Applied Mathematical Modelling 37 (2013) 9191–9202

Contents lists available at SciVerse ScienceDirect

Applied Mathematical Modelling

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

A memetic algorithm approach for solving the

multidimensional multi-way number partitioning problem
Petrică C. Pop a,⇑, Oliviu Matei b
a
Technical University of Cluj-Napoca, North University Center Baia Mare, Department of Mathematics and Computer Science, Baia Mare, Romania
b
Technical University of Cluj-Napoca, North University Center Baia Mare, Department of Electrical Engineering, Baia Mare, Romania

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

Article history: In this paper, we describe a generalization of the multidimensional two-way number par-
Received 16 May 2012 titioning problem (MDTWNPP) where a set of vectors has to be partitioned into p sets
Received in revised form 23 October 2012 (parts) such that the sums per every coordinate should be exactly or approximately equal.
Accepted 29 March 2013
We will call this generalization the multidimensional multi-way number partitioning
Available online 7 May 2013
problem (MDMWNPP). Also, an efficient memetic algorithm (MA) heuristic is developed
to solve the multidimensional multi-way number partitioning problem obtained by com-
Keywords:
bining a genetic algorithm (GA) with a powerful local search (LS) procedure. The perfor-
Number partitioning
Genetic algorithms
mances of our memetic algorithm have been compared with the existing numerical
Local search results obtained by CPLEX based on an integer linear programming formulation of the
Memetic algorithm problem. The solution reveals that our proposed methodology performs very well in terms
Combinatorial optimization of both quality of the solutions obtained and the computational time compared with the
previous method of solving the multidimensional two-way number partitioning problem.
Ó 2013 Elsevier Inc. All rights reserved.

1. Introduction

Number partitioning problem is a classical, challenging and surprisingly difficult problem in combinatorial optimization
and it is defined as follows: given a set S of n integers, the two-way number partitioning problem (TWNPP) asks for a division
of S into two subsets such that the sums of number in each subset are as close as possible (equal or approximately equal).
Though the number partitioning problem is NP-complete (see [1]), there have been proposed heuristic algorithms that
solve the problem in many instances either optimally or approximately. This is one of the reasons for which the problem
has been called ’’The Easiest Hard Problem’’ by Hayes [2].
A variation of the number partitioning problem is the 3-partition problem, in which a set of numbers S must be parti-
tioned into triples such that the sums in each subset to be equal or approximately equal.
The number partitioning problem has drawn a lot of attention due to its theoretical aspects and properties and important
real-world applications in multiprocessor scheduling, the minimization of VLSI circuit size and delay, public key cryptogra-
phy, voting manipulation, etc. For a more detailed description of the applications we refer to [3,4].
There are several ways to solve the TWNPP in exponential time in n: the most naive algorithm would be to cycle through
all the subsets of n numbers and for every possible subset S1 and for its corresponding complementary S2 ¼ S n S1 calculate
their sums. Obviously, this algorithm is impracticable for large instances, since its time complexity is Oð2n Þ. A better expo-
nential time algorithm which runs in Oð2n=2 Þ was described by Horowitz and Sahni [5].

⇑ Corresponding author.
E-mail address: [email protected] (P.C. Pop).

0307-904X/$ - see front matter Ó 2013 Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.apm.2013.03.075
9192 P.C. Pop, O. Matei / Applied Mathematical Modelling 37 (2013) 9191–9202

Various heuristic algorithms have been developed for solving the TWNPP including: a natural greedy algorithm obtained
by sorting the numbers in decreasing order and then assigning each number in turn to the subset with the smaller sum so
far; a complete greedy algorithm described by Korf [6] where based on a binary tree each level assigns a different number
and each branch point alternately assigns that number to one subset or the other; the set differencing heuristic introduced
by Karmarkar and Karp [7] that repeatedly replaces the two largest numbers with their difference, inserting the new number
in the sorted order until there is only one number left which is the final partition difference, the complete Karmarkar-Karp
algorithm developed by Korf [6], a hybrid recursive algorithm obtained by combining several existing algorithms with some
new extensions developed by Korf [8] in the case of the multi-way partition and tested on the three, four and five-way par-
titioning. Alidaee et al. [9] presented a new modeling of the multi-way partition problem as an unconstrained quadratic bin-
ary program and solved it by efficient metaheuristic algorithms.
Several metaheuristic approaches have been proposed for solving the two-way number partitioning problem including a
Simulated Annealing algorithm by Johnsonn et al. [10], genetic algorithm by Ruml et al. [11], GRASP by Arguello et al. [12],
Tabu Search by Glover and Laguna [13], memetic algorithm by Berretta et al. [14], etc.
The multidimensional two-way number partitioning problem (MDTWNPP) was introduced by Kojic [15] and is a gener-
alization of the TWNPP in which given a set of vectors we are looking for a partition of the vectors into two subsets such that
the sums per every coordinate should be as close as possible.
The MDTWNPP is NP-hard, as it reduces when the vectors have dimension one to the TWNPP, which is known to be an
NP-hard problem.
Kojic [15] described as well an integer programming formulation and tested the model on randomly generated sets using
CPLEX. The obtained experimental results show that the MDTWNPP is very hard to solve even in the case of medium size
instances. To the best of our knowledge, this is the only existing approach for solving the MDTWNPP.
The aim of this paper is to describe a generalization of MDTWNPP called the multidimensional multi-way number par-
titioning problem (MDMWNPP) where a set of vectors has to be divided into a collection of mutually exclusive and collec-
tively exhaustive subsets such that the sums per every coordinate in each of the subsets are as nearly equal as possible. In
addition, we develop an efficient memetic algorithm based heuristic, obtained by combining a genetic algorithm with a pow-
erful local search procedure, for solving the MDMWNPP. The results of extensive computational experiments in the case of
multidimensional two, three and four-way partitioning are presented and analyzed. In the case of MDTWNPP the results re-
veal that our proposed methodology performs very well in terms of both quality of the solutions obtained and the compu-
tational time compared with the previous method introduced by Kojic [15].

2. Definition of the problem

Given a set of n vectors of dimension m

S ¼ fv i j v i ¼ ðv i1 ; v i2 ; . . . ; v im Þ; i 2 f1; . . . ; ng; m 2 Ng;
then according to Kojic [15] the multidimensional two-way number partitioning problem consists in splitting the elements
of S into two sets, S1 and S2 such that

1. S1 [ S2 ¼ S and S1 \ S2 ¼ ;;
2. the sums of elements in the subsets S1 and S2 are equal or almost equal for all the coordinates.

If we introduce the variable t that denotes the greatest difference in sums per coordinate, i.e.
( )
X X

t ¼ max  v ij  v ij j j 2 f1; . . . ; mg ;
 i2S i2S
1 2

then the objective function of the MDTWNPP is to minimize t. If min t ¼ 0 then the partition will be called perfect partition for
obvious reasons.
Next we define the multidimensional multi-way number partitioning problem (MDMWNPP) as a generalization of
MDTWNPP where a set of vectors is partitioned into a given number of subsets rather than into two subsets.
Let again S be a set of n vectors of dimension m and p 2 N; p P 2, then the multidimensional multi-way number parti-
tioning problem consists in splitting the elements of S into p subsets, S1 ; S2 ; . . . ; Sp such that.

1. S1 [ S2 [ . . . [ Sp ¼ S and Si \ Sj ¼ ;, for all i; j 2 f1; . . . ; pg and i – j;

2. the sums of elements in the subsets S1 ; S2 ; . . . ; Sp are equal or almost equal for all the coordinates.

In particular, if the the set of vectors is partitioned into two subsets we get the MDTWNPP. For partitioning into more than
two subsets, the objective function to be minimized is the greatest difference between maximum and minimum subset sums
per every coordinate. Introducing the variable t denoting the greatest difference between maximum and minimum subset
sums per every coordinate, i.e.
 P.C. Pop, O. Matei / Applied Mathematical Modelling 37 (2013) 9191–9202 9193

( ( ) ( ) )
 X X

t ¼ max max

v ij j l 2 f1; . . . ; pg  min v ij j l 2 f1; . . . ; pg j j 2 f1; . . . ; mg ;
i2S l i2S l

then the objective function of the MDMWNPP is to minimize t.

Example. Let S ¼ fð1; 3Þ; ð4; 4Þ; ð3; 2Þ; ð2; 5Þ; ð2; 1Þg and we want to partition its elements into three subsets S1 ; S2 and S3 .
We can do this partition in several ways, some candidates are:

 S1 ¼ fð1; 3Þg; S2 ¼ fð4; 4Þg; S3 ¼ fð3; 2Þ; ð2; 5Þ; ð2; 1Þg, then the sums are ð1; 3Þ; ð4; 4Þ; ð7; 2Þ, the difference between the
maximum and minimum values per coordinates is ð6; 2Þ and t ¼ 6;
 S1 ¼ fð1; 3Þ; ð3; 2Þg; S2 ¼ fð4; 4Þ; ð2; 5Þg; S3 ¼ fð2; 1Þg, then the sums are ð4; 1Þ; ð6; 9Þ; ð2; 1Þ, the difference between the
maximum and minimum values per coordinates is ð4; 10Þ and t ¼ 10;
 S1 ¼ fð1; 3Þg; S2 ¼ fð4; 4Þ; ð2; 1Þg; S3 ¼ fð3; 2Þ; ð2; 5Þg, then the sums are ð1; 3Þ; ð6; 3Þ; ð5; 3Þ, the difference between the
maximum and minimum values per coordinates is ð5; 1Þ and t ¼ 5;
 S1 ¼ fð2; 5Þg; S2 ¼ fð4; 4Þ; ð2; 1Þg; S3 ¼ fð3; 2Þ; ð1; 3Þg, then the sums are ð2; 5Þ; ð6; 3Þ; ð4; 1Þ, the difference between the
maximum and minimum values per coordinates is ð4; 4Þ and t ¼ 4;
 S1 ¼ fð4; 4Þg; S2 ¼ fð1; 3Þ; ð2; 1Þg; S3 ¼ fð3; 2Þ; ð2; 5Þg, then the sums are ð4; 4Þ; ð3; 2Þ; ð5; 3Þ, the difference between the
maximum and minimum values per coordinates is ð2; 2Þ and t ¼ 2.

Therefore, the minimum of the maximal elements of the listed candidates is in the fifth case with min t ¼ 2.

3. The memetic algorithm for solving the MDMWNPP

Memetic algorithms have been introduced by Mascato [16] to denote a family of metaheuristic algorithms that emphasis
on the use of a population-based approach with separate individual learning or local improvement procedures for problem
search. Therefore a memetic algorithm (MA) is a genetic algorithm (GA) hybridized with a local search procedure applied to
all the individuals in order to intensify the search space.
Genetic algorithms are not well suited for fine-tuning structures which are close to optimal solutions. Therefore, embed-
ding of local improvement operators into the recombination step of a GA is essential in order to obtain a competitive GA.
Memetic algorithms have been recognized as a powerful algorithmic paradigm for evolutionary computing, being applied
successfully to solve combinatorial optimization problems such as the VRP (Vehicle Routing Problem) [17,18] and the CARP
(Capacitated Arc Routing Problem) [19], the generalized traveling salesman problem [20], etc.
Our effective heuristic algorithm for solving the MDMWNPP is a memetic algorithm, which combines the power of ge-
netic algorithm with that of local search. The details of the proposed MA are formally expressed as:

Step 1 Construct the initial population (see Section 3.1.2).

Step 2 Use local search procedures to replace each of the individuals of the initial population by the local optimum (see
Section 3.2).
Step 3 Use genetic operators: crossover and mutation to produce the non-optimized next generation (see Section 3.1.4).
Step 4 Use local search procedures to replace each of the current generation solution by the local optimum (see
Section 3.2).
Step 5 Repeat Steps 3 and 4 until a termination condition is reached.

The general scheme of our heuristic is given in Fig. 1.

Next we give the description of our memetic algorithm for solving the multidimensional multi-way number partitioning
problem.

3.1. The genetic algorithm

3.1.1. Representation
It is known that a good representation scheme is important for the performance of the GA and it should
define noteworthy crossover, mutation and other specific genetic operators to the problem in order to minimize
the computational effort within these procedures. Designing such a representation is a hard problem in evolutionary
computation.
In order to meet this requirement we use an efficient representation in which the solution structure is a fixed size ordered
structure (n-dimensional vector) of integer numbers from the interval ½1; p. These integer numbers identify the set partitions
as assigned to the vectors’ (see Fig. 2). This representation ensures that the set of vectors belonging to the set S is partitioned
into p subsets S1 ; S2 ; . . . ; Sp .
9194 P.C. Pop, O. Matei / Applied Mathematical Modelling 37 (2013) 9191–9202

Fig. 1. Generic form of the proposed memetic algorithm.

Fig. 2. Representation of an individual’s chromosome.

3.1.2. Initial population

The construction of the initial population is of great importance to the performance of GA, since it contains most of the
material the final best solution is made of.
Experiments have been carried out with two different ways of generating the initial population:

(1) A common method of population generation is random generation. Each gene for a chromosome assumes a value of
P
i; i 2 f1; . . . ; pg, with probability pi , where pi¼1 pi ¼ 1. This approach is efficient and provides a population covering the
feasible region but it may lead to large values of the objective function yielding poor performance of the GA algorithm.
(2) Another method considered is based on generating the initial population partially randomly and partially based on the
problem structure. A random number q 2 f2; . . . ; ng is generated and then for the vectors belonging to f2; . . . ; qg the
genes are generated randomly and the other vectors are partitioned iteratively such that by adding each vector we
reduce the greatest difference in sums per coordinate.

Generating the population using as well the information about the problem structure permitted us to improve the initial
population by 50% with respect to the fitness value in comparison to the randomly generation of the initial population.

3.1.3. The fitness value

Every solution has a fitness value assigned to it, which measures its quality. The fitness function will use an evaluation
function to measure a value of worth for the individual so that they can be compared against each other and basically deter-
mines which possible solutions get passed onto multiply and mutate into the next generation of solutions. The rest of the
genetic algorithm will discard any solutions with a ‘‘poor’’ fitness value and accept any with a ‘‘good’’ fitness value.
In our case, the fitness value of the MDMWNPP, for a given partition of the vectors into p subsets is given by the greatest
difference between maximum and minimum subset sums per every coordinate. The aim is to find the partition that mini-
mize this value denoted in the previous section by t.
 P.C. Pop, O. Matei / Applied Mathematical Modelling 37 (2013) 9191–9202 9195

3.1.4. Genetic operators

Genetic operators are used in genetic algorithms to bring diversity (mutation-like operators) and to combine existing
solutions into others (crossover-like operators). The main difference among them is that mutation operators operate on
one chromosome, while the crossover operators are binary operators.
Crossover operator
During each successive generation, a proportion of the existing population is selected to breed a new generation. The
crossover operator requires some strategy to select two parents from the previous generation. In our case we selected the
two parents using the binary tournament method, where two solutions, called parents, are picked from the population, their
fitness is compared and the better solution is chosen for a reproductive trial. In order to produce a child, two binary tour-
naments are held, each of which produces one parent.
We experimented both single and double point crossover. Since there was not a big difference in the results we got from
both methods, we decided to use single point crossover. The crossover point is determined randomly by generating a random
number between 1 and n  1. We decided upon crossover rate of 85 % by testing the program with different values. This
means that 85% of the new generation will be formed with crossover and 15 % will be copied to the new generation.
Mutation operator
Mutation is a genetic operator that alters one ore more genes in a chromosome from its initial state. This can result in
entirely new gene values being added to the gene pool. With these new gene values, the genetic algorithm may be able
to arrive at better solution than was previously possible. Mutation is an important part of the genetic search as helps to pre-
vent the population from stagnating at any local optima and its purpose is to maintain diversity within the population and to
inhibit the premature convergence.
We consider a mutation operator that changes the new offspring by flipping values from i to j, where i; j 2 f1; . . . ; pg.
Mutation can occur at each value position in the string with 10 % probability.

3.1.5. Selection
Selection is the stage of a genetic algorithm in which individuals are chosen from a population for later breeding (cross-
over or mutation). The selection process is deterministic and is based on the fitness of the population. In our algorithm we
investigated and used the properties of ðl; kÞ selection, where l parents produce k (k  l) and only the offspring undergo
selection. In other words, the lifetime of every individual is limited to only one generation. The limited life span allows to
forget the inappropriate internal parameter settings. This may lead to short periods of recession, but it avoids long stagnation
phases due to unadapted strategy parameters, because the new population is not attracted into local optima by the old
population.
However, as the individuals evolve towards optima, the difference between the new population and the old one is getting
smaller, because the offspring are always in the vicinity of their parents. This is the mechanism which makes the MA benefit
of elitism. The high value of k ¼ 10  l increases the chances that a optimum is reached rather than a new population gets out
of it.

3.1.6. Genetic parameters

The genetic parameters are very important for the success of a GA, equally important as the other aspects, such as the
representation of the individuals, the initial population and the genetic operators. Based on preliminary computational
experiments, we set the following genetic parameters:

 The population size l has been set to 10 times the number of the vectors. This turned out to be the best number of indi-
viduals in a generation.
 The intermediate population size k was chosen ten times the size of the population: k ¼ 10  l.
 Mutation probability was set at 10%.
 The maximum number of generations (epochs) in our algorithm was set to 10,000.

In our algorithm the termination strategy is based on a maximum number of generations to be run if there is no improve-
ment in the objective function for a sequence of 15 consecutive generations.

3.2. Local improvement procedure

Computational experiments showed that our proposed GA involving just the crossover and the mutation operators is
effective in producing good solutions. However, based on the fact that classical GAs are not aggressive enough for some com-
binatorial optimization problems, we improved our GA algorithm by combining with local search procedures.
A local search heuristic tries to improve a solution by moving to a better neighbor solution. Whenever the neighboring
solution is better than the current solution, it replaces the current solution. When no better neighbor solution can be found,
the search terminates.
For each solution belonging to the initial population and for each new child obtained using the genetic operators, we use a
local improvement procedure that runs several local search heuristics sequentially. Once an improvement move is found, it
is immediately executed, meaning that the first improvement strategy is used rather than best-improvement strategy.
9196 P.C. Pop, O. Matei / Applied Mathematical Modelling 37 (2013) 9191–9202

In our algorithm we used the k-change neighborhood local search heuristic where up to k bits are changed at a time in a
complementary manner. We apply it for k ¼ 1; 2; 3 as follows:

 1-change neighbor. We select randomly an entry from the string representation. Suppose that the entry has a value
i 2 f1; . . . ; pg than change its value randomly with another one j 2 f1; . . . ; pg; j – i, meaning that we assign a vector from
a partition to another partition. If there is an acceptable quality gain after the change, then it is accepted. The complexity
of this procedure is OðnÞ, where n is the number of vectors. The 1-change procedure is repeated as long as improvements
are achieved.
 2-change neighbor. We select randomly two different entries from the string representation and swap their values, mean-
ing that we interchanged two vectors belonging to different partitions. If there is an acceptable quality gain after the
swap, then it is accepted. The complexity of this procedure is Oðn2 Þ and again the procedure is repeated as long as
improvements are achieved.
 3-change neighbor. We select randomly three different entries from the string representation and swap their values,
meaning that we interchanged three vectors belonging to different partitions. The complexity of this procedure is
Oðn3 Þ and again the procedure is repeated as long as improvements are achieved.

Our improvement procedure applies all the described local search heuristics cyclically. In the case of the MDTWNPP, we
apply successively the 1, 2 and 3-change neighborhoods, in this order.

4. Computational results

In this section we present computational results in order to asses the effectiveness of our proposed memetic algorithm for
solving the multidimensional multi-way number partitioning problem: two, three and four-way partitioning and we com-
pare them with the results Kojic [15] using CPLEX. It is worth to mention that while the MA was designed for our specific
problem and provides a suboptimal solution whose optimality cannot be checked, CPLEX is an optimization software pack-
age solving optimally linear programming problems, integer (mixed) programming problems, convex and non-convex qua-
dratic programming problems, semidefinite programming problems, etc. and can provide a verification. (See Figs. 3, 4).
We conducted our computational experiments for solving the MDMWNPP on a set of instances generated randomly and
following the general format n  m, where n represents the number of elements (vectors) and m represents the dimension of
the vectors. We consider for each n  m five instances denoted by a, b, c, d and e.
These instances were used by Kojic [15] in her computational experiments in the case of the multidimensional two-way
number partitioning problem (MDTWNPP).
In our computational experiments we performed 10 independent runs for each instance.
The testing machine was an an Intel Dual-Core 1,6 GHz and 1 GB RAM with Windows XP Professional as operating sys-
tem. The algorithm was developed in Java, JDK 1.6.
The following two tables show the computational results obtained with our memetic algorithm in comparison with those
obtained by Kojic [15] using CPLEX for instances containing between 50 and 300 vectors and with the dimension between 2
and 20 in the case of the MDTWNPP.

Fig. 3. MA versus GA and CPLEX for instances containing 400 vectors with the dimension between 10 and 20.
 P.C. Pop, O. Matei / Applied Mathematical Modelling 37 (2013) 9191–9202 9197

Fig. 4. MA versus GA and CPLEX for instances containing 500 vectors with the dimension between 10 and 20.

Table 1
Computational results for instances containing 50 and 100 vectors with the dimension between 2 and 20.

Problem Results of CPLEX Results of MA Problem Results of CPLEX Results of MA

instance instance
Best sol. time Best sol. Avg. sol. Avg. Best sol. time Best sol. Avg. sol. Avg.
time time
50_2a 3.204 552.60 5.438 5.502 35.73 100_2a 19.513 482.98 19.513 20.152 302.37
50_2b 9.193 63.75 9.193 9.546 32.38 100_2b 3.915 266.72 4.362 4.849 386.35
50_2c 9.191 161.9 9.191 9.756 38.98 100_2c 7.975 836.79 5.472 6.302 362.38
50_2d 4.753 92.01 4.753 5.023 42.65 100_2d 3.055 766.01 3.055 4.523 372.37
50_2e 6.719 1340.84 7.241 7.892 44.71 100_2e 2.077 1399.84 3.372 4.241 365.48
50_3a 303.581 465.91 283.563 289.541 162.34 100_3a 130.255 975.09 124.566. 130.332 397.28
50_3b 350.828 430.88 314.392 325.493 165.28 100_3b 97.935 80.71 98.372 108.283 392.64
50_3c 152.089 774.38 156.287 160.388 168.62 100_3c 263.199 383.98 256.822. 268.379 390.71
50_3d 102.516 1378.31 101.283 110.228 172.27 100_3d 247.043 1523.05 245.388 263.992 396.12
50_3e 217.903 312.16 198.854 212.211 171.64 100_3e 153.615 462.47 152.393 172.391 397.62
50_4a 909.765 1663.43 783.046 821.247 273.31 100_4a 520.496 664.68 521.283 531.635 402.83
50_4b 1272.224 1581.08 1017.283 1182.235 261.94 100_4b 1021.628 835.00 1021.628 1082.114 407.56
50_4c 461.161 1275.95 456.392 458.380 287.36 100_4c 908.24 232.91 900.394 910.838 402.66
50_4d 1024.681 827.7 987.27 1002.832 279.02 100_4d 1096.223 1678.59 1102.226 1132.432 411.42
50_4e 1199.574 498.3 1187.273 1201.398 281.07 100_4e 517.443 102.99 502.277 521.843 410.96
50_5a 926.164 281.77 918.278 925.226 286.45 100_5a 2441.489 570.23 2347.346 2536.142 521.76
50_5b 3202.875 1192.94 3002.869 3112.633 290.02 100_5b 2825.848 284.94 2539.387 2573.377 526.83
50_5c 2696.703 143.42 2563.904 2655.387 291.28 100_5c 2833.222 1236.07 2783.479 2837.741 534.21
50_5d 2275.792 357.38 2183.634. 2217.288 296.43 100_5d 2975.937 15.03 2637.388 2717.689 532.87
50_5e 4823.935 197.43 4689.382 4881.732 296.45 100_5e 4160.207 1072.36 4001.352 4112.556 540.02
50_10a 16176.578 1432.87 15722.29 16234.725 452.41 100_10a 17699.079 46.88 15792.742 16353.822 601.22
50_10b 19560.318 5.97 19560.318 20182.273 462.16 100_10b 18993.443 704.88 18367.389 18928.321 609.16
50_10c 17757.097 1308.68 15823.832. 16373.839 467.82 100_10c 15386.568 703.95 14628.836 15262.796 628.36
50_10d 14925.023 1104.7 14925.023 15025.637 425.76 100_10d 18276.246 337.87 16892.521 17066.389 615.27
50_10e 15369.009 472.44 14527.381 15263.833 476.45 100_10e 16516.277 689.61 15738.892 15938.556 612.67
50_15a 33208.019 1777.83 30728.546 32383.653 683.21 100_15a 32143.237 1082.82 30286.522 31762.362 642.37
50_15b 35003.301 1121.87 34362.391 35572.124 674.57 100_15b 28723.793 202.41 27893.837 28638.321 642.35
50_15c 29920.923 1443.76 28736.382 30826.276 684.65 100_15c 33363.206 1157.74 32982.347 33829.226 650.17
50_15d 21652.841 1594.98 20356.836 21723.622 690.25 100_15d 30706.17 1098.49 30706.17 31126.352 652.45
50_15e 31800.69 332.25 29018.285. 30025.245 692.46 100_15e 30253.623 280.81 28393.876 29731.532 659.81
50_20a 52826.34 71.83 50647.836 52876.443 635.26 100_20a 49992.607 1643.4 46782.412 47383.933 827.37
50_20b 51917.902 1378.83 50382.384 51662.929 657.41 100_20b 46691.489 1518.05 45673.837 46377.318 824.18
50_20c 50560.864 493.27 51829.374 51983.338 624.67 100_20c 45739.714 1042.18 44839.372 45211.776 826.31
50_20d 53955.965 166.26 51538.574 52632.307 635.82 100_20d 45371.992 406.33 45728.972 46182.389 834.38
50_20e 48281.499 234.75 47829.865 48292.728 652.36 100_20e 52315.704 1271.84 50732.852 51822.651 836.27
9198 P.C. Pop, O. Matei / Applied Mathematical Modelling 37 (2013) 9191–9202

Table 2
Computational results for instances containing 200 and 300 vectors with the dimension between 2 and 20.

Problem Results of CPLEX Results of MA Problem Results of CPLEX Results of MA

instance instance
Best sol. time Best sol. Avg. sol. Avg. Best sol. time Best sol. Avg. sol. Avg.
time time
200_2a 11.463 766.53 8.241 9.739 235.46 300_2a 2.744 1753.38 3.627 3.746 529.37
200_2b 3.919 915.91 4.211 4.764 253.46 300_2b 6.958 682.5 5.377 5.824 530.29
200_2c 0 704.92 2.382 2.653 246.48 300_2c 2.73 1273.13 7.821 7.937 540.26
200_2d 2.691 1146.03 4.271 4.365 252.83 300_2d 0.881 1446.18 1.718 1.975 541.28
200_2e 0.971 661.39 1.822 1.927 255.61 300_2e 1.522 1647.92 1.522 1.661 543.14
200_3a 181.368 1773.64 152.832 158.021 261.26 300_3a 6.1 941.22 5.241 5.568 542.1
200_3b 137.584 256.78 102.653 115.874 267.36 300_3b 110.139 1097.1 98.28 102.631 654.39
200_3c 3.059 537.67 10.283 14.983 271.27 300_3c 226.933 459.49 210.082 215.822 560.18
200_3d 224.645 1129.01 189.732 200.648 273.37 300_3d 137.587 952.38 135.85 136.026 563.1
200_3e 120.542 844.37 102.762 112.390 278.37 300_3e 188.581 263.35 152.39 162.710 570.8
200_4a 537.018 372.95 521.761 528.301 273.37 300_4a 15.268 537.98 416.387 421.386 575.47
200_4b 1188.248 32.82 1067.442 1152.223 271.27 300_4b 1068.095 293.00 563.076 578.102 582.65
200_4c 6.109 469.97 7.566 7.738 271.98 300_4c 900.62 355.95 870.065 892.651 586.9
200_4d 1094.743 92.75 992.55 996.365 277.28 300_4d 1004.401 972.48 820.382 856.309 592.3
200_4e 1264.715 24.06 1288.543 1290.846 281.28 300_4e 908.869 449.44 873.648 891.280 592.21
200_5a 1931.064 440.7 1836.645 1903.833 290.3 300_5a 1847.76 320.52 1500.277 1578.098 591.13
200_5b 2734.271 188.68 2583.557 2648.890 293.62 300_5b 4195.209 138.36 3647.364 3748.481 592.14
200_5c 3576.93 236.02 3476.526 3567.447 286.21 300_5c 2658.01 1621.06 2502.882 2675.087 598.63
200_5d 2782.748 58.78 2538.372 2738.474 297.29 300_5d 2396.939 1339.89 2037.532 2283.816 601.27
200_5e 3798.611 98.85 3578.36 3647.498 300.24 300_5e 2499.651 135.62 2103.37 2326.145 603.45
200_10a 16530.321 146.75 16450.302 16635.145 378.37 300_10a 16112.376 40.06 12839.361 14263.705 700.27
200_10b 19616.619 151.34 18360.366 18837.398 384.12 300_10b 19954.971 205.00 17282.376 18262.881 708.27
200_10c 16158.656 759.02 15830.382 15928.005 403.38 300_10c 15996.203 184.77 15996.203 16002.847 710.28
200_10d 17399.449 1640.67 15482.076 16377.385 410.27 300_10d 20282.178 1768.84 18293.364 18937.982 715.34
200_10e 18107.353 151.13 18002.322 18272.902 420.39 300_10e 19620.941 70.3 17823.355 18373.755 721.28
200_15a 35139.957 669.95 31026.391 33384.028 480.2 300_15a 37524.309 1678.59 35672.273 35852.406 842.19
200_15b 34575.029 649.44 32948.021 33262.398 483.41 300_15b 34673.445 737.63 30748.487 31934.125 851.82
200_15c 35016.095 934.43 30464.392 32443.839 482.43 300_15c 30553.455 208.9 28938.388 29461.056 853.2
200_15d 33160.395 742.52 32647.473 33023.752 489.38 300_15d 36264.63 179.13 35483.364 35884.485 861.2
200_15e 29600.126 493.5 27483.366 28363.822 510.28 300_15e 32237.793 186.39 30823.478 31262.384 865.38
200_20a 44991.718 872.96 41937.357 42393.927 520.39 300_20a 47297.493 1302.81 45362.379 45965.495 1002.3
200_20b 49884.338 377.89 48393.228 49272.912 523.39 300_20b 44127.831 940.69 44017.288 44938.205 1034.21
200_20c 48451.593 334.35 45627.277 47397.803 530.38 300_20c 43594.894 1033.94 40382.273 41540.227 1057.12
200_20d 43631.462 382.39 43251.189 43526.398 531.29 300_20d 48814.817 1338.33 45637.146 46779.021 1127.44
200_20e 41768.116 1247.7 40272.654 41435.037 539.27 300_20e 50067.495 799.1 50067.495 50829.114 1183.39

The first and the seventh columns in the tables give the instances, the second, third, eighth and ninth columns provide the
results obtained by Kojic [15] using CPLEX: the best solution and the necessary computational time in order to get it and the
forth, fifth, sixth and the last three columns provide the results obtained by our novel memetic algorithm: the best solution,
the average solution and the required time to get these average solutions. Because CPLEX did not finish its work in any con-
sidered instance in the tables are provided the best solutions obtained for a maximum of 30 min run for each test.
Analyzing the computational results from Tables 1 and 2, it should be noted that our approach compares favorable with
the approach provided by Kojic [15] with respect to the best solution values: in 107 out of 140 instances we have been able
to improve the objective function of the MDTWNPP, in 11 out of 140 we obtained the same solutions and in 22 out 140 in-
stances our solutions are higher than those obtained using CPLEX [15]. As well, we can observe that in 85 out of 140 in-
stances, even the average solutions provided by our MA algorithm are better than the solutions provided by Kojic using
CPLEX.
Tables 3 and 4 depict the results of the memetic algorithm along with a comparison of the genetic algorithm and those
obtained by Kojic [15] using CPLEX for instances containing between 400 and 500 vectors and with the dimension between 2
and 20 in the case of the multidimensional two-way number partitioning problem.
The first columns describe the instances, next two columns give the results of CPLEX: the best solution and the necessary
computational time in order to get it, next four columns give the results of the GA alone: the best solution, the relative gap
(as a percentage) between the best solution provided by CPLEX and the best solution provided by the GA, the average solu-
tion and the required time to get these average solutions and the last four columns give the results of the proposed MA: the
best solution, the relative gap (as a percentage) between the best solution provided by CPLEX and the best solution provided
by the MA, the average solution and the required time to get these average solutions.
Analyzing the results presented in Tables 3 and 4, we observe that our proposed heuristics MA performs very well in
terms of both solution quality and computational times in comparison with the approach provided by Kojic [15]: in 60
 P.C. Pop, O. Matei / Applied Mathematical Modelling 37 (2013) 9191–9202 9199

Table 3
Computational results for instances containing 400 vectors with the dimension between 2 and 20.

Problem Results of CPLEX Results of GA Results of MA

instance
Best time Best Gap % Average Average Best Gap Average Average
solution solution solution time solution % solution time
400_2a 12.592 1708.5 7.438 0.409 10.785 148.53 7.216 0.426 9.452 121.44
400_2b 1.53 830.32 1.53 0 1.982 146.9 1.53 0 1.604 132.38
400_2c 4.354 956.34 4.354 0 5.188 131.65 3.711 0.147 4.115 128.32
400_2d 4.42 151.52 4.775 0.08 5.546 135.8 4.42 0 5.168 130.29
400_2e 5.185 1427.76 5.185 0 5.966 156.38 4.839 0.066 5.104 149.32
400_3a 194.372 704.74 78.283 0.595 189.555 146.87 72.352 0.627 108.211 142.12
400_3b 175.9 478.81 129.352 0.264 193.818 154.38 121.235 0.31 167.233 123.81
400_3c 220.83 735.1 165.45 0.25 215.337 157.81 158.902 0.28 201.837 117.34
400_3d 257.849 80.78 178.345 0.308 292.76 158.27 127.746 0.504 228.287 134.18
400_3e 194.681 176.9 194.681 0 236.764 157.36 192.921 0.09 199.244 121.29
400_4a 1409.159 518.64 834.456 0.407 1008.728 188.28 821.822 0.416 892.384 156.39
400_4b 749.984 1237.82 683.273 0.088 856.671 192.28 662.002 0.117 664.623 136.34
400_4c 914.349 200.49 914.349 0 1132.511 191.27 901.255 0.014 905.115 176.35
400_4d 941.826 446.63 846.273 0.101 1149.915 193.24 842.721 0.105 1003.576 164.58
400_4e 902.761 576.52 627.75 0.304 896.435 195.46 622.382 0.31 723.197 176.59
400_5a 3737.767 79.9 3547.291 0.051 4912.599 262.89 3281.229 0.122 3385.147 221.72
400_5b 1494.313 1225.9 1453.29 0.027 1848.155 272.26 1421.065 0.049 1566.948 254.36
400_5c 2680.899 53.76 2640.928 0.014 3282.496 256.89 2638.989 0.015 2687.392 219.77
400_5d 2539.113 135.63 2423.948 0.045 3160.938 287.27 2421.321 0.046 2641.911 287.27
400_5e 1634.702 65.29 1489.725 0.088 2084.417 276.54 1465.103 0.103 1502.449 252.16
400_10a 14836.579 1622.34 7728.546 0.479 11235.401 356.27 7701.829 0.48 7838.938 340.66
400_10b 17918.141 1215.03 10918.141 0.39 15417.294 376.89 10827.980 0.395 11550.494 354.75
400_10c 21213.818 1703.88 9208.251 0.565 19088.936 324.5 9202.612 0.566 10144.09 290.43
400_10d 15212.906 1283.81 8212.906 0.46 14562.008 352.28 8212.906 0.46 8654.046 336.52
400_10e 16369.531 1530.48 13332.372 0.185 14009.345 342.56 12839.782 0.215 12992.594 320.45
400_15a 37574.022 926.03 29529.332 0.214 35754.439 321.35 29431.928 0.216 31043.73 304.65
400_15b 34390.093 62.52 21390.093 0.378 38253.683 326.28 21278.924 0.381 22411.202 279.04
400_15c 37161.817 1463.43 28621.829 0.229 40813.076 378.29 28425.883 0.235 29609.658 356.84
400_15d 30019.198 1203.22 16223.857 0.459 24117.878 381.1 16172.368 0.461 16293.763 366.54
400_15e 32561.093 26.19 30261.649 0.07 42612.356 378.29 30127.653 0.074 30466.451 321.76
400_20a 41974.284 767.3 28363.836 0.324 39800.183 390.28 28134.232 0.329 29765.538 366.02
400_20b 46751.348 354.05 38275.503 0.181 48713.445 368.45 37938.930 0.188 38740.68 354.77
400_20c 47259.514 313.95 32748.920 0.307 63799.921 372.32 32762.871 0.306 33469.524 342.41
400_20d 51544.421 31.38 36728.927 0.287 58840.664 310.38 36726.269 0.287 37467.746 289.36
400_20e 48792.272 251.36 43788.54 0.102 53750.069 302.65 43765.028 0.103 44902.726 263.47

out of 70 instances we have been able to improve the value of the objective function of the MDTWNPP, in 7 out of 70 in-
stances we obtained the same solutions and in the case of the instance 500 2a; 500 3e and 500 5c the solutions provided
by the MA are higher than the solutions obtained using CPLEX [15]. In addition, in 55 out of 70 instances even the average
solutions provided by the MA improved the quality of the solutions provided by CPLEX.
As well we observe that the performance of the MA approach is superior to the GA in terms of both solution quality and
computational times for MDTWNPP: in 63 out of 70 we have been able to improve the value of the objective function of the
MDTWNPP, in 6 out of 70 instances we obtained the same solutions and in the case of the instance 400 20c the solution pro-
vided by MA is higher that the solution obtained using just the GA. The reason for obtaining better results in shorter amount
of time is that when the search technique is incorporated in GA then the solution space in better searched.
According to the experimental results we can conclude that the local search heuristic is playing an important role in the
GA process.
In the next two figures we represent the average solutions obtained using our proposed MA-based heuristic in compar-
ison with the average solutions provided by the GA alone and the best solutions provided by CPLEX for instances containing
400 and 500 vectors with the dimension between 10 and 20.
Regarding the computational times, it is difficult to make a fair comparison between the approach of Kojic using CPLEX
and our method because they have been evaluated on different computers, they are implemented in different languages and
in addition Kojic did not mention the exact model of processor used in her computational experiments. The running time of
our MA is proportional with the number of generations. However, from Tables 1–4, it should be noted that in average our
heuristic is faster than the approach developed by Kojic [15].
In Fig. 5, we present a comparison of the average computational times of the MA versus GA, for instances containing be-
tween 400 and 500 vectors and with the dimension between 2 and 20.
9200 P.C. Pop, O. Matei / Applied Mathematical Modelling 37 (2013) 9191–9202

Table 4
Computational results for instances containing 500 vectors with the dimension between 2 and 20.

Problem Results of CPLEX Results of GA Results of MA

instance
Best time Best Gap % Average Average Best Gap % Average Average
solution solution solution time solution solution time
500_2a 0.62 1745.39 1.354 1.18 1.544 547.38 1.354 1.18 1.438 544.65
500_2b 1.879 1797.03 2.213 -0.177 2.653 546.39 1.879 0 2.242 492.29
500_2c 2.013 1076.04 2.839 -0.41 3.048 548.49 2.013 0 2.539 478.44
500_2d 1.003 1687.53 1.003 0 1.202 539.3 1.003 0 1.017 421.26
500_2e 0.913 1726.6 1.234 -0.351 1.077 529.39 0.913 0 1.006 382.48
500_3a 213.539 1698.09 124.54 0.416 219.304 580.52 124.54 0.416 165.983 524.42
500_3b 142.16 806.88 132.878 0.065 149.71 589.3 132.878 0.065 139.076 502.22
500_3c 270.729 466.64 221.463 0.182 270.558 540.41 219.763 0.188 260.085 522.79
500_3d 13.736 1134.38 12.9485 0.057 14.1217 592.46 12.1321 0.116 13.589 520.04
500_3e 7.625 1091.25 12.453 -0.63 14.94 521.43 8.746 0.147 9.225 381.33
500_4a 1562.920 171.29 1243.564 0.204 1864.105 638.37 1232.920 0.211 1749.848 567.08
500_4b 1186.722 429.41 980.253 0.174 1399.302 621.37 977.271 0.176 1082.888 465.59
500_4c 1093.756 63.74 837.384 0.234 1190.668 640.83 832.244 0.239 910.455 652.82
500_4d 4.58 714.37 4.58 0 5.961 610.27 4.580 0 4.869 552.16
500_4e 1410.694 106.31 938.384 0.334 949.861 634.52 937.823 0.335 1213.708 541.36
500_5a 3665.484 36.99 1922.135 0.475 3598.075 672.39 1903.802 0.48 2241.563 615.22
500_5b 4448.741 21.29 2833.853 0.363 4928.04 590.46 2824.579 0.365 3296.385 500.85
500_5c 3511.837 523.39 3928.203 -0.118 4367.228 696.59 3839.652 0.09 4176.629 590.72
500_5d 2597.644 81.42 1823.239 0.298 2383.455 621.47 1821.393 0.298 2028.585 621.47
500_5e 2572.429 955.58 1822.343 0.291 2448.885 654.73 1816.680 0.293 2023.652 579.12
500_10a 19183.301 1718.29 12938.304 0.325 18317.79 736.74 12892.029 0.328 13213.687 637.01
500_10b 12161.350 128.48 10393.382 0.145 11712.657 701.29 10387.625 0.145 10625.234 622.27
500_10c 16594.760 368.13 10283.385 0.38 10645.003 692.03 10282.295 0.38 10324.87 612.28
500_10d 20284.381 1699.01 16378.394 0.192 20874.022 678.9 16328.752 0.195 17283.48 678.90
500_10e 15548.670 1680.47 14950.76 0.038 16200.764 732.39 14738.344 0.052 15928.492 635.37
500_15a 30316.775 1055.81 20394.564 0.327 23566.966 720.31 20332.662 0.329 21087.687 682.31
500_15b 31878.383 1591.08 28348.563 0.11 39679.188 743.86 28283.569 0.112 29115.211 721.62
500_15c 32792.472 803.77 25484.567 0.222 41151.566 783.09 25461.223 0.223 26118.369 719.81
500_15d 35555.260 881.27 27394.64 0.229 46711.844 810.28 27283.392 0.232 28970.01 724.37
500_15e 30806.719 455.06 21849.57 0.29 22171.456 807.62 21652.005 0.297 22028.363 728.90
500_20a 48281.977 1000.12 32934.495 0.317 45219.074 843.3 32897.388 0.318 33238.612 843.30
500_20b 54921.900 237.63 38494.084 0.299 45750.657 873.39 38467.134 0.299 39675.389 818.39
500_20c 41578.884 1382.98 39495.452 0.05 41516.187 812.73 38763.093 0.067 39474.872 736.90
500_20d 54293.200 1728.58 43840.674 0.192 57730.837 843.3 43637.932 0.193 43763.0979 792.38
500_20e 41092.622 1713.03 40352.904 0.018 53286.063 921.83 40341.827 0.018 40654.702 918.22

Fig. 5. Running time comparison of the GA and MA.

 P.C. Pop, O. Matei / Applied Mathematical Modelling 37 (2013) 9191–9202 9201

Table 5
Computational results three-way and four way multidimensional partition.

Problem instance Results of MA Problem instance Results of MA

Best sol. Avg. sol. Avg. time Best sol. Avg. sol. Avg. time
50_2a 84.5 86.2 182.45 50_2a 370.7 391.4 342.37
50_3a 329.9 334.4 202.34 50_3a 650.3 678.3 536.24
50_4a 3370.6 3382.5 673.83 50_4a 801.2 836.7 1023.39
50_5a 4106.4 4125.8 781.82 50_5a 1006.4 1094.4 1243.28
50_10a 37485.4 37521.6 1189.38 50_10a 41829.1 42005.6 1647.76
50_15a 55960.2 56015.2 1212.27 50_15a 55960.2 56034.7 1843.92
50_20a 102394.8 102652.0 1235.22 50_20a 123484.8 123627.8 2135.48
100_2a 161.7 178.1 342.39 100_2a 662.2 687.2 564.02
100_3a 510.2 531.3 428.38 100_3a 1193.3 1213.7 847.37
100_4a 833.9 867.5 673.22 100_4a 1728.5 1924.6 922.14
100_5a 6169.4 6224.5 834.62 100_5a 8272.2 8356.8 1972.39
100_10a 46690.6 47004.8 1436.08 100_10a 74638.6 75034.8 2819.32
100_15a 96661.7 96827.3 2073.76 100_15a 122421.3 122892.7 3193.84
100_20a 112838.5 113112.5 2564.38 100_20a 174383.1 174981.6 4392.01

We can seen that the MA is more efficient that the GA as almost all the results are achieved in shorter time.
Table 5 shows the results obtained using our proposed MA in the case of the three-way and four-way multidimensional
number partitioning problem for instances containing 50, respectively 100 vectors with dimension between 2 and 20.
In the first part of Table 5 we presented the best solutions, average solutions and the average computational times in the
case of the three-way multidimensional partition problem and in the second part the results obtained in the case of the four-
way multidimensional partition problem.
Taking into account the computational results presented in Tables 1–5, we can conclude that our proposed MA-based
heuristic leads to good results within reasonable times and outperforms the approach introduced by Kojic [15] and the ge-
netic algorithm alone. The success of our algorithm relies on the combination of the genetic algorithm with the local search
procedure.

5. Conclusions

In this paper, we considered a generalization of the of the multidimensional two-way number partitioning problem
(MDTWNPP) where a set of vectors has to be partitioned into p sets (parts) such that the sums per every coordinate should
be exactly or approximately equal, called the multidimensional multi-way number partitioning problem (MDMWNPP).
We developed an efficient memetic algorithm for solving the MDMWNPP and in particular the MDTWNPP, that combines
a genetic algorithm with a powerful local search procedure consisting of three local search heuristics.
The extensive computational results show that our memetic algorithm is robust and compares favorably in comparison to
the existing approach and the genetic algorithm alone.
In the future, we plan to explore the possibility of building a parallel implementation of the system in order to improve
the execution time. In addition, we will need to asses the generality and scalability of the proposed heuristic by testing it on
larger instances.

Acknowledgments

This work was supported by a grant of the Romanian National Authority for Scientific Research, CNCS - UEFISCDI, project
number PN-II-RU-TE-2011–3-0113. The authors are grateful to the anonymous referees for reading the manuscript very
carefully and providing constructive comments which helped to improve substantially the paper.

References

[1] M.R. Garey, D.S. Johnson, Computers and Intractability. A Guide to the Theory of NP-Completeness, W.H. Freeman, New York, 1997.
[2] B. Hayes, The easiest hard problem, Am. Sci. 90 (2002) 113–117.
[3] E. Coffman, G.S. Lueker, Probabilistic Analysis of Packing and Partitioning Algorithms, John Wiley & Sons, New York, 1991.
[4] T. Walsh, Where are the really hard manipulation problems? The phase transition in manipulating the veto rule, in: Proceedings of IJCAI-09 (2009) pp.
324–329.
[5] E. Horowitz, S. Sahni, Computing partitions with applications to the Knapsack problem, J. ACM 21 (2) (1974) 277–292.
[6] R.E. Korf, A complete anytime algorithm for number partitioning, Art. Intel. 106 (2) (1998) 181–203.
[7] N. Karmarkar, R.M. Karp, The differencing method of set partitioning, Technical Report UCB/CSD 82/113, Computer Science Division, University of
California, Berkeley, 1982.
[8] R.E. Korf, A hybrid recursive multi-way number partitioning algorithm, in: Proceedings of the Twenty-Second International Joint Conference on
Artificial Intelligence (2011) pp. 591–596.
[9] B. Alidaee, F. Glover, G. Kochenberger, C. Rego, A new modeling and solution approach for the number partitioning problem, J. Appl. math. decis. sci. 9
(2) (2005) 135–145.
9202 P.C. Pop, O. Matei / Applied Mathematical Modelling 37 (2013) 9191–9202

[10] D.S. Johnson, C.R. Aragon, L.A. McGeoch, C. Schevon, Optimization by simulated annealing: an experimental evaluation. Part II: Graph coloring and
number partitioning, Oper. Res. 39 (3) (1991) 378–406.
[11] W. Ruml, J.T. Ngo, J. Marks, S.M. Shieber, Easily searched encodings for number partitioning, J. Optim. Theory Appl. 89 (2) (1996) 251–291.
[12] M.F. Arguello, T.A. Feo, O. Goldschmidt, Randomized methods for the number partitioning problem, Comput. Oper. Res. 23 (2) (1996) 103–111.
[13] F. Glover, M. Laguna, Tabu Search, Kluwer Academic Publishers, Norwell, Massachusetts, USA, 1997.
[14] R.E. Berretta, P. Moscato, C. Cotta, Enhancing a memetic algorithms’ performance using a matching-based recombination algorithm: results on the
number partitioning problem, in: M.G.C. Resende, J. Souza (Eds.), Metaheuristics: Computer Decision-Making, Kluwer, 2004.
[15] J. Kojic, Integer linear programming model for multidimensional two-way number partitioning problem, Comput. Math. Appl. 60 (2010) 2302–2308.
[16] P. Moscato, On evolution, search, optimization, genetic algorithms and martial arts: towards memetic algorithms, Caltech Concurrent Computation
Program, Report 826, 1989.
[17] j.e. Mendoza, B. Castanier, C. Gueret, A.L. Medaglia, N. Velasco, A memetic algorithm for the multi-compartment vehicle routing problem with
stochastic demand, Comput. Oper. Res. 37 (2010) 1886–1898.
[18] S.U. Ngueveu, C. Prins, R.W. Calvo, An effective memetic algorithm for the cumulative capacitated vehicle routing problem, Comput. Oper. Res. 37
(2010) 1877–1885.
[19] C. Prins, S. Bouchenoua, A memetic algorithm solving the vrp, the carp and general routing problems with nodes, edges and arcs, Stud. Fuzziness Soft
Comput. 166 (2005) 65–85.
[20] B. Bontoux, C. Artigues, D. Feillet, A memetic algorithm with a large neighborhood crossover operator for the generalized traveling salesman problem,
Comput. Oper. Res. 37 (11) (2010) 1844–1852.