A diploid genetic algorithm for solving the multidimensional

multi-way number partitioning problem

Adrian Petrovan Petrica C. Pop∗
Technical University of Cluj-Napoca, Technical University of Cluj-Napoca,
North University Center at Baia Mare, North University Center at Baia Mare,
Department of Electronics and Computer Science Department of Mathematics and Computer Science
Baia Mare, Romania Baia Mare, Romania
[email protected] [email protected]

Cosmin Sabo Oliviu Matei

Technical University of Cluj-Napoca, Technical University of Cluj-Napoca,
North University Center at Baia Mare, North University Center at Baia Mare,
Department of Mathematics and Computer Science Department of Electronics and Computer Science
Baia Mare, Romania Baia Mare, Romania
[email protected] [email protected]
ABSTRACT number partitioning problem. In Genetic and Evolutionary Computation
In this paper, we explore the Multidimensional Multi-way Number Conference Companion (GECCO ’23 Companion), July 15–19, 2023, Lisbon,
Portugal. ACM, New York, NY, USA, 4 pages. https://doi.org/10.1145/3583133.
Partitioning Problem (MDMWNPP), which is a more complex ver-
3590672
sion of the number partitioning problem. In MDMWNPP, we aim
to partition a set of vectors into a given number of subsets, such
that the sums of each subset’s elements are as close to equal as
possible for all vector coordinates. To address this problem, we pro-
pose a diploid genetic algorithm (DGA) that maintains population
diversity and enhances the performance of genetic algorithms (GA).
1 INTRODUCTION
We also integrate an efficient local search (LS) procedure to guide The number partitioning problem (NPP) is a classical combinatorial
the search towards promising solution regions. Our approach is optimization problem, which aims to partition a set 𝑆 of positive
compared to the classical (haploid) GA and state-of-the-art results integers into two subsets 𝑆 1 and 𝑆 2 such that the difference between
for MDMWNPP on 96 benchmark instances from the literature. the sum of elements in 𝑆 1 and the sum of elements in 𝑆 2 is minimized.
The results indicate that our method outperforms the classical GA NPP is a 𝑁 𝑃 − ℎ𝑎𝑟𝑑 problem, but for many instances it was solved
and is highly competitive with the state-of-the-art approaches. optimally or approximately by different heuristic approaches. For
this reason, NPP was called "the easiest hard problem" [8].
CCS CONCEPTS The multidimensional two-way number partitioning problem
was introduced by Kojić [6], denoted by MDTWNPP, where instead
• Computing methodologies → Genetic algorithms; • Theory
of the set of numbers we have a set of vectors of fixed dimension
of computation → Bio-inspired optimization; • Mathematics of
𝑚, with 𝑚 ≥ 2, that has to be partitioned such that the sums for
computing → Combinatorial optimization; • Software and its
each coordinate must be nearly equal as possible. In addition, Ko-
engineering → Genetic programming.
jić [6] provided an integer programming formulation of the and a
KEYWORDS collection of 210 instances, that were solved with CPLEX with an
upper time limit to 30 minutes for each instance. Pop and Matei
multi-way multi-dimensional number partitioning problem, genetic [14] extended the MDTWNPP by partitioning the set of vectors into
algorithms, diploid genetic algorithms, local search more than two subsets (𝑘 ≥ 2) and they called the problem the mul-
ACM Reference Format: tidimensional multi-way number partitioning problem, denoted by
Adrian Petrovan, Petrica C. Pop, Cosmin Sabo, and Oliviu Matei. 2023. MDMWNPP. The same authors developed a GA [13] and a memetic
A diploid genetic algorithm for solving the multidimensional multi-way algorithm [14] for solving the MDMWNPP. Faria et al. described
∗ Corresponding author Petrica C. Pop
in [4] a mixed integer programming formulation (MIP) to model
the MDMWNPP and provided extensive computational results for
Permission to make digital or hard copies of part or all of this work for personal or different values of 𝑘, 𝑘 ∈ {2, 3, 4}. Santucci et al. [16] proposed
classroom use is granted without fee provided that copies are not made or distributed an improved memetic algebraic differential evolution algorithm
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored. for solving the multidimensional two-way number partitioning
For all other uses, contact the owner/author(s). problem. Their computational results show that their proposed
GECCO ’23 Companion, July 15–19, 2023, Lisbon, Portugal approach is the state-of-the-art for MDTWNPP. Recently, Nikolic
© 2023 Copyright held by the owner/author(s).
ACM ISBN 979-8-4007-0120-7/23/07. et al. [9] proposed a novel MIP formulation of the MDMWNPP and
https://doi.org/10.1145/3583133.3590672 provided extensive computational results.

231
GECCO ’23 Companion, July 15–19, 2023, Lisbon, Portugal Adrian Petrovan, Petrica C. Pop, Cosmin Sabo, and Oliviu Matei

Since their introduction by Holland in 1975, the genetic algo- 3 DESCRIPTION OF THE DIPLOID GENETIC
rithms (GAs) have been applied successfully to solve various com- ALGORITHM
plex optimization problems, being nowadays one of the most pop-
DGAs have a structure which is quite similar to classical GAs, with
ular and used metaheuristic for providing high-quality solutions
two main distinctions: the first one concerns the representation
to optimization and search problems. It is well-known that if a
of the individuals and their evaluation and the second difference
GA loses the diversity of the population then it will be trapped
concerns the genetic operators: crossover and mutation.
into a local optimum. Several methods have been investigated to
In the following, we give a brief description of our DGA for
overcome this drawback, such as: methods based on increasing the
solving the MDMWNPP.
population randomness by using high mutation rate or injection
Representation. In the case of DGA the representation each indi-
of new random individuals [7], niching methods [3], the use of
viduals consists of two coupled chromosomes: 𝐼 = (𝐶 1, 𝐶 2 ), where
multiploid individuals considered by Collingwood et al. [2], etc.
𝐶𝑖 are the cromosomes, with 𝑖 ∈ {1, 2}. In the case of the investi-
DGAs have been inspired by solutions from biological evolution,
gated problem, in order to minimize the computational effort, the
namely the most complex organism have a diploid or multiploid
representation of each chromosome is made by a fixed size ordered
structure in their genetic representation and dominance schemes
structure (n-dimensional vector) of integer numbers from the in-
for gene selection [5] and they are characterized by the fact that
terval [1, 𝑘]. By this values we identify the partition assigned for
each of their individuals is composed by two chromosomes. DGAs
every vector belonging to 𝑆.
have been applied successfully in the case of dynamic optimization
Initial population. Due to the fact that a well-structured popu-
problems [1, 18, 19], etc., and as well to static optimization problems,
lation is of great importance to the performance of any GA, and
see for further information [11, 12, 15]. Recently, Petrovan et al.
based on the preliminary experiments that have been carried out,
[10] provided a comprehensive comparison in terms of the quality
the initial population was generated such that each subset 𝑆 𝑗 with
of the achieved solutions, the corresponding execution time and
𝑗 ∈ {1, ...𝑘 } contains at least one vector. The generation of the ini-
impact of the genetic operators on the quality of the results between
tial was done as follows: first 𝑘 position of an 𝑛 sized chromosome
the Haploid Genetic Algorithms (HGAs) and DGAs.
is populated by a permutation of values from 1 to 𝑘. Next 𝑛 − 𝑘
The aim of this paper is to describe a new efficient solution
positions are generated randomly with values between 1 and 𝑘.
approach for dealing with MDMWNPP, whose main feature is the
The dominance scheme. The dominance scheme in the case of
use of a DGA characterized by the use of two coupled chromosomes
DGA determines the way in which is constructed the phenotype of
for the representation of the individuals and a dominance scheme.
a given individual composed by two coupled chromosomes. In the
A local search procedure is coupled with the developed DGA to
literature have been proposed different approaches for describing
guide the search towards promising regions of the solution space.
the dominance scheme [17–19]. In this work, we considered the
The present paper is structured as follows: the second section
dominance scheme proposed by Shabash and Wiese [17], called
defines the investigated problem, Section 3 describes our novel
best-chromosome-wins, that instead of analyzing the dominance
solution approach, namely a diploid genetic algorithm (DGA) com-
of each locus, we take into consideration the entire chromosome
bined with a powerful local search (LS) procedure. In Section 4, we
and the better chromosome is selected to be dominant.
make a comparative analysis of the efficiency of our novel proposed
The fitness value. In our case, for a diploid individual a fitness
solution approach in comparison to the classical (haploid) GA and
value is calculated for each chromosome. These values for a given
the existing methods from the literature. Finally in Section 5, we
partitioning of the vectors into 𝑘 subsets are determined by the
formulate some conclusions and present future research directions.
the greatest difference between maximum and minimum subset
sums per every coordinate, i.e. 𝑟 (𝑆). Further, the fitness of an in-
2 PROBLEM DEFINITION dividual is represented in our implementation based on the used
Given a set of 𝑛 vectors of dimension 𝑚: dominance scheme, namely the fitness of the chromosome selected
𝑆 = {𝑣𝑖 = (𝑣𝑖1, 𝑣𝑖2, ..., 𝑣𝑖𝑚 ) | 𝑖 ∈ {1, ..., 𝑛}, 𝑚 ≥ 2} to be dominant.
Genetic operators. The processes carried out within the genetic
then the multidimensional multi-way number partitioning problem algorithm are performed by means of genetic operators and aim
(MDMWNPP) aims in dividing the vectors from 𝑆 into 𝑘 subsets, at reaching the exit condition from the iterative routine. However,
𝑆 1, 𝑆 2, ..., 𝑆𝑘 satisfying the following properties: maintaining the diversity of the population is essential in determin-
1. 𝑆 1 ∪ 𝑆 2 ∪ ... ∪ 𝑆𝑘 = 𝑆 ing the optimal or suboptimal solutions. In this regard, the correct
2. 𝑆𝑝 ∩ 𝑆𝑞 = ∅, for all 𝑝, 𝑞 ∈ 1, ..., 𝑛 with 𝑝 ≠ 𝑞 choice of crossover and mutation operator will make de difference
such that the sums of the elements in the subsets 𝑆 1, 𝑆 2, ..., 𝑆𝑘 are in the performance of the developed algorithm.
equal or almost equal for all the coordinates of the vectors. Crossover operator. In our GA, we have chosen to use the uniform
Given a partition of 𝑆 into the subsets 𝑆 1, 𝑆 2, ..., 𝑆𝑘 , we intro- crossover whose principle in the case of diploid individuals is the
duce the variable 𝑟 (𝑆) that denotes the greatest difference between following: after selecting the two parents which will be used in the
maximum and minimum subset sums per every coordinate, i.e. reproductive process, a mapping binary vector of length equal to
that of the chromosomes is generated and it will be used to form two
𝑟 (𝑆 ) = max
n
max
n ∑︁ o
𝑣𝑖 𝑗 − min
n ∑︁
𝑣𝑖 𝑗
oo
offspring by interleaving chromosomes of the two parents. When
𝑗 ∈{1,...,𝑚} 𝑖 ∈{1,...,𝑘 }
𝑖 ∈𝑆𝑘
𝑖 ∈{1,...,𝑘 }
𝑖 ∈𝑆𝑘 an 1 bit appears in the mapping vector then the chromosomes of
the two parents change the values between them, otherwise the
and then the objective of MDMWNPP consists in minimizing 𝑟 (𝑆).

232
A diploid genetic algorithm for solving the MDMWNP problem GECCO ’23 Companion, July 15–19, 2023, Lisbon, Portugal

values remain unchanged. Based on preliminary experiments, we Our novel DGA and HGA were implemented in Python and have
decided by testing different values to a crossover rate of 90%, this been evaluated on a PC with AMD Ryzen 9, 12-Core 3.8GHz, 64 GB
means 90% of the new generation will be obtained by crossover and RAM, Windows 11 Education operating system. In our proposed
10% are elites that will be copied to the new generation. algorithms for every instance we performed 30 independent trials.
Mutation operator. In our approach, we used an inversion muta- In Tables 1-2 , we report the solutions obtained by our DGA and
tion technique as follows: for each individual selected to be mutated, HGA for solving 96 instances of the MDMWNPP for four different
are generated two integers between 1 and 𝑛 and values between number of subsets 𝑘 = 5, 𝑘 = 10, 𝑘 = 20 and 𝑘 = 30 and compare
the two values are inverted for each chromosome of the individual. them against the existing results from the literature, provided by
Mutation can occur at each individual in the population with a Nikolic et al. [9].
probability of maximum 1%. The structure of the Tables 1-2 is as follows: the first three
Selection. As we have previously mentioned, in our implementa- columns and the columns 10 - 12 provide information concern-
tion we used the tournament selection method, which is a deter- ing the considered instances, the following two columns 4 - 5 and
ministic selection process based on the fitness of the individual. In the columns 13 - 14 provide the best found solutions (𝑟 ) reported
our algorithm we used the following principle: the lifetime of every by Nikolic et al. [9] achieved by solving the MIP formulation of the
individual is limited to one generation excepting the ones selected MDMWNPP by CPLEX, together with the CPU processing times
as elites which are copied to the new population. As the individu- (𝑡) in seconds. The columns 6 - 7 and 15 - 16 display the best found
als evolve towards optima, the difference between populations in solutions (𝑟 ) and the average running times (𝑡𝑎𝑣𝑔 ) achieved by our
successive epochs becomes smaller. HGA and the columns 8 - 9 and 17 - 18 display the best found
Genetic parameters. Based on preliminary experiments, we set solutions (𝑟 ) and the average running times (𝑡𝑎𝑣𝑔 ) achieved by our
the following values for the genetic parameters: the population size DGA. The best solutions are highlighted in bold.
was determined by using a multiplication factor 𝜇 = 300 of the Summarizing the results presented in Tables 1-2, we can no-
number of vectors in the instance, the mutation probability was set tice that the results obtained by our developed DGA are highly
at 1%, the elitism factor was set to 10%, the LS procedure selection competitive against the results reported by Nikolic et al. [9] and
factor was set at 30% and the termination strategy was set at 300 outperforms the results achieved by the HGA.
epoch without improvement in the objective function.
The local search procedure. In our approach, the individuals 5 CONCLUSIONS AND FURTHER RESEARCH
of the population are sorted in ascending order by fitness value. DIRECTIONS
Further, two LS techniques: Swap-change neighbour and 1-change
The aim of this paper was to solve the multidimensional multi-way
neighbour will be applied to the sorted individuals. The number of
number partitioning problem using a diploid genetic algorithm. We
individuals to whom local searches will be applied is proportional to
use diploidy to enhance the performance of the classical GAs by
a previously chosen factor. Our implementation use both presented
maintaining diversity of the population, and we combine the DGA
LS techniques, one after other.
with a local search procedure, in order to obtain a powerful and
effective novel solution approach for solving the MDMWNPP.
It is our objective to carry on our research and to investigate
4 COMPUTATIONAL RESULTS some other dominance schemes and local search strategies in order
to improve the efficiency of the proposed algorithm. We also intend
This section is dedicated to the preliminary experimental results ob-
to evaluate the generality and scalability of our developed diploid
tained by our proposed diploid genetic algorithm. In order to assess
genetic algorithm by evaluating it on larger instances.
the efficiency of our developed algorithm, we used the instances
originally proposed by Kojić [6] and extended by Nikolic et al. [9].
These instances contain five types of instances, denoted by A, B,
C, D and E, being generated five different matrices of dimension REFERENCES
500 × 200. Each entry of the matrices is a float number having four [1] Harsh Bhasin, Gitanshu Behal, Nimish Aggarwal, Raj Kumar Saini, and Shivani
decimals, and all the entries are generated randomly. Choudhary. 2016. On the applicability of diploid genetic algorithms in dynamic
environments. Soft Computing 20, 9 (2016), 3403–3410.
For evaluating the performance of our developed DGA, we com- [2] Emma Collingwood, David Corne, and Peter Ross. 1996. Useful diversity via
pared it against the classical HGA and to the results reported by multiploidy. In Proceedings of IEEE International Conference on Evolutionary Com-
Nikolic et al. [9] obtained by solving the MIP formulation of the putation. IEEE, 810–813.
[3] Jong De. 1975. KA An Analysis of the Behavior of a Class of Genetic Adaptative
MDMWNPP by CPLEX with a time limit of 1200 seconds. Systems. Ph. D. Dissertation. PhD thesis, University of Michigan.
The classical HGA uses a representation of the chromosome [4] Alexandre Frias Faria, Sergio Ricardo de Souza, and Elisangela Martins de Sa.
similar to each of the coupled chromosomes in the DGA, but the 2021. A mixed-integer linear programming model to solve the Multidimensional
Multi-Way Number Partitioning Problem. Computers & Operations Research 127
number of individuals in HGA was selected twice as the individuals (2021), 105133.
in DGA, keeping into account that each of the individuals in DGA is [5] Daniel L Hartl and Elizabeth W Jones. 1998. Genetics: principles and analysis.
Jones and Barlett Publishers.
composed by two chromosomes and evaluating a diploid individual [6] Jelena Kojić. 2010. Integer linear programming model for multidimensional two-
requires twice the processing time as a haploid individual, while way number partitioning problem. Computers & Mathematics with Applications
storing a diploid individual requires twice the memory as storing a 60, 8 (2010), 2302–2308.
[7] Oliviu Matei, Petrică C Pop, Jozsef Laszlo Sas, and Camelia Chira. 2015. An
haploid individual. All the other genetic operators used in the HGA improved immigration memetic algorithm for solving the heterogeneous fixed
are similar to those used in DGA. fleet vehicle routing problem. Neurocomputing 150 (2015), 58–66.

233
GECCO ’23 Companion, July 15–19, 2023, Lisbon, Portugal Adrian Petrovan, Petrica C. Pop, Cosmin Sabo, and Oliviu Matei

Table 1: Comparison results obtained by our proposed DGA against the results obtained by the HGA and the results obtained
by Nikolic et al. [9] using CPLEX for solving the MDMWNP problem in the case of instances of type A

Instance Nikolici [9] HGA DGA Instance Nikolici [9] HGA DGA
n m k 𝑟 𝑡 [s] 𝑟 𝑡𝑎𝑣𝑔 [s] 𝑟 𝑡𝑎𝑣𝑔 [s] n m k 𝑟 𝑡 [s] 𝑟 𝑡𝑎𝑣𝑔 [s] 𝑟 𝑡𝑎𝑣𝑔 [s]
30 2 5 1356.036 1200 3416.25 139.44 2197.068 93.62 40 10 5 55902.036 1200 65374.107 244.73 64149.876 131.57
30 2 10 15805.822 549.61 15972.238 176.13 16530.324 118.89 40 10 10 85086.796 1200 83717.814 435.93 79719.121 245.89
30 2 20 62110.281 1196.86 62110.281 116.75 62110.281 73.88 40 10 20 98692.56 1200 100309.459 705.86 95325.359 1077.29
30 2 30 94160.9 0.2 94160.9 53.21 94160.9 0.008 40 10 30 118822.794 1200 118493.449 394.9 118493.449 621.61
30 5 5 21117.612 204.85 31445.773 110.44 26051.691 104.65 40 20 5 116379.918 1200 113252.476 190.97 107894.551 130.58
30 5 10 50024.911 1200 55314.771 224.79 56089.272 111.17 40 20 10 140111.38 1200 121474.304 445.07 121474.304 197.22
30 5 20 90730.863 657.3 90730.863 178.71 90730.863 81.44 40 20 20 129228.716 1200 124038.81 500.35 120706.033 962.01
30 5 30 95895.434 1.21 95895.434 53.24 95895.434 0.012 40 20 30 144504.175 1200 140761.519 301.33 140761.519 454.48
30 10 5 62020.545 1200 69223.681 94.27 65712.679 71.78 50 2 5 2972.38 1200 2038.7 83.34 1819.704 168.38
30 10 10 82943.184 1200 83248.831 172.15 87387.897 106.52 50 2 10 9213.474 1200 7059.887 469.46 4474.539 7010.12
30 10 20 117217.197 1200 115493.647 179.49 115493.647 84.68 50 2 20 16728.958 1200 20569.542 242.5 15743.526 397.97
30 10 30 95973.224 2.47 95973.224 52.62 95973.224 0.009 50 2 30 41790.384 1200 46304.26 120.37 46121.941 241.66
30 20 5 114675.691 1200 120179.98 96.43 112147.446 66.6 50 5 5 26344.391 1200 30585.461 76.67 24771.402 165.71
30 20 10 131281.205 1200 124086.565 177.47 117869.271 98.85 50 5 10 51325.464 1200 45984.376 395.93 42466.117 389
30 20 20 140793.801 1200 139596.355 207.62 139594.03 105.47 50 5 20 72237.17 1200 93853.307 301.32 65339.243 503.67
30 20 30 98548.27 9.22 98548.27 47.52 98548.27 0.008 50 5 30 99117.746 1200 95846.766 144.31 95595.778 305.78
40 2 5 2356.018 1200 2214.377 256.36 1434.04 188.34 50 10 5 64304.349 1200 74020.335 101.12 63757.9 108.71
40 2 10 7650.608 1200 8842.657 368.29 7650.608 226.11 50 10 10 101400.02 1200 83430.353 423.14 78296.654 346.45
40 2 20 35259.236 1200 35307.02 510.96 35259.236 835.74 50 10 20 107742.195 1200 105868.841 342.98 105374.199 366.57
40 2 30 59599.299 59.22 59599.299 312.55 59599.299 432.28 50 10 30 125454.434 1200 123217.296 202.83 121840.287 317.16
40 5 5 22708.596 1200 33590.995 202.48 27177.394 191.79 50 20 5 109890.706 1200 111298.965 88.44 127784.308 174.51
40 5 10 53098.727 1200 49913.639 457.47 44082.467 228.39 50 20 10 144054.429 1200 129068.12 403.14 127950.537 335.96
40 5 20 70895.778 1200 71618.785 573.05 70895.778 1001.5 50 20 20 152283.617 1200 141275.385 233.41 140845.285 455.3
40 5 30 102479.974 1200 102479.974 355.13 102479.974 535.12 50 20 30 151382.038 1200 143483.721 146.84 142420.493 292.12

Table 2: Comparison results obtained by our proposed DGA against the results obtained by the HGA and the results obtained
by Nikolic et al. [9] using CPLEX for solving the MDMWNP problem in the case of instances of type B

Instance Nikolici [9] HGA DGA Instance Nikolici [9] HGA DGA
n m k 𝑟 𝑡 [s] 𝑟 𝑡𝑎𝑣𝑔 [s] 𝑟 𝑡𝑎𝑣𝑔 [s] n m k 𝑟 𝑡 [s] 𝑟 𝑡𝑎𝑣𝑔 [s] 𝑟 𝑡𝑎𝑣𝑔 [s]
30 2 5 1351.374 346.56 3811.574 140.62 2118.734 84.42 40 10 5 66108.425 1200 65223.089 176.2 65083.281 135.06
30 2 10 14338.294 72.1 16248.899 166.37 14338.294 98.08 40 10 10 86479.276 1200 86932.186 378.21 93559.603 228.93
30 2 20 63290.646 19.59 63290.646 142.14 63290.646 78.71 40 10 20 99276.092 1200 100990.455 558.87 97777.15 750.42
30 2 30 94629.75 0.37 94629.75 53.17 94629.75 0.009 40 10 30 131606.339 1200 131606.339 379.62 130517.914 650.54
30 5 5 22230.203 253.05 39201.056 105.56 22230.203 133.4 40 20 5 117125.746 1200 120571.568 200.41 115091.636 130.98
30 5 10 51908.359 829.52 53662.784 188 51908.359 114.53 40 20 10 134880.367 1200 130985.178 371.94 122910.719 192.73
30 5 20 102068.245 1200 102068.245 207.22 102068.245 88.01 40 20 20 132814.682 1200 130121.956 460.37 129885.769 678.18
30 5 30 97263.902 1.9 97263.902 53.23 97263.902 0.009 40 20 30 143862.081 1200 143862.081 353.79 143862.081 602.54
30 10 5 60879.343 1200 68411.878 110.59 66655.225 70.36 50 2 5 1892.319 1200 1696.536 108.99 1268.117 181.53
30 10 10 78854.718 1164.2 83841.518 178.22 78854.718 87.95 50 2 10 8904.598 1200 8713.726 408.66 5402.037 478.52
30 10 20 129726.102 1200 129244.214 162.31 130595.956 95.97 50 2 20 24097.44 1200 43617.561 285.41 17536.766 457.38
30 10 30 97396.324 4.39 97396.324 52.54 97396.324 0.009 50 2 30 46906.656 1200 46906.656 172.94 46906.656 205.17
30 20 5 107852.462 1200 120086.7 90.43 110721.107 75.7 50 5 5 27627.385 1200 26181.842 96.9 26181.842 169.61
30 20 10 124363.794 1200 121002.568 182.53 114760.87 95.53 50 5 10 54875.399 1200 47178.232 376.08 41287.616 420.68
30 20 20 144450.92 1200 145255.766 196.55 144194.698 97.1 50 5 20 74205.39 1200 71116.818 242.22 96835.355 492.34
30 20 30 98462.28 11.52 98462.28 47.45 98462.28 0.01 50 5 30 94336.443 1200 101730.707 142.65 104333.034 242.23
40 2 5 1485.355 1200 2099.586 232.69 2113.876 163.47 50 10 5 59746.907 1200 78616.244 76.43 75933.766 121.26
40 2 10 4030.423 1200 9657.649 416.43 7962.639 240.91 50 10 10 110606.209 1200 88871.931 378.67 85463.377 470.03
40 2 20 27900.963 596.02 28473.103 541.39 27900.963 685.46 50 10 20 115328.168 1200 115840.701 280.77 114836.139 472.49
40 2 30 70356.887 43.72 70356.887 310.76 70356.887 490.96 50 10 30 129899.7 1200 125176.226 188.1 125176.226 281.08
40 5 5 22708.596 1200 33590.995 202.48 27177.394 191.79 50 20 5 123875.647 1200 134621.529 87.07 123875.647 259.84
40 5 10 51657.648 1200 45259.349 449.79 44216.99 239.18 50 20 10 154426.38 1200 136373.714 341.95 131640.013 342.64
40 5 20 72079.643 1200 77624.431 548.27 72079.643 727.39 50 20 20 159637.004 1200.01 166308.515 237.85 148290.229 305.14
40 5 30 105617.486 1200 105617.486 308.86 105617.486 477.71 50 20 30 155334.547 1200 147424.711 200.78 143141.443 363.45

[8] Stephan Mertens. 2006. Number partitioning. Computational complexity and [15] Petrică C Pop, Oliviu Matei, Cosmin Sabo, and Adrian Petrovan. 2018. A two-level
statistical physics (2006), 125. solution approach for solving the generalized minimum spanning tree problem.
[9] Bojan Nikolic, Marko Djukanovic, and Dragan Matic. 2022. New mixed-integer European Journal of Operational Research 265, 2 (2018), 478–487.
linear programming model for solving the multidimensional multi-way number [16] Valentino Santucci, Marco Baioletti, and Gabriele Di Bari. 2021. An improved
partitioning problem. Computational and Applied Mathematics 41, 3 (2022), 1–72. memetic algebraic differential evolution for solving the multidimensional two-
[10] Adrian Petrovan, Oliviu Matei, and Petrică Pop. 2023. A Comparative Study be- way number partitioning problem. Expert Systems with Applications 178 (2021),
tween Haploid Genetic Algorithms and Diploid Genetic Algorithms. Carpathian 114938.
Journal of Mathematics 39, 2 (2023), 433–458. [17] Boris Shabash and Kay C Wiese. 2015. Diploidy in evolutionary algorithms for dy-
[11] Petrică Pop, Oliviu Matei, and Camelia Pintea. 2018. A two-level diploid genetic namic optimization problems: a best-chromosome-wins dominance mechanism.
based algorithm for solving the family traveling salesman problem. In Proceedings International Journal of Intelligent Computing and Cybernetics (2015).
of the Genetic and Evolutionary Computation Conference. 340–346. [18] Shengxiang Yang. 2006. On the design of diploid genetic algorithms for problem
[12] Petrica Pop, Matei Oliviu, and Cosmin Sabo. 2017. A hybrid diploid genetic based optimization in dynamic environments. In 2006 IEEE International Conference on
algorithm for solving the generalized traveling salesman problem. In International Evolutionary Computation. IEEE, 1362–1369.
Conference on Hybrid Artificial Intelligence Systems. Springer, 149–160. [19] Shengxiang Yang. 2007. Learning the dominance in diploid genetic algorithms
[13] Petrica C Pop and Oliviu Matei. 2013. A genetic algorithm approach for the mul- for changing optimization problems. (2007).
tidimensional two-way number partitioning problem. In International conference
on learning and intelligent optimization. Springer, 81–86.
[14] Petrică C Pop and Oliviu Matei. 2013. A memetic algorithm approach for solving
the multidimensional multi-way number partitioning problem. Applied Mathe-
matical Modelling 37, 22 (2013), 9191–9202.

234