) are
where ρ is a constant parameter that controls the decay of selected randomly of each colon of Ab1 . It means insertion
the inverse exponential function and f ∗ is the fitness of c or swap procedure is repeated three times.

Table 11 Hypermutation process of clones of Ab1

Antibodies Rand / Method Selected bits Mutated Clones
Ab1 - Clone1 = 01011101 0,235 / Insertion 5; 6; 7 01010011
Ab1 - Clone2 = 01011101 0,672 / Swap (1,6); (7,8); (2,3) 10111010
Ab1 - Clone3 = 01011101 0,03 / Insertion 1; 4; 8 11001100

Table 12 Mutated forms of clones

Clones S1 and S2 Mutated forms of S1 and S2
S1 = {(1 4 2),(6 5 9),(2 4 5)} S1 = {(1 4 2),(6 5 9),(2 2 5),(9 1 3)}
Clone 1
S2 = {(7 3 1),(4 8 6),(2 2 5),(9 1 3),(8 1 6)} S2 = {(7 3 1),(4 8 6),(2 4 5),(8 1 6)}
S1 = {(1 4 2),(6 5 9),(2 4 5)} S1 = {(7 3 1),(9 1 3),(8 1 6)}
Clone 2
S2 = {(7 3 1),(4 8 6),(2 2 5),(9 1 3),(8 1 6)} S2 = {(1 4 2),(4 8 6),(2 2 5),(6 5 9),(2 4 5)}
S1 = {(1 4 2),(6 5 9),(2 4 5)} S1 = S1 = {(4 8 6),(6 5 9),(2 4 5),(8 1 6)}
Clone 3
S2 = {(7 3 1),(4 8 6),(2 2 5),(9 1 3),(8 1 6)} S2 = {(1 4 2),(7 3 1),(2 2 5),(9 1 3)}

123
7508 Arabian Journal for Science and Engineering (2018) 43:7499–7520

Table 13 Sum of coordinates and affinity values mutated clones Table 14 An example initial solution of SA
Clones Sum of dimensions Affinity Initial solution Sum of coordinates Fitness value

Clone 1 {(18,14,19); (21,14,18)} max(3,0,1) = 3 2573|1468 {(21,18,23),(18,10,14)} 9

Clone 2 {(24,5,10); (15,21,27)} max(9,16,17) = 17
Clone 3 {(20, 18, 26); (19,10,11)} max(1,8,7) = 8
Bold value refers to the best individual solution the system to stick at a local optimum value and to search
other regions of the solution space.
In this study, dual local search is used to improve the per-
Three clones of the Ab1 is presented in column 1 of formance of SA for solving MDTWNPP. In the first local
Table 11. The bold or colored bits are selected to muta- search, insertion method is used to balance the number of vec-
tion. According to random generated number, the insertion or tors in the subsets. In the second local search, swap method
swap operator is chosen to mutate the antibody. If the number is used to place the vectors in the correct subset. SA is a good
is smaller than 0.3, then insertion method else swap method algorithm for global search. With this modification, SA has
is applied. This value is selected experimentally. According been able to make a better local search for MDTWNPP and its
to insertion method, the selected bit value has been changed. performance is increased. The proposed SA consists of three
If its value is 0 then it is turned to 1 or just the opposite is steps. These are: generating initial solution, implementing
performed. According to swap method, selected bit pairs are dual local search and determining to accept a worse solution
replaced with each other. The pair bits are shown in same as iteration best.
color. Randomly selected bits are given in column 3. At last,
the mutated clones are given column 4.
Affinity calculations of mutated clones are shown in detail 8.1 Generating Initial Solution of SA
in Table 12 where the first column gives the number of clones,
the second column indicates S1 and S2 subsets of the Ab1 and SA starts with an initial solution and tries to improve this
the third column represents the mutated forms of S1 and S2 . solution with local search. An example representation of an
The selected bits are also bolded or colored like Table 11. The initial solution for SA is given in Table 14. The problem set
sum of coordinates and affinity values are given in Table 13. and representation of the solution is like Table 1.
As shown in Table 13, all clones affinities are better than
the original Ab. The clone having the lowest affinity value is
selected to replace the original Ab. For this example, Clone 1 8.2 Implementing Dual Local Search of SA
is selected. In CSA, the cloning, hyper mutation and selecting
the clone with the best affinity processes are repeated until In the first local search, first a random subset is selected and
an iteration number is met. then a vector is selected randomly from the selected subset.
The selected vector is removed from the selected subset and
inserted to the other subset. This local search continues until
M neighbors are evaluated. If a better solution than the cur-
8 Simulated Annealing with Dual Local rent best iteration solution is founded it is accepted as the best
Search iteration solution. An example of first local search is shown
in Table 15.
SA is one of the most common metaheuristics used in the The second local search begins with the best solution
optimization literature. Basic principles of SA are based on obtained from the first local search. The aim of this local
the foundations of Metropolis et al. [33] where the anneal- search is to replace the vectors in the optimum subset. In
ing process is discussed. Kirkpatrick et al. [34] used these this local search, one vector is randomly selected from each
foundations to model and solve big-sized optimization prob- subset and the selected vectors are mutually replaced. An
lems in reasonable times without a guarantee of finding the example of the second local search is shown in Table 16.
global optimal solution. The basic parameters of SA are start-
ing temperature TS , cooling plan CP , and number of neighbor
solution M. SA starts with an initial random solution and a Table 15 An example of the first local search of SA
starting temperature TS , to solve the problem and the neigh- Solution/ Randomly Randomly Neighbor
borhood of this solution is searched for better solutions until fitness value selected sub- selected vec- solution/
a termination criteria (end temperature TE ) is met. A differ- set tor fitness value
ence of SA from other methods is that SA may accept a bad 2573|1468/9 S1 3 257|14683/1
solution with a probability at every iteration in order to avoid

123
Arabian Journal for Science and Engineering (2018) 43:7499–7520 7509

Table 16 An example of the second local search of SA proportion of current population size to be added for diver-
Solution/ Randomly Randomly Neighbor sity = 0.10, affinity proportionate mutation parameter ρ = 0.5.
fitness value selected vec- selected vec- solution/ SA is initialized with parameters Ts = 500000, C p = 0.99
tor from S1 tor from S1 fitness value and M = set size of the problem.
The experimental results achieved are shown at Tables 17,
2573|1468/9 2 1 1573|2468/7
18, 19, 20, 21, 22 and 23, respectively. The first column of
the tables shows the instances, the second column gives the
best results found with CPLEX and the metaheuristics, the
8.3 Determining to Accept a Worse Solution as third column represents the average results of 20 individual
Iteration Best tests found with the metaheuristics and on the fourth column
the standard deviation of the individual tests are given for the
One of the issues that affect the performance of a metaheuris- metaheuristics.
tic on a particular problem is way how often the solution is For the experiments with set size 50 in Table 17, when the
trapped into local optimum. The methodology used in SA to best results are considered GGA, SA and MFMBO obtained
avoid premature convergence is to accept a worse solution the best results for 7, 9 and 3 datasets, respectively. If the aver-
than the current iteration best solution with a probability that age results are analyzed it may be seen that GGA, SA yielded
is an exponential function of the difference between the qual- the best average result for 7 and 11 datasets, respectively.
ity of solutions and the current temperature at that iteration. The number of datasets GGA and SA yields the best devia-
tion values are 8 and 10, respectively. As a general evaluation
of this experiment one may figure out that the performance
9 Computational Results of SA increases proportionally with the problem complexity.
For the experiments with set size 100 in Table 18, when
GGA is implemented using C# Microsoft Visual Studio the best results are considered GGA, SA obtained the best
Ultimate 2013, MFMBO is coded using C++ Qt CSA is devel- results for 5, 13 datasets, respectively. If the average results
oped using MATLAB 2012. All experiments are carried out are analyzed it may be seen that GGA, SA yielded the best
using a computer with an Intel Core i7 [email protected] Ghz average result for 8 and 10 datasets, respectively. The devi-
processor with 16 GBs of memory operated on Microsoft ations of GGA and SA are the best for 11 and 8 datasets,
Windows 8 system. In order to evaluate the performance respectively. The analysis of these experiments show that SA
of three metaheuristics and ILP model, authors generated utilized the best performance for the best results, average
126 datasets with different coordinates, sizes because of results and deviations of the experimental results. Also it is
there is not a common benchmarking dataset for MDTWNNP clear that the success of SA increases with the dimension of
[35]. All random numbers generated are allowed to have the problem which is a desired aspect of a metaheuristic.
three decimal digits. The set sizes for the instances are If the experiments for set size 200 in Table 19 are inves-
determined to be {50, 100, 200, 300, 400, 500, 1000}. The tigated GGA, SA and MFMBO obtained the best results for
number of coordinates for every set size is considered as 6, 11 and 1 datasets, respectively. When the average results
{2, 3, 5, 10, 15, 20}. Three individual instances are gen- are analyzed it may be seen that GGA, SA yielded the best
erated for every set size and for every number of coor- average result for 7 and 11 datasets, respectively. The devi-
dinates. Every instance is named as set size_# of coordi- ations of GGA, SA and MFMBO are the best for 8, 9 and
nates_instance(a,b,c). For example 100_3_b denotes to the 1 datasets, respectively. Resembling the same results with
instance with set size 100, number of coordinates 3 and the previous experiments SA performs better results for the
instance order b. The stopping criterion for all metaheuristic big-sized datasets than other metaheuristics.
methods is considered to be 900 s and all metaheuristics are The results for the experiments with set size 300 are given
run for 20 individual tests for all instances. On the other hand, in Table 20. When the best results are analyzed it can be
CPLEX solver is run once with 1800 s stopping criterion for seen that GGA, SA and MFMBO yielded the best results for
all instances. 9, 7 and 2 datasets, respectively. If the average results are
Once the experiments begin, the parameter optimization is considered, GGA and SA delivers the best values for 8 and
done for all metaheuristics. GGA is implemented with param- 10 datasets, respectively. From another point of view, if the
eters number of population = 1000, crossover ratio = 0.9 and deviations are taken into account GGA and SA produces best
mutation ratio = 0.9. MFMBO is executed with parameters values for 8 and 10 datasets.
number of birds = 25, number of neighbor = 11, number According to the experimental results with set size 400
of flap = 20, number of neighbor share = 1 and number of given at Table 21, GGA and SA showed a performance to
flocks = 500. CSA is run with parameters number of ini- give the best results for both 9 datasets. When the average
tial network cells = 1000, multiplying factor β = 0.003, results are considered, GGA and SA gave the best average

123
 7510

123
Table 17 Results for instances with set size 50
Instances Best Average Deviation
CPLEX GGA SA MFMBO CSA GGA SA MFMBO CSA GGA SA MFMBO CSA

50_2_a 0.820 0.04 0.030 0.774 0.174 0.083 0.142 3.109 0.755 0.040 0.065 1.524 0.539
50_2_b 0.366 0.028 0.048 1.182 0.358 0.089 0.120 3.494 0.935 0.042 0.058 1.431 0.407
50_2_c 0.299 0.063 0.015 1.531 0.257 0.119 0.120 2.710 0.868 0.042 0.075 0.973 0.382
50_3_a 15.921 1.562 2.969 40.953 6.052 5.886 5.705 62.923 17.728 2.204 1.965 12.935 7.696
50_3_b 14.177 1.846 2.511 8.454 3.927 4.250 5.608 36.119 17.879 1.396 2.370 15.112 7.408
50_3_c 16.864 2.494 2.969 37.700 7.253 4.908 5.714 48.533 17.269 1.541 1.960 13.117 7.708
50_5_a 126.771 88.043 57.876 152.863 168.076 106.631 106.230 310.918 294.283 15.221 21.342 91.422 59.899
50_5_b 202.924 41.349 74.915 346.287 155.677 109.579 110.962 423.717 257.172 33.349 25.545 67.259 43.048
50_5_c 162.646 90.692 100.086 263.500 96.502 118.529 136.185 414.745 274.373 16.907 21.491 93.571 72.984
50_10_a 1563.681 1175.304 1189.547 1716.535 1781.765 1559.859 1402.989 2349.671 2197.029 161.044 122.285 264.843 215.328
50_10_b 1726.904 1335.264 1043.625 1779.232 1854.323 1450.463 1285.027 2216.981 2108.441 87.090 128.685 222.599 186.557
50_10_c 1335.882 1328.598 1289.195 1897.140 1719.052 1561.888 1396.791 2215.245 2164.283 146.778 107.388 268.481 204.430
50_15_a 3018.781 3267.319 3018.781 4070.789 4009.632 3828.420 3177.846 4375.418 4697.607 358.882 131.138 178.525 340.135
50_15_b 3067.279 3018.903 2927.760 3067.279 3841.900 3675.191 3096.582 4280.040 4467.522 462.160 163.820 483.683 355.981
50_15_c 3302.693 3362.359 3153.811 3866.280 3475.641 3874.340 3233.442 4246.718 4527.269 288.676 80.972 210.637 562.891
50_20_a 5646.702 5656.961 5646.702 5646.702 5885.587 6293.094 5646.702 6374.412 6864.552 334.084 0.000 424.952 351.592
50_20_b 5557.627 5658.978 5557.627 5327.761 6262.376 6111.578 5557.627 6111.689 6901.479 337.234 0.000 425.776 424.048
50_20_c 5169.570 5258.486 5169.570 4412.499 6067.112 6148.508 5169.570 6101.246 6780.829 421.653 0.000 652.320 366.852
Bold values refer to the best individual solution
Arabian Journal for Science and Engineering (2018) 43:7499–7520
 Table 18 Results for instances with set size 100
Instances Best Average Deviation
CPLEX GGA SA MFMBO CSA GGA SA MFMBO CSA GGA SA MFMBO CSA

100_2_a 0.587 0.037 0.019 0.333 0.427 0.075 0.065 1.676 0.739 0.04 0.054 0.635 0.223
100_2_b 0.313 0.014 0.006 0.412 0.140 0.072 0.074 1.468 0.604 0.044 0.044 0.607 0.265
100_2_c 0.363 0.025 0.007 0.115 0.117 0.065 0.084 1.553 0.611 0.029 0.064 0.751 0.321
100_3_a 15.081 1.387 2.726 13.024 8.825 3.196 6.121 27.938 15.807 1.777 2.210 8.762 6.253
100_3_b 26.616 3.100 1.943 9.278 10.070 4.534 4.549 31.884 19.069 1.286 1.971 14.404 4.793
Arabian Journal for Science and Engineering (2018) 43:7499–7520

100_3_c 15.249 1.465 1.529 8.643 5.337 3.972 5.497 24.736 13.537 1.271 1.994 6.769 5.245
100_5_a 216.272 76.953 68.980 205.425 173.113 110.406 121.979 288.620 286.686 19.869 28.956 61.830 51.062
100_5_b 259.931 68.427 110.870 173.435 218.905 107.048 143.105 243.510 281.498 23.96 24.191 54.477 36.502
100_5_c 231.023 72.469 70.895 164.683 157.205 127.017 141.288 284.293 252.281 22.258 27.926 58.171 68.052
100_10_a 1413.181 1131.413 1075.446 1487.141 1406.129 1350.747 1281.782 1847.311 2094.149 116.368 134.345 257.155 338.703
100_10_b 1443.46 952.433 915.827 1203.163 1799.46 1314.531 1274.031 1653.701 2170.197 206.230 187.591 192.519 268.015
100_10_c 1449.052 964.268 1192.830 1007.107 1687.306 1363.998 1339.027 1767.388 2249.354 190.887 107.011 311.158 211.084
100_15_a 3304.117 3135.350 2645.627 3489.455 4350.489 3525.495 3002.132 3837.825 4920.642 219.090 189.724 270.238 275.03
100_15_b 3056.405 2889.974 2769.881 3225.045 4080.905 3319.477 2981.924 3693.849 4837.997 268.254 156.524 322.547 378.503
100_15_c 3181.991 2294.433 2415.282 3226.187 3906.193 3108.409 2898.782 3663.094 4750.758 324.855 194.296 242.754 366.305
100_20_a 4943.964 4555.801 4476.502 5291.334 6569.404 5168.517 4715.180 5551.415 7485.181 381.128 164.828 234.033 515.355
100_20_b 4613.200 4422.483 4057.718 5130.222 6375.971 5234.149 4478.760 5448.002 7118.519 357.971 252.971 253.129 397.188
100_20_c 4350.876 4669.931 3222.774 4286.002 6360.697 5235.724 4483.664 5307.752 7259.757 390.247 472.052 487.332 401.286
Bold values refer to the best individual solution

123
7511
 7512

123
Table 19 Results for instances with set size 200
Instances Best Average Deviation
CPLEX GGA SA MFMBO CSA GGA SA MFMBO CSA GGA SA MFMBO CSA

200_2_a 0.480 0.009 0.052 0.260 0.180 0.064 0.103 0.866 0.528 0.040 0.035 0.378 0.232
200_2_b 0.822 0.035 0.048 0.154 0.242 0.080 0.111 0.604 0.730 0.033 0.064 0.215 0.283
200_2_c 0.636 0.020 0.038 0.018 0.146 0.085 0.099 0.444 0.507 0.045 0.037 0.400 0.257
200_3_a 20.144 2.435 2.412 6.174 8.740 5.177 5.131 16.079 15.791 1.258 2.240 6.826 4.506
200_3_b 34.713 2.627 1.247 7.795 8.289 3.862 3.803 14.499 16.404 0.940 1.970 5.677 6.268
200_3_c 16.747 2.389 2.025 4.067 7.373 4.119 4.222 10.973 14.367 1.125 1.164 3.474 3.819
200_5_a 290.317 73.723 81.420 155.739 190.301 117.922 130.235 213.371 274.357 24.034 26.393 35.106 55.251
200_5_b 177.133 75.512 72.314 104.960 116.945 129.478 113.136 219.337 255.533 27.952 26.579 63.560 74.538
200_5_c 176.920 72.133 38.638 138.623 129.920 106.066 99.922 202.865 221.832 18.455 32.993 50.171 54.094
200_10_a 1562.503 1049.530 1134.800 1396.105 1747.567 1355.753 1344.711 1707.080 2185.537 185.688 140.876 192.664 203.428
200_10_b 1405.135 956.929 1147.554 1278.109 1649.380 1269.061 1420.278 1574.231 2189.357 148.555 116.866 137.573 285.986
200_10_c 1594.059 1094.773 1152.489 1490.464 1531.827 1330.498 1393.134 1691.590 2280.598 114.488 107.167 148.408 304.039
200_15_a 3223.486 2718.032 2698.994 3370.608 3341.696 3060.899 2974.254 3759.343 5030.571 220.052 169.189 203.088 766.090
200_15_b 3271.084 3014.508 2192.314 3125.574 4778.486 3222.296 2905.825 3552.118 5361.366 99.270 324.392 259.087 340.262
200_15_c 2933.267 2481.017 2146.385 3100.857 3924.787 2983.037 2885.117 3358.357 5180.018 228.175 326.139 203.264 488.383
200_20_a 4528.352 4458.439 4106.615 4669.363 7272.377 4964.359 4488.545 5177.919 8199.702 334.449 187.171 224.695 462.637
200_20_b 4953.884 4268.625 3723.423 4424.274 7072.755 4858.343 4503.077 4969.566 8341.041 262.750 367.228 299.941 554.198
200_20_c 5073.884 4365.803 4232.243 4405.410 7142.044 4881.211 4713.758 5201.637 7984.126 271.320 232.552 384.918 529.182
Bold values refer to the best individual solution
Arabian Journal for Science and Engineering (2018) 43:7499–7520
 Table 20 Results for instances with set size 300
Instances Best Average Deviation
CPLEX GGA SA MFMBO CSA GGA SA MFMBO CSA GGA SA MFMBO CSA

300_2_a 1.003 0.031 0.056 0.194 0.379 0.087 0.073 0.513 0.710 0.039 0.014 0.269 0.237
300_2_b 1.061 0.015 0.011 0.139 0.059 0.053 0.065 0.852 0.474 0.022 0.029 0.501 0.205
300_2_c 0.491 0.009 0.022 0.143 0.109 0.062 0.067 0.464 0.448 0.034 0.036 0.215 0.272
300_3_a 13.359 1.299 1.277 3.001 6.119 3.977 4.782 12.288 15.689 1.910 1.794 5.208 7.127
300_3_b 15.397 1.958 0.914 2.631 11.963 5.071 4.192 12.136 15.231 1.605 1.670 4.731 2.952
Arabian Journal for Science and Engineering (2018) 43:7499–7520

300_3_c 22.108 1.511 1.726 6.239 8.422 4.222 4.015 12.264 18.391 1.747 1.300 4.534 6.832
300_5_a 235.993 78.249 45.785 151.617 177.573 116.933 126.190 235.718 255.244 23.161 37.405 39.836 45.959
300_5_b 198.827 58.514 67.484 115.874 142.544 120.156 112.127 190.059 235.162 34.000 22.209 43.950 52.772
300_5_c 201.081 71.507 87.465 119.539 157.270 120.828 123.832 185.349 239.455 22.649 17.491 42.596 50.626
300_10_a 1600.501 1161.359 1225.771 1516.081 1493.839 1366.332 1375.611 1734.594 2441.357 134.885 126.850 150.839 355.561
300_10_b 1211.643 1017.839 1023.663 952.016 1719.849 1281.413 1275.348 1531.982 2261.841 153.772 154.373 301.352 225.055
300_10_c 1439.701 907.481 1147.318 1099.552 1593.338 1250.915 1352.103 1432.054 2239.176 159.381 120.349 187.945 266.652
300_15_a 2833.666 2575.428 2472.876 2563.757 5184.172 3080.485 2978.084 3453.729 5559.460 248.674 312.260 377.207 377.066
300_15_b 3066.962 2635.511 1939.744 2854.334 4608.820 3056.794 2940.795 3444.819 5294.132 223.088 411.033 255.883 406.560
300_15_c 3219.413 2628.103 2500.951 2493.521 4795.620 2991.105 3090.022 3237.849 5603.266 246.297 242.304 382.799 362.760
300_20_a 4520.479 3724.087 3781.884 4266.824 8508.286 4577.688 4349.201 5090.336 8930.889 378.082 248.726 350.587 312.256
300_20_b 5377.245 3638.537 4097.179 4826.285 7176.819 4516.910 4464.700 5291.548 8801.875 402.767 223.338 293.471 726.131
300_20_c 4629.251 4061.494 4049.986 4349.067 7271.521 4624.728 4411.991 5000.837 8732.549 309.705 362.005 361.590 598.006
Bold values refer to the best individual solution

123
7513
 7514

123
Table 21 Results for instances with set size 400
Instances Best Average Deviation
CPLEX GGA SA MFMBO CSA GGA SA MFMBO CSA GGA SA MFMBO CSA

400_2_a 0.763 0.039 0.027 0.050 0.294 0.076 0.110 0.360 0.605 0.028 0.041 0.210 0.268
400_2_b 0.921 0.007 0.019 0.103 0.077 0.078 0.089 0.488 0.480 0.045 0.050 0.224 0.286
400_2_c 0.784 0.004 0.009 0.272 0.216 0.068 0.105 0.548 0.541 0.034 0.047 0.244 0.216
400_3_a 26.286 2.105 3.800 4.014 13.106 5.358 5.545 10.608 18.255 1.765 1.050 3.545 3.995
400_3_b 14.376 1.014 1.700 5.914 6.123 3.989 4.245 10.891 12.540 1.578 1.445 3.366 4.821
400_3_c 11.917 2.689 1.450 4.762 6.024 4.865 4.797 12.452 14.790 1.301 1.774 5.830 6.036
400_5_a 200.652 64.931 84.194 111.214 165.269 106.332 122.440 204.165 273.568 23.765 17.693 59.031 45.593
400_5_b 202.975 88.465 50.490 149.553 157.551 111.402 112.144 202.122 243.298 10.494 28.529 38.077 47.712
400_5_c 177.156 77.054 43.088 177.373 161.029 115.654 111.586 220.220 265.589 20.209 27.195 21.591 42.755
400_10_a 1525.539 960.587 1155.768 1270.186 1895.196 1170.272 1390.221 1605.951 2384.695 128.224 145.996 140.717 303.974
400_10_b 1544.819 1218.227 819.333 1242.115 2088.952 1334.915 1239.794 1632.292 2436.273 70.232 260.641 194.322 195.763
400_10_c 1475.285 955.831 1118.181 1163.416 1925.563 1243.658 1368.804 1534.602 2364.944 197.461 135.487 212.394 217.582
400_15_a 3078.730 2809.403 2196.625 3202.180 5084.131 3080.064 2944.260 3438.418 5760.112 154.868 280.543 133.350 343.453
400_15_b 2976.724 2479.535 2646.329 3113.688 5086.257 3011.878 3017.999 3474.726 5718.315 259.147 216.999 267.444 288.397
400_15_c 3013.114 2847.162 1877.136 3092.129 4162.375 3151.915 2920.405 3670.391 5545.030 123.529 439.518 229.803 549.167
400_20_a 5064.953 4017.978 4412.946 4985.473 8837.862 4668.974 4687.353 5470.120 9696.369 399.797 205.480 286.947 486.761
400_20_b 4982.146 4051.608 4011.380 4492.011 8588.503 4451.193 4528.142 5167.287 9646.422 335.804 267.302 389.448 440.852
400_20_c 4822.463 4220.332 4171.581 4559.854 8463.921 4610.120 4624.602 5246.735 9117.066 247.387 267.877 363.060 444.856
Bold values refer to the best individual solution
Arabian Journal for Science and Engineering (2018) 43:7499–7520
Arabian Journal for Science and Engineering (2018) 43:7499–7520 7515

results for 13 and 5 datasets. Finally, if the deviations of the greedy genetic operations that provides local search more
results are investigated, GGA, SA and MFMBO obtained the effectively. SA searches the global space effectively since
best deviations for 10, 7 and 1 datasets. worse solutions may be accepted with a probability
As it is shown at Table 22, GGA, SA and MFMBO per- which avoids the algorithm from premature solutions.
formed best results for 8, 9 and 1 datasets, respectively. For solving MDTWNNP, the local search performance
Also GA and SA produced best average results for 10 and 8 of SA is developed using a dual local search. Accord-
datasets, respectively. Lastly, GGA, SA and MFMBO yielded ing to the whole experimental results, CPLEX, GGA,
best deviations for 10, 7 and 1 datasets. In this experiment, SA and MFMBO yielded best results for 2, 56, 65, 7
SA and GA showed a competitive performance. datasets, respectively. Considering the average results
The last experimental results are carried out for set size of 20 individual runs of the metaheuristics, GGA and
1000 which is the biggest size problem. As it is shown at SA apparently yields the best averages for 65 and 61
Table 23, GGA and SA yielded best solutions for 11 and dataset, respectively. From the standard deviations point
7 datasets, respectively. Besides GA and SA produced best of view which emphasizes the robustness and consistency
average results for 12 and 6 datasets, respectively. Finally, of the metaheuristics, again GGA and SA show better
GGA and SA produced best deviations for 10, 7 and 1 datasets. performance than the other metaheuristics with smallest
According to the whole experimental results, GGA and standard deviation value for 60 and 63 out of 126 datasets,
SA showed better performance than MFMBO and CSA for respectively. Also, it may be seen at Fig. 4 that GGA and
MDTWNNP. Also, if the performance of GGA and SA is SA have the best rankings for three number of coordinates
compared it may be seen these two metaheuristics are in each.
competition for all datasets. In order to demonstrate this com- • The experimental results show that, MFMBO is the third
petition, the ranking sequences of the best results, average successful metaheuristic for the MDTWNNP. MFMBO
results and standard deviations for all coordinates are given obtained better results than CPLEX, GGA, SA and CSA
in Figs. 4, 5 and 6, respectively. for 94, 8, 6 and 91 problem instances, respectively.
It may be seen from Fig. 4 that when the best result rank- Considering the average results, it may be suggested
ings are considered GGA has the best ranking for number of that MFMBO shows better performance for problem
coordinates 3, 5 and 10 while SA has the best ranking for instances with coordinates more than 3. It may be seen in
15 and 20. This may be evaluated as the performance of SA Fig. 5 that the standard deviation of MFMBO mutually
increases proportionally to the number of coordinates which proportional to the coordinate of the problem. As a result
directly refers to the complexity of the problem. of this, when the search space of the problem decreases
According to the ranking of the average results for all coor- the MFMBO can get stuck in the local minimum.
dinates given at Fig. 5, GGA and SA has the smallest ranking • In spite of the fact that CSA shows the worst performance
for 3 different coordinate values each. But, the performance among the metaheuristics for 64 problem instances espe-
of SA increases proportionally to the number of coordinates. cially for those with lower coordinates, CSA yields better
The robustness of a metaheuristic algorithm may be con- results than CPLEX. When the averages and standard
sidered with the deviations of the results. When the results deviations are investigated, it may be pointed out that
in Fig. 6 are analyzed, GGA showed the best performance CSA is more successful for instances with lower coor-
for number of coordinates 2, 3 and 5 besides SA showed the dinates and its success is mutually proportional to the
best performance for the number of coordinates 10, 15 and complexity of the problem.
20. These results may be interpreted as SA and GGA are the
most robust and successful metaheuristics for MDTWNNP. In order to show the validity and statistical significance of
From another point of view if the performance of GGA and the problem results, the obtained results are used to carry out
SA is investigated one may see that the performance of SA analysis of variance (ANOVA) statistical tests for all coor-
increases proportionally to the number of coordinates for dinates of the problem for α = 0.05 [36]. According to the
ranking analysis which is a desirable aspect of a metaheuris- theory of ANOVA, the greater the F value is the better the sig-
tic. nificance is visible [37]. The results of the one way ANOVA
As a conclusion of the whole experimental results one may statistical tests for all coordinates show that there is a statis-
figure out the following inferences: tical significance among the experimental results.

• GA’s global search ability is strong for solving discrete 10 Conclusions

problems but its ability is defected by premature and
weak local search ability. For solving this an improved In this study, the MDTWNPP which is a generalization of
metaheuristic approach GGA is proposed which uses TWNPP and which has not yet been studied widely in the

123
 7516

123
Table 22 Results for instances with set size 500
Instances Best Average Deviation
CPLEX GGA SA MFMBO CSA GGA SA MFMBO CSA GGA SA MFMBO CSA

500_2_a 0.246 0.051 0.023 0.260 0.111 0.105 0.086 0.648 0.357 0.041 0.038 0.187 0.180
500_2_b 0.365 0.024 0.032 0.113 0.106 0.054 0.083 0.533 0.345 0.033 0.045 0.219 0.164
500_2_c 0.511 0.021 0.020 0.055 0.306 0.073 0.080 0.523 0.604 0.039 0.042 0.322 0.176
500_3_a 16.520 2.148 2.286 5.038 4.030 3.900 4.729 11.751 14.920 1.442 1.283 5.483 6.661
500_3_b 34.473 2.011 2.183 4.363 5.647 4.817 5.236 14.311 14.164 1.619 1.834 6.840 7.311
500_3_c 16.613 1.419 1.826 5.014 4.865 3.594 4.501 13.554 15.841 1.361 1.409 5.122 5.882
500_5_a 211.984 45.161 100.914 88.436 197.070 101.046 131.330 205.179 270.955 23.782 20.444 48.193 50.291
500_5_b 171.979 78.032 64.597 105.056 193.017 111.668 99.875 219.367 235.349 17.529 17.909 55.277 21.103
500_5_c 370.073 75.371 61.968 127.594 174.362 108.289 111.348 202.065 283.176 20.998 26.452 43.204 84.438
500_10_a 1465.149 1083.687 1115.696 956.925 2022.469 1310.163 1362.214 1663.466 2500.446 114.287 130.930 273.090 228.784
500_10_b 1326.769 914.213 1124.989 1468.970 1920.211 1351.332 1311.855 1690.494 2305.864 209.644 126.403 191.774 239.338
500_10_c 1200.920 1160.576 1132.072 1579.452 2122.029 1361.009 1327.527 1778.334 2569.673 102.173 104.971 162.444 227.854
500_15_a 3358.478 1846.034 2538.042 2657.453 4863.564 2910.141 3081.690 3318.080 5990.452 415.392 234.487 362.885 532.883
500_15_b 3230.867 2651.251 2375.215 2922.338 5636.049 2924.607 2796.782 3466.906 6217.006 175.126 285.250 410.178 327.431
500_15_c 2663.838 2689.847 2587.569 2782.071 5056.987 2997.357 2885.722 3464.731 5968.206 179.208 160.775 347.348 550.860
500_20_a 5071.810 4289.652 4164.257 4743.356 9578.465 4551.408 4549.765 5476.985 10111.590 140.727 219.137 374.152 401.899
500_20_b 4969.659 3968.144 4129.131 4750.797 8720.732 4599.623 4513.112 5383.943 9953.851 279.136 241.615 316.826 620.844
500_20_c 4948.700 4161.747 4078.767 4877.736 7943.481 4555.822 4648.106 5236.794 9749.920 248.908 304.356 245.652 798.700
Bold values refer to the best individual solution
Arabian Journal for Science and Engineering (2018) 43:7499–7520
 Table 23 Results for instances with set size 1000
Instances Best Average Deviation
CPLEX GGA SA MFMBO CSA GGA SA MFMBO CSA GGA SA MFMBO CSA

1000_2_a 0.359 0.019 0.017 0.233 0.195 0.078 0.080 0.864 0.401 0.036 0.036 0.388 0.114
1000_2_b 0.886 0.016 0.047 0.086 0.234 0.051 0.084 0.826 0.517 0.030 0.020 0.600 0.206
1000_2_c 0.824 0.040 0.040 0.636 0.101 0.093 0.079 1.150 0.369 0.034 0.033 0.411 0.170
1000_3_a 9.059 1.835 2.331 3.443 6.182 4.221 4.954 17.599 15.733 1.659 1.428 7.529 5.656
1000_3_b 18.066 1.187 2.398 5.621 5.078 3.477 4.157 18.383 14.902 1.479 0.892 6.363 6.209
Arabian Journal for Science and Engineering (2018) 43:7499–7520

1000_3_c 17.182 0.861 2.055 7.538 6.250 3.740 4.726 19.760 13.386 2.125 1.493 6.496 4.803
1000_5_a 256.460 76.756 36.633 123.782 149.372 109.980 108.118 263.440 243.047 28.693 31.799 80.366 68.297
1000_5_b 237.897 63.176 69.926 133.674 165.909 99.110 102.799 258.453 253.417 23.218 19.974 65.077 50.024
1000_5_c 298.445 68.388 71.879 212.803 184.674 110.227 116.638 266.157 236.641 24.024 25.097 32.991 48.992
1000_10_a 1596.560 912.229 1134.993 1505.398 1709.698 1259.331 1311.152 1959.973 2931.211 131.293 97.303 224.103 436.822
1000_10_b 1665.834 1100.642 1078.371 1308.367 2553.090 1283.285 1363.930 1738.728 2942.765 127.399 144.061 205.121 232.509
1000_10_c 1618.678 1009.872 1083.297 1401.823 2700.234 1222.703 1277.526 1924.847 3037.665 144.440 99.334 292.918 216.934
1000_15_a 3710.317 2707.085 2286.340 3563.963 5871.338 3057.931 2946.571 3994.004 7276.557 173.865 341.013 224.895 671.056
1000_15_b 2711.326 2614.125 2421.917 3463.570 6523.187 2910.773 2854.385 4140.494 7615.466 191.384 190.401 338.144 521.414
1000_15_c 3318.914 2115.464 2545.730 3305.721 6413.521 2884.633 2925.176 4067.102 7284.090 318.517 239.981 374.847 571.060
1000_20_a 5832.677 3975.607 4043.885 5114.241 11223.015 4474.750 4532.871 6233.983 12813.856 322.987 313.989 597.222 977.497
1000_20_b 4601.663 4526.584 4268.352 6093.989 9826.043 4701.921 4576.172 6567.718 12857.850 108.775 212.721 215.474 1350.696
1000_20_c 5538.170 4107.453 4130.397 5136.738 11164.659 4508.219 4501.347 6549.728 13383.304 243.431 154.515 562.735 1253.081
Bold values refer to the best individual solution

123
7517
7518 Arabian Journal for Science and Engineering (2018) 43:7499–7520

0
2 Coordinates 3 Coordinates 5 Coordinates 10 Coordinates 15 Coordinates 20 Coordinates
GGA SA MFMBO CPLEX CSA

Fig. 4 The best results ranking sequence of CPLEX and the metaheuristics for all coordinates

0
2 Coordinates 3 Coordinates 5 Coordinates 10 Coordinates 15 Coordinates 20 Coordinates
GGA SA MFMBO CSA

Fig. 5 The average results ranking sequence of the metaheuristics for all coordinates

0
2 Coordinates 3 Coordinates 5 Coordinates 10 Coordinates 15 Coordinates 20 Coordinates

GGA SA MFMBO CSA

Fig. 6 The standard deviation ranking sequence of the metaheuristics for all coordinates

123
Arabian Journal for Science and Engineering (2018) 43:7499–7520 7519

literature is considered. The main objective of the paper is to 4. Kojić, J.: Integer linear programming model for multidimensional
analyze the performances of four metaheuristic approaches two-way number partitioning problem. Comput. Math. Appl. 60(8),
2302–2308 (2010)
GGA, SA, MFMBO and CSA on this problem. First, 126 5. Rodriguez, F.J.; Glover, F.; García-Martínez, C.; Martí, R.; Lozano,
problem instances with different set size and number of coor- M.: GRASP with exterior path-relinking and restricted local search
dinates parameters are generated randomly. These problems for the multidimensional two-way number partitioning problem.
are solved using the integer linear model presented at [4] Comput. Oper. Res. 78, 243–254 (2017)
6. Bauke, H.; Franz, S.; Mertens, S.: Number partitioning as a random
using CPLEX solver with a time limit of 1800 s. The best energy model. J. Stat. Mech. Theory Exp. 2004(04), P04003 (2004)
results found with CPLEX in this time limit are recorded. 7. Dell’Amico, M.; Iori, M.; Martello, S.; Monaci, M.: Heuristic and
Then, GGA, SA, MFMBO and CSA algorithms are consid- exact algorithms for the identical parallel machine scheduling prob-
ered and some contributions are made for these algorithms lem. INFORMS J. Comput. 20(3), 333–344 (2008)
8. Merkle, R.; Hellman, M.: Hiding information and signatures in
in order to achieve best performance for MDTWNPP. All
trapdoor knapsacks. IEEE Trans. Inf. Theory 24(5), 525–530
instances are solved for 900 s with these four metaheuris- (1978)
tics for 20 times and the best results, the average of the 9. Hayes, B.: Computing science: the easiest hard problem. Am. Sci.
results and standard deviations of the results are recorded. 90(2), 113–117 (2002)
10. Horowitz, E.; Sahni, S.: Computing partitions with applications to
The main performance criteria taken into account is the num-
the knapsack problem. J. ACM (JACM) 21(2), 277–292 (1974)
ber of instances the metaheuristics outperform the best result 11. Korf, R.E.: A complete anytime algorithm for number partitioning.
found by CPLEX. Here, it is found out that GGA and SA have Artif. Intell. 106(2), 181–203 (1998)
achieved better results with outperforming CPLEX, MFMBO 12. Karmarkar, N.; Karp, R.M.: The Difierencing Method of Set Par-
titioning. Technical Report UCB/CSD 82/113, Computer Science
and CSA. When the performance of GGA and SA is compared Division, University of California, Berkeley (1982)
to each other, both algorithms obtained approximately same 13. Pedroso, J.P.; Kubo, M.: Heuristics and exact methods for number
number of best results. If a more detailed analysis is carried partitioning. Eur. J. Oper. Res. 202(1), 73–81 (2010)
out between two metaheuristics, GGA shows better perfor- 14. Alidaee, B.; Glover, F.; Kochenberger, G.A.; Rego, C.: A new mod-
eling and solution approach for the number partitioning problem.
mance than SA for datasets with the number of coordinates Adv. Decis. Sci. 2005(2), 113–121 (2005)
2, 3 and 5 and SA shows better performance than SA for 15. Berretta, R.; Cotta, C.; Moscato, P.: Enhancing the performance
datasets with the number of coordinates 10, 15 and 20. Con- of memetic algorithms by using a matching-based recombination
sidering the whole set of instances, MFMBO and CSA has algorithm. In: Resende, M., Pinho de Sousa, J. (eds.) Metaheuris-
tics: Computer Decision-Making, pp. 65–90. Springer, Boston
outperformed CPLEX for 94 and 64 instances, respectively. (2003)
Finally, ANOVA statistical tests for each coordinate values are 16. Johnson, D.S.; Aragon, C.R.; McGeoch, L.A.; Schevon, C.: Opti-
carried out to show that the results between all metaheuristics mization by simulated annealing: an experimental evaluation; part
significant difference. II, graph coloring and number partitioning. Oper. Res. 39(3), 378–
406 (1991)
Hybrid algorithms in the literature are generally proposed 17. Koyutürk, M.: Hypergraph Based Declustering for Multi-disk
for improving the performance and solution quality on the Databases. Doctoral dissertation, Bilkent University (2000)
specific problem. As a future study, the datasets of this study 18. Argüello, M.F.; Feo, T.A.; Goldschmidt, O.: Randomized methods
will be solved using a hybrid algorithm. for the number partitioning problem. Comput. Oper. Res. 23(2),
103–111 (1996)
19. Fuksz, L.; Pop, P.; Zelina, I.: Heuristic algorithms for solving the
Compliance with ethical standards bi-dimensional two-way number partitioning problem. Stud. Univ.
Babes-Bolyai Ser. Inform. LVIII 3, 17–28 (2013)
20. Pop, P.C.; Matei, O.: A memetic algorithm approach for solving the
multidimensional multi-way number partitioning problem. Appl.
Conflict of interest: The authors declare that they have no conflict of Math. Model. 37(22), 9191–9202 (2013)
interest. 21. Kratica, J.; Kojić, J.; Savić, A.: Two metaheuristic approaches for
solving multidimensional two-way number partitioning problem.
Comput. Oper. Res. 46, 59–68 (2014)
22. Holland, J.H.: Hierarchical Descriptions, Universal Spaces and
Adaptive Systems (No. 08226-4-T). Michigan Univ Ann Arbor
Logic of Computers Group (1968)
References 23. Pham, D.T.; Karaboga, D.: Intelligent Optimization Techniques.
Springer, London (2000)
1. Mertens, S.: The easiest hard problem: number partitioning. Com- 24. Sivanandam, S.; Deepa, S.: Introduction to Genetic Algorithms.
put. Complex. Stat. Phys. 125(2), 125–139 (2006) Springer, Berlin (2007)
2. Mertens, S.: Phase transition in the number partitioning problem. 25. Duman, E.; Uysal, M.; Alkaya, A.F.: Migrating birds optimization:
Phys. Rev. Lett. 81(20), 4281 (1998) a new metaheuristic approach and its performance on quadratic
3. Hacibeyoglu, M.; Tongur, V.; Alaykiran, K.: Solving the bi- assignment problem. Inf. Sci. 217, 65–77 (2012)
dimensional two-way number partitioning problem with heuristic 26. Duman, E.; Elikucuk, I.: Applying migrating birds optimization to
algorithms. In: 2014 IEEE 8th International Conference on Appli- credit card fraud detection. In: Pacific-Asia Conference on Knowl-
cation of Information and Communication Technologies (AICT), edge Discovery and Data Mining, pp. 416–427. Springer, Berlin,
pp. 1–5 (2014) Heidelberg (2013)

123
7520 Arabian Journal for Science and Engineering (2018) 43:7499–7520

27. Gao, K.Z.; Suganthan, P.N.; Chua, T.J.: An enhanced migrating 33. Metropolis, N.; Rosenbluth, A.W.; Rosenbluth, M.N.; Teller,
birds optimization algorithm for no-wait flow shop scheduling A.H.; Teller, E.: Equation of state calculations by fast computing
problem. In: 2013 IEEE Symposium on Computational Intelligence machines. J. Chem. Phys. 21(6), 1087–1092 (1953)
in Scheduling (SCIS), pp. 9–13. IEEE (2013) 34. Kirkpatrick, S.; Gelatt, C.D.; Vecchi, M.P.: Optimization by simu-
28. Pan, Q.K.; Dong, Y.: An improved migrating birds optimisation for lated annealing. Science 220(4598), 671–680 (1983)
a hybrid flowshop scheduling with total flowtime minimisation. Inf. 35. Mehmet, H.: Multidimensional two-way number partition-
Sci. 277, 643–655 (2014) ing dataset. https://konya.edu.tr/storage/files/people/MDTWNPP/
29. De Castro, L.N.; Timmis, J.: Artificial Immune Systems: A New dataset.rar (2017)
Computational Intelligence Approach. Springer, Berlin (2002) 36. Semwal, V.B.; Mondal, K.; Nandi, G.C.: Robust and accurate fea-
30. De Castro, L.N.; Von Zuben, F.J.: Learning and optimization using ture selection for humanoid push recovery and classification: deep
the clonal selection principle. IEEE Trans. Evol. Comput. 6(3), learning approach. Neural Comput. Appl. 28(3), 565–574 (2017)
239–251 (2002) 37. Semwal, V.B.; Singha, J.; Sharma, P.K.; Chauhan, A.; Behera, B.:
31. Campelo, F.; Guimarães, F.G.; Igarashi, H.; Ramírez, J.A.: A clonal An optimized feature selection technique based on incremental
selection algorithm for optimization in electromagnetics. IEEE feature analysis for bio-metric gait data classification. Multimed.
Trans. Magn. 41(5), 1736–1739 (2005) Tools Appl. 76, 1–19 (2016)
32. Cruz-Cortés, N.; Trejo-Pérez, D.; Coello, C.A.C.: Handling con-
straints in global optimization using an artificial immune system.
In: International Conference on Artificial Immune Systems, pp.
234–247. Springer, Berlin, Heidelberg (2005).

123