Computers and Operations Research 127 (2021) 105133

Contents lists available at ScienceDirect

Computers and Operations Research

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

A mixed-integer linear programming model to solve the

Multidimensional Multi-Way Number Partitioning Problem
Alexandre Frias Faria a, Sérgio Ricardo de Souza a,⇑, Elisangela Martins de Sá a
a
Federal Center of Technological Education of Minas Gerais (CEFET-MG), Av. Amazonas, 7675 – Nova Gameleira, CEP 30510-000 Belo Horizonte, MG, Brazil

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

Article history: This paper addresses a mixed-integer linear programming model for solving the Multidimensional Multi-
Received 5 February 2020 Way Number Partitioning Problem (MDMWNPP), the most general version of the family of Number
Revised 30 July 2020 Partitioning Problems. First, a contextualization concerning the Two-Way Number Partitioning
Accepted 24 October 2020
Problem (TWNPP), the Multi-Way Number Partitioning Problem (MWNPP), the Multidimensional Two-
Available online 17 November 2020
Way Number Partitioning Problem (MDTWNPP), and the approached problem are presented.
MDMWNPP is a generalization of MDTWNPP for the number k of subsets greater than two. After the pro-
Keywords:
posed model is introduced, some properties are shown, and the main constraints are proved. To validate
Mathematical Programming
Combinatorial Optimization
the consistency of the proposed mathematical formulation, a comparison, using instances from the liter-
Number Partitioning Problem ature, between the proposed model for k ¼ 2 and a model from the literature for MDTWNPP, is per-
Multidimensional Multi-Way Number formed. The proposed model showed competitive results against this specialized model and, besides,
Partitioning Problem solved instances of the problem for values of k equals to 3 and 4.
Ó 2020 Elsevier Ltd. All rights reserved.

1. Introduction and other NP-complete problems. Horowitz and Sahni (1974) and
Schroeppel and Shamir (1981) proposed the first trivially accurate
This article addresses the Multidimensional Multi-Way Number algorithms adaptable to TWNPP, both for solving the Knapsack
Partitioning Problem (MDMWNPP), the most general version of the Problem. Korf (1998) introduced two proposals for applying
family of Number Partitioning Problems described in the scientific heuristics with exact methods for the TWNPP solution. The first,
literature. To make clear this statement, it is necessary to review called Complete Greedy Algorithm (CGA), uses the Longest Pro-
the previous related problems to contextualize this study. How- cessing Time heuristic (LPT), first introduced in Graham (1969);
ever, for its presentation, some definitions are initially required. the second, called Complete Karmarkar-Karp Algorithm (CKK), uses
Throughout this text, a partition of a set X is a collection of pair- the Differencing Method heuristic, better known as the Karmarkar-
wise disjoint nonempty subsets whose union forms X. Let k be a Karp Heuristic (KKH), presented in Karmarkar and Karp (1982).
fixed positive integer. Then, a k-partition of a set X is a partition The first generalization of TWNPP is the Multi-Way Number
of this set, with exactly k nonempty subsets. In this text, subsets Partitioning Problem (MWNPP). Let In be the set indexing the ele-
belonging to the partition are called parts. The notation ments of the numerical sequence V and a fixed positive integer
Im ¼ f y 2 Z : 1 6 y 6 mg is used to denote the closed set of all inte- k P 2. This problem consists of finding a k-partition of In , in the
gers between 1 and m. form fA1 ; A2 ; . . . ; Ak g, such that all elements of
nP P o
i2A1 v i ; . . . ; i2Ak v i
The seminal version of this family of problems is the Two-Way fit within the shortest possible interval.
Number Partitioning Problem (TWNPP). Let In be the set indexing
Thus, concerning TWNPP, MWNPP expands the number of parts
the elements of the numerical sequence
in which the set In must be distributed, from 2 to a generic number
V ¼ fv 1 ; v 2 ; . . . ; v i ; . . . ; v n g. TWNPP consists of finding a 2-
P P  k. MWNPP appears explicitly stated in Karmarkar and Karp (1982),
 
partition fA1 ; A2 g of In such that  i2A1 v i  i2A2 v i  is minimized. when analyzing the Differencing Method. This heuristic consists of
Karp (1972) formally listed TWNPP as a basic NP-complete prob- dividing the largest numbers into distinct parts, inserting the dif-
lem and proved a serie of equivalence relations between TWNPP ferences between the elements removed in the unallocated set,
as long as it is nonempty. Gent and Walsh (1998) show that
⇑ Corresponding author. MWNPP is a difficult problem for solving by general-purpose meta-
E-mail addresses: [email protected] (A.F. Faria), sergio@dppg. heuristics such as Genetic Algorithms, Simulated Annealing, and
cefetmg.br (S.R. de Souza), [email protected] (E.M. de Sá). others. In many cases, these methods lose in runtime and perfor-

https://doi.org/10.1016/j.cor.2020.105133
0305-0548/Ó 2020 Elsevier Ltd. All rights reserved.
Alexandre Frias Faria, Sérgio Ricardo de Souza and Elisangela Martins de Sá Computers and Operations Research 127 (2021) 105133

mance concerning KKH and even LPT. Korf (1998) presents exact appears in Pop and Matei (2013) with the resolution of the
algorithms to solve MWNPP, which incorporated adapted versions three-way ðk ¼ 3Þ and four-way ðk ¼ 4Þ cases using Memetic
of KKH and LPT as internal functions. Algorithm. It should be noted that this problem is still little stud-
The first improvement of these methods comes with the Recur- ied in the literature and does not have a proposed mathematical
sive Number Partitioning (RNP) algorithm proposed by Korf (2009), formulation. Further work on MDMWNPP and its main properties
which works with the resolution of smaller subproblems derived are found in Faria et al. (2019). Table 1 shows a briefing of the
from MWNPP. Through successive MWNPP conversions from a main articles concerning MDTWNPP and MDMWNPP. We high-
(k  1)-partition to a k-partition, Moffitt (2013) proposes an algo- lighted that this class of problems is solved mainly by meta-
rithm based on the resolution of smaller subproblems. Currently, heuristics. The only formulation in integer mathematical
the state-of-the-art for MWNPP is the Sequential Number Parti- programming is due to Kojić (2010), for the case k ¼ 2, i.e., for
tioning (SNP) algorithm, presented in Korf et al. (2014), and the MDTWNPP.
Cached Iterative Weakening (CIW) algorithm, presented in The MDMWNPP family of problems has important real-world
Schreiber and Korf (2014). Schreiber (2014) consolidates a thor- applications such as multiprocessor scheduling, minimization of
ough analysis of these algorithms. VLSI circuit size and delay, public-key cryptography (Pop and
The second generalization of TWNPP is the Multidimensional Matei, 2013), and assigning persons to fair teams (Hacibeyoglu
Two-Way Number Partitioning Problem (MDTWNPP). This gener- et al., 2018). Another application is the financial analysis concern-
alization considers V as a sequence of vectors, each one with ing the fair distribution of indivisible assets among the heirs, with-
dimension m. Let In be a set indexing the m-dimensional vectors out necessarily selling them. When the multidimensional factor is
of the sequence V. The problem consists of finding a 2-partition included, the problem is strongly related to the clusterization
fA1 ; A2 g of In such that the diameter induced by the infinity norm methods.
nP P o
i2A1 v i ; i2A2 v i
of the set be minimized. This generalization The current article proposes a mixed-integer programming
model for MDMWNPP. This is, to our best knowledge, the first
was initially proposed in Kojić (2010). This article also introduces
mathematical model proposed for this problem, and is, therefore,
a mixed-integer linear programming model for MDTWNPP, which
the main contribution of this article. For completeness and valida-
is solved using CPLEX. To the best of our knowledge, this is the only
tion, this model is compared to the mathematical model presented
mathematical model for this problem proposed to date. Fuksz et al.
in Kojić (2010) for MDTWNPP, i.e., valid only for k ¼ 2, with the
(2013) presents three heuristic algorithms for solving the Bi-
same set of instances used by this author and Pop and Matei
dimensional Two-way Number Partitioning Problem: a greedy
(2013). The analysis is further complemented by the solution,
algorithm, a genetic algorithm (GA), and a hybrid algorithm com-
using the proposed mathematical model, of this same set of
bining the genetic algorithm and the Variable Neighborhood
instances, for k ¼ 3 and k ¼ 4.
Search (VNS) metaheuristic. According to the authors, the hybrid
The paper is organized as follows: Section 2 presents the formal
algorithm GA-VNS outperforms both the greedy algorithm and
statement of MDMWNPP and numerical examples for better
the genetic algorithm when performed in isolation. Pop and
understanding its meaning. Section 3 uses the objective function
Matei (2013) approach MDTWNPP using populational-based meta-
from Pop and Matei (2013) to deduce the constraints of the pro-
heuristics for its solution, such as Memetic Algorithm (MA) and
posed mathematical model. Section 4 shows the relationship
Genetic Algorithm (GA). Kratica et al. (2014) present implementa-
between this proposed mathematical model and the mathematical
tions of the Variable Neighborhood Search (VNS) and
formulation presented in Kojić (2010) for the MDTWNPP. Section 5
Electromagnetism-like (EM) metaheuristics for the MDTWNPP
presents the computational results for the solution of the instances
solution and compare the results with those presented in Kojić
introduced in Kojić (2010). Finally, Section 6 concludes the article,
(2010) and Pop and Matei (2013). The results show that the EM
discussing the results obtained.
metaheuristic has performance slightly superior to the others. In
turn, Rodriguez et al. (2017) solve MDTWNPP using a hybrid meta-
heuristic involving Greedy Randomized Adaptive Search Procedure 2. Problem statement
(GRASP) and Exterior Path-relinking procedure. The results
obtained from the same set of instances used in Kojić (2010), Pop The Multidimensional Multi-Way Number Partitioning Problem
and Matei (2013), and Kratica et al. (2014) show the superiority (MDMWNPP) is originally introduced in Pop and Matei (2013). The
of this proposed hybrid method over the previously existing meth- formal statement is presented in the Definition 1.
ods. Hacibeyoglu et al. (2018) developed four metaheuristics to
solve MDTWNPP: Genetic Algorithm (GA), Simulated Annealing Definition 1 (MDMWNPP). Define v i ¼ ðv i1 ; v i2 ; . . . ; v il ; . . . ; v im Þ 2
(SA), Migrating Birds Optimization Algorithm (MBO), and Clonal Rm . Let V ¼ fv i gi2In be a sequence of vectors and k P 2 a positive
Selection Algorithm (CSA). The results showed that the GA and integer. The Multidimensional Multi-Way Number Partitioning Prob-
the SA implementations outperform the results from the other pro- lem (MDMWNPP) consists of finding a k-partition of In , in the form
  S
posed algorithms and the results from the mixed-integer program- Aj j2I , where 8j 2 Ik ; Aj – £, Aj ¼ In and Aj1 \ Aj2 ¼ £,
k
j2Ik
ming model proposed in Kojić (2010), in terms of the best solution
1 6 j1 < j2 6 k, which minimizes the diameter induced by the infinity
value obtained within a given time limit. Santucci et al. (2019) nP o
introduces the Memetic Algebraic Differential Evolution Algorithm norm of the multiset i2Aj v i . The objective function of this
j2Ik
for the binary search space (MADEB) and applies it to solve problem can be given by:
MDTWNPP. According to the authors, this algorithm surpasses
8 8 9 8 99
state-of-the-art algorithms for solving MDTWNPP instances.
   < <X = < X ==
The third generalization of TWNPP is the Multidimensional 
f Aj j2I ¼ max max v il ;  min: v il ;; ð1Þ
Multi-Way Number Partitioning Problem (MDMWNPP), proposed k l2Im : j1 2Ik : j2 2Ik 
i2Aj i2Aj
1 2
by Pop and Matei (2013). Let In be a set indexing the m-
dimensional vectors of the sequence V. This problem consists of Given a k-partition, Example 1 explains how to determine the
finding a k-partition fA1 ; A2 ; . . . ; Ak g of In such that the diameter value of the objective function (1).
induced by the infinity norm of the set
nP P P o
i2A1 v i ; i2A2 v i ; . . . ; i2Ak is minimized. MDMWNPP originally Example 1. Consider the sequence:
2
Alexandre Frias Faria, Sérgio Ricardo de Souza and Elisangela Martins de Sá Computers and Operations Research 127 (2021) 105133

Table 1
Summary of the bibliography for the Multidimensional Multi-Way Number Partitioning Problem.

Item Author Title Year Dim Part Method Type

1. Kojić (2010) Integer linear programming model for multidimensional two-way number partitioning 2010 MD 2W ILP ILP-
problem Exact
2. Pop and Matei A genetic algorithm approach for the multidimensional two-way number partitioning 2013 MD 2W GA Meta
(2013) problem
3. Pop and Matei A memetic algorithm approach for solving the multidimensional multi-way number 2013 MD kW MA Meta
(2013) partitioning problem
4. Kratica et al. Two metaheuristic approaches for solving multidimensional two-way number 2014 MD 2W VNS and EM Meta
(2014) partitioning problem
5. Rodriguez et al. GRASP with exterior path-relinking and restricted local search for the multidimensional 2017 MD 2W GRASP Meta
(2017) two-way number partitioning problem
6. Hacibeyoglu et al. A Comparative Analysis of Metaheuristic Approaches for Multidimensional Two-Way 2018 MD 2W GA, SA, MBO Meta
(2018) Number Partitioning Problem and CSA
7. Santucci et al. Algebraic Differential Evolution and VNS for the Multidimensional Two-Way Number 2019 MD 2W ADE + VNS Meta
(2019) Partitioning Problem


V ¼ ð14; 48; 23Þ1 ; ð87; 61; 48Þ2 ; ð76; 14; 23Þ3 ; Expression (4) gives rise to:
ð24; 25; 33Þ4 ; ð84; 13; 49Þ5 ; ð25; 48; 78Þ6 ;  8 9
 < X  
ð56; 14; 73Þ7 ; ð55; 21; 20Þ8 ; ð16; 13; 86Þ9 ; ð74; 55; 31Þ10 g  X =
 max 
j1 ;j2 2Ik : v il  v il
  6 y 8l 2 I m :
; ð5Þ
where n ¼ 10 and m ¼ 3. Given a solution  i2Aj i2Aj  
1 2
zfflfflfflfflfflfflffl}|fflfflfflfflfflfflffl{ zfflfflfflfflffl}|fflfflfflfflffl{ zfflfflfflfflfflffl}|fflfflfflfflfflffl{
S ¼ f1; 3; 8; 9g ; f2; 4; 7g ; f5; 6; 10g , it is implied that k ¼ 3. The
A1 A2 A3 which implies:
calculus of the objective function is then performed as follows. First,
Table 2 divides the V vectors according to the solution S and sums 8 9
the vectors contained in the same part, generating a set of 3 vectors. < X X =
Then, using the vectors from the last row of Table 2, the difference  y 6 max  v il  v il ; 6 y 8l 2 Im : ð6Þ
j1 ;j2 2Ik : 
i2A i2A
between the largest and the smallest element for each coordinate is cal- j1 j2

culated. Then, f183  161; 116  96; 158  152g ¼ f22; 20; 6g. The Using the same idea of the previous step:
largest of these values, i.e., 22, is the diameter of the set and the value
of the objective function. For the values of k 2 f3; 4; 5g and the same V
sequence, Table 3 shows optimal and feasible solutions. X X
y6 v il  v il 6 y 8l 2 Im ; 8j1 ; j2 2 Ik  Ik : ð7Þ
i2Aj i2Aj
1 2

3. Proposed mathematical model

Let now xji be a binary variable, such that xji ¼ 1 if the i index of
the i-th vector belongs to Aj part; and xji ¼ 0, otherwise. From
Consider the objective function introduced by Definition 1 and
given by Expression (1). From this expression, we will determine Expression (7), we get:
the sets of constraints to define a Mixed-Integer Linear Program-
ming (MILP) formulation for the Multidimensional Multi-Way
Number Partitioning Problem (MDMWNPP). X
n X
n

Let y 2 Rþ be the diameter to be minimized. Then, using the y6 v il xj i 1

v il xj i 6 y 8l 2 Im ; 8j1 ; j2 2 Ik  Ik :
2
ð8Þ
i¼1 i¼1
property:
  Putting v il in evidence in the Expression (8):
max ai  aj  ¼ max ai  min aj ð2Þ
i;j2In i2In j2In

we get:
8 8 9 8 99 X
n

< <X = < X == y6 v il xj1 i  xj2 i 6 y 8l 2 Im ; 8j1 ; j2 2 Ik  Ik : ð9Þ
y ¼ max max v il ;  min: v il ;; ð3Þ i¼1
l2Im : j1 2Ik : j2 2Ik 
i2A j1 i2A j2 It is possible to reduce the number of constraints in Expression
(9) to less than half without loss of generality. To achieve this
8 8 99 reduction, first consider that, when j1 ¼ j2 , the inequalities of
< < X X ==

¼ max  max   v il  v il ;; ð4Þ Expression (9) are always valid because, if y P 0, then
l2Im :j1 ;j2 2Ik :   y 6 0 6 y. Second, the feasibility of the family of inequalities (9)
i2Aj i2Aj
1 2
for the pair ðj1 ; j2 Þ implies the feasibility for the pair ðj2 ; j1 Þ due to
symmetry. The equivalent Expression (10) is obtained by multiply-
ing Expressions (9) by ð1Þ:
Table 2
Evaluating the objective function.

v i 2 A1 v i 2 A2 v i 2 A3 X
n
 X
n

ð14; 48; 23Þ1 ð87; 61; 48Þ2 ð84; 13; 49Þ5 y6 v il xj1 i  xj2 i 6 y () y6 v il xj2 i  xj1 i 6 y ð10Þ
ð76; 14; 23Þ3 ð24; 25; 33Þ4 ð25; 48; 78Þ6 i¼1 i¼1
ð55; 21; 20Þ8 ð56; 14; 73Þ7 ð74; 55; 31Þ10
After the deduction of the main constraints, we arrive at the
ð16; 13; 86Þ9
proposed mathematical model for MDMWNPP, with
ð161; 96; 152Þ ð167; 100; 154Þ ð183; 116; 158Þ
ðmkðk  1Þ þ k þ nÞ constraints and ðnk þ 1Þ variables, given by:
3
Alexandre Frias Faria, Sérgio Ricardo de Souza and Elisangela Martins de Sá Computers and Operations Research 127 (2021) 105133

Table 3
Feasible solutions vs Optimal solutions.

k Feasible f ðÞ Optimal f ðÞ

3 f1; 2; 3; 9g; f4; 7; 10g; f5; 6; 8g 54 f1; 3; 8; 9g; f2; 4; 7g; f5; 6; 10g 22
4 f1; 9; 10g; f2; 6g; f3; 4; 5g; f7; 8g 81 f1; 4; 5g; f2; 7g; f3; 6g; f8; 9; 10g 44
5 f1; 4; 5g; f2g; f3; 6g; f7; 10g; f8; 9g 59 f1; 5g; f2; 4g; f3; 6g; f7; 8g; f9; 10g 51

ðMODk Þ min y ð11aÞ Define the decision variable xi as:

X
k
s:t: xji ¼ 1; 8i 2 In ð11bÞ 1; if i 2 A1
xi ¼ ð12Þ
j¼1
0; if i 2 A2
X
n
xji P 1; 8j 2 Ik ð11cÞ and the variable t as:
i¼1  
Xn
 X X 

v il xj1 i  xj2 i 6 y; 81 6 j1 < j2 6 k; 8l 2 Im ð11dÞ t ¼ max wil 
l2Im 
wil 

ð13Þ
i¼1 i2A i2A
1 2
Xn

v il xj2 i  xj1 i 6 y; 81 6 j1 < j2 6 k; 8l 2 Im ð11eÞ Clearly, t denotes the largest difference between the sums per
i¼1 coordinate. The MILP from Kojić (2010) is, thus:
y 2 Rþ ð11fÞ 
MODKojic min t ð14aÞ
xji 2 f0;1g; 8ð j;iÞ 2 Ik  In ð11gÞ
X
n
s:t:  0:5t þ wil xi 6 0:5sl ; 8l 2 I m ð14bÞ
i¼1
The objective function (11a) minimizes the diameter of the
nP ok X
n
0:5t þ wil xi P 0:5sl ; 8l 2 I m ð14cÞ
multiset i2Aj v i , given by the nonnegative real variable y.
j¼1 i¼1
The equations (11b) ensure that the parts form a collection of xi 2 f0; 1g; 8i 2 I n ð14dÞ
pairwise disjoint subsets. The inequalities (11c) ensure that each
j part is nonempty. The inequalities (11d) and (11e) calculate Expression (14a) defines the objective function of this prob-
the difference between the sums of each pair of distinct parts. lem. Kojić (2010) shows how to determine constraints (14b)
Each one of these differences is limited by the variable y. The and (14c) from the definition of xi and t variables, presented in
membership relation (11f) ensures that the diameter is nonneg- Expressions (12) and (13), respectively. Kojić (2010) highlights
ative. The membership relation (11g) indicates that the variables that the proposed model has n binary variables and 2m
xij are binary variables such that xji ¼ 1, if the index vector i of constraints.
the V sequence belongs to the index part j, and xji ¼ 0
otherwise. 4.2. Comparison between models (MODk ), for k ¼ 2, and (MODKojic )

Consider the MILP for MDMWNPP, given by model (MODk ) and

4. Comparison between models shown by Expressions (11). For k ¼ 2, this mathematical model
becomes:
This section presents an analysis of the MILP for the Mul-
tidimensional Two-Way Number Partitioning Problem ðMODk¼2 Þ min y ð15aÞ
(MDTWNPP) presented in Kojić (2010) and a comparison s:t: x1;i þ x2;i ¼ 1; 8i 2 I n ð15bÞ
between this model and the proposed MILP shown by Expres- Xn
sion (11). x1;i P 1 ð15cÞ
i¼1
Xn
4.1. The mathematical model of Kojić (2010) x2;i P 1 ð15dÞ
i¼1
According to Kojić (2010), the mathematical model of Xn

MDTWNPP is defined as follows. Let V ¼ fwi gi2In be a sequence v il x1;i  x2;i 6 y; 8l 2 I m ð15eÞ
i¼1
of real vectors such that n ¼ jVj and wi 2 Rm ; 8i 2 In . The goal is
Xn

to partition it into two subsets A1 and A2 , so that A1 [ A2 ¼ In v il x2;i  x1;i 6 y; 8l 2 I m ð15fÞ
and A1 \ A2 ¼ £, in order to minimize the difference between i¼1
the sum of the elements of each subset. Let also wil be the l-th y 2 Rþ ð15gÞ
coordinate of the i-th element wi ¼ ðwi1 ; . . . ; wim Þ in the sequence
xji 2 f0; 1g; 8ð j; iÞ 2 I2  In ð15hÞ
V, and:
P
X
n Define sl ¼ ni¼1 v il . Performing the replacement x1;i ¼ 1  x2;i ,
sl ¼ wil ; 8l 2 I m : based on constraints (15b), the model ðMODk Þ for k ¼ 2 result
i¼1 in:

4
Alexandre Frias Faria, Sérgio Ricardo de Souza and Elisangela Martins de Sá Computers and Operations Research 127 (2021) 105133

ðMODk¼2 Þ min y ð16aÞ function value. In this case, the default optimal solution is also
Xn xi ¼ 0; 8i 2 In , since this value is the initial relaxation with the
s:t:: x1;i P 1 ð16bÞ objective function equal to maxfjai jg.
i¼1 Looking at the instances used in Kojić (2010), we note that all
Xn
numbers are strictly positive, i.e., all vector coordinates are such
x1;i 6 n  1 ð16cÞ
i¼1
that v il > 0. This hypothesis will be evaluated next.
y Xn
s
 þ v il x1;i 6 l ; 8l 2 I m ð16dÞ Proposition 1. Let x 2 Rm m
þ and y 2 Rþ be vectors with strictly
2 i¼1 2
positive coordinates. Then:
y Xn
s
þ v il x1;i P l ; 8l 2 I m ð16eÞ jjx  yjj1 < jjx þ yjj1
2 i¼1 2
y 2 Rþ ð16fÞ
Proof. Consider a fixed coordinate xl > 0 and yl > 0 of the given
x1;i 2 f0; 1g; 8i 2 I n ð16gÞ
vectors. We want to prove that jxl  yl j < jxl þ yl j, 8l 2 Im . Thus, by
The mathematical model shown by Expressions (16) differs applying the maximum on both sides, the proof concludes.
from the model proposed in Kojić (2010) only due to the con-
According to the definition of absolute value, if xl P yl , then
straints (16b) and (16c). These constraints are useful for preventing
jxl  yl j ¼ ðxl  yl Þ. Once yl e xl are strictly positives, we get:
a solution with all vectors contained in the same part and, thus,
avoiding a j partition from occurring such that Aj ¼ £. Previous jxl  yl j ¼ ðxl  yl Þ < ðxl  yl þ 2yl Þ ¼ ðxl þ yl Þ ¼ jxl þ yl j
articles dealing with MDTWNPP and MDMWNPP, however, do
Otherwise, if xl < yl , then:
not incorporate the property Aj – £ as part of describing these
problems. The current article, on the other hand, incorporates it jxl  yl j ¼ ðyl  xl Þ < ðyl  xl þ 2xl Þ ¼ ðyl þ xl Þ ¼ jyl þ xl j
as being necessary to the complete description of these problems.
Hence:
Without enforcing that the subsets must be non-empty, the correct
description of these problems should be ‘‘to divide in at most 2 sub- jxl  yl j < jxl þ yl j; 8l 2 Im
sets” or another similar statement.
Taking the maximum on both sides:
If the instances are generated with vectors in Rm containing
negative numbers, the model proposed by Kojić (2010), without maxfjxl  yl jg < maxfjxl þ yl jg
l2Im l2Im
constraints (16b) and (16c), would lead to a wrong optimal solu-
tion. This fact is shown in Example 2. Thus:
jjx  yjj1 < jjx þ yjj1
Example 2. Let
which concludes the proof.
Now, let fA1 ; A2 g be a 2-partition of In . Proposition 1 can be used
V ¼ fð3; 3Þ; ð1; 1Þ; ð1; 1Þ; ð1; 1Þg to conclude that:
be a set of vectors. Thus, n ¼ 4 and m ¼ 2. The optimal solution of X X X X X
MDTWNPP, i.e, for k ¼ 2, for this instance is: vi  vi < vi þ vi ¼ vi
i2A1 i2A2 1 i2A1 i2A2 1 i2In 1
V1 ¼ fð3; 3Þ; ð1; 1Þ; ð1; 1Þg;
Consequently, the constraints (16c) and (16d) are not necessary
V2 ¼ fð1; 1Þg
in this case because any 2-partition of In will have an objective
However, from the application of the mathematical model proposed by function value strictly smaller than a solution containing all vec-
Kojić (2010), the result is: tors in a single part, as occurred in Example 2. In conclusion, the
model presented in Kojić (2010) can be used for testing, as long
V1 ¼ fð3; 3Þ; ð1; 1Þ; ð1; 1Þ; ð1; 1Þg;
as the hypothesis v il > 0; 8ði; lÞ 2 In  Im is assumed.
V2 ¼ £
Note, therefore, that this solution contains an empty part, that is, the 5. Computational experiments
partition V 2 is an empty set, and, on the other hand, the partition V 1
contains all the elements of own instance V, i.e., V 1 ¼ V and V 2 ¼ £. The computational experiments were performed on a computer
with an Intel Core i7-3770 CPU, 3.4 GHz, 8-core, 32 GB RAM, and
In fact, there are numerous examples of instances where the
64-bit Ubuntu 16.04 operating system. The mathematical model
formulation of Kojić (2010) fails, such as:
(MODk ), shown by Expressions (11), was implemented using CPLEX
vi ¼ ð1; 1Þ; 8i 2 In1 solver version 12.7, with the default configuration, in C++ language
V¼ ð17Þ
vn ¼ ðn  1; n  1Þ
and gcc compiler version 5:4:0.
V ¼ fð1; 3Þ; ð1; 2Þ; ð0; 1Þ; ð6; 5Þ; ð3; 3Þ; ð3; 2Þg ð18Þ
The set of instances used in the experiments was originally gen-
V ¼ fða1 ; 0; 0; 0Þ; ð0; a2 ; 0; 0Þ; ð0; 0; a3 ; 0Þ; ð0; 0; 0; a4 Þg; a1 ; a2 ; a3 ; a4 2 R: ð19Þ
erated for the paper Kojić (2010) and provided to us by authors of
If the sum of the vectors is equal to the null vector, then the the papers Kratica et al. (2014) and Pop and Matei (2013). All
value of the objective function is zero, since the first integer solu- papers cited in Table 1, except Hacibeyoglu et al. (2018), used this
tion found is always xi ¼ 0; 8i 2 In . This solution satisfies the con- same dataset for their experiments. In this set of instances, each
straints (16d) and (16e), as well as reaching a MILP optimality sequence V contains n vectors with dimension m, being
criterion, even if one of the parts is empty. Among the examples n 2 f50; 100; 200; 300; 400; 500g and m 2 f2; 3; 4; 5; 10; 15; 20g.
presented by Expressions (17), (18), and (19), the one associated There are 5 different kinds of instances, named as
with Expression (19) is the instance with the most difficult solu- r 2 fa; b; c; d; eg, for each combination ðn; mÞ, totalizing 210 possi-
tion because all of its possible partitions have the same objective ble different sequences. The description of the instances is com-

5
Alexandre Frias Faria, Sérgio Ricardo de Souza and Elisangela Martins de Sá Computers and Operations Research 127 (2021) 105133

Table 4 Status 2 fInfeasible; Feasible; Optimalg; a real number UB

Results for the model (MODKojic ) (upper bound); a real number LB (lower bound); and the running
Status Measure UB LB Time time in seconds. Each line represents an instance solved. The same
Feasible: 209 Minimum 0.25 0 1271 computational test environment was used to solve both (MODKojic )
Optimal: 1 1st Quartile 121.06 0 1797 and (MODk ) models. Thus, we did not use the results presented in
Median 1899.38 0 1800 Kojić (2010) for comparison purposes since, for allowing a fair
Average 14481.81 1263 1795 evaluation, model (MODKojic ) was again solved in our computa-
3rd Quartile 29184.12 0 1800
Maximum 58791.40 50556 1800
tional environment.
To better organize the article, Appendix A, containing Tables
A.13, A.14, A.15, A.16, A.17, and A.18, presents the comparison,
for k ¼ 2, between the formulations (MODk ) and (MODKojic ), shown
by expressions (11) and (14), respectively. Additionally, Appendix
Table 5 B, with Tables B.19, B.20, B.21, B.22, B.23, and B.24, shows the
Results for the model (MODk ) using k ¼ 2
results obtained by the model (MODk ) for k ¼ 3 and k ¼ 4. On the
Status Measure UB LB Time other hand, Tables 4, 5, 6, and 7 summarize the results described
Feasible: 210 Minimum 0.13 0 1761 in these tables cited before, aiming to support the analysis. Its mea-
1st Quartile 146.02 0 1798 sures are calculated over all instances. In particular, Tables 4 and 5
Median 1951.33 0 1799 resume the results described in Appendix A, concerning the models
Average 14303.46 1248 1797
3rd Quartile 29153.33 0 1800
(MODk ) and (MODKojic ) for k ¼ 2; and Tables 6 and 7 summarizes
Maximum 56535.80 45184 1800 the results from Appendix B using the model (MODk ) for, respec-
tively, k ¼ 3 and k ¼ 4.
Table 4 reports that an optimal solution was found by model
(MODKojic ), and, according to Table A.13, this optimal solution is
due to the instance 2 50 20c. Note that this only instance for which
an optimal solution was found addresses the least number of vec-
Table 6 tors to be partitioned (n ¼ 50), but the vectors of this instance have
Results for the model (MODk ) using k ¼ 3
the largest dimension considered in the set of instances (m ¼ 20).
Status Measure UB LB Time In this instance, the models (MODKojic ) and (MODk ) have the same
Feasible: 210 Minimum 39.71 0.0 1707 UB value, but the model (MODk ) does not certify the optimality
1st Quartile 1749.71 0.0 1799 of its solution. On the other hand, Tables 4 and 5 allow evaluating
Median 9102.22 0.0 1799 the performance of these two models. Regarding the average of the
Average 28504.28 377.1 1789
UB values for all 210 instances tested, it is concluded that the
3rd Quartile 59631.55 0.0 1800
Maximum 96176.20 18418.0 1800 model (MODk ) is better than the model (MODKojic ). Note that the
difference between the values is small for all measures of these
two tables.
Another descriptive analysis of the results from Appendix A is
given in Table 8. This table contains the number of solutions found
by the model (MODk ), classified, concerning the solutions from the
Table 7
model (MODKojic ), as better, worse, or equal. From this table, the
Results for the model (MODk ) using k ¼ 4
quality of the results from the model (MODk ) is slightly superior
Status Measure UB LB Time to those from the model (MODKojic ), once the balance between the
Feasible: 210 Minimum 254.9 0.0 1755 number of better and worse solutions for the values of UB and LB
1st Quartile 4213.8 0.0 1799 shows favorable to the solutions from the model proposed in this
Median 15769.0 0.0 1799
article.
Average 37841.8 443.1 1798
3rd Quartile 77292.1 0.0 1799 To verify if there is a statistical difference between the UB and
Maximum 113213.0 19955.0 1800 LB values obtained by models (MODk ) and (MODKojic ), as well as
concerning the runtime for the experimental tests with these mod-
els, we used the Paired Wilcoxon signed-rank test with continuity
correction (Dalgaard, 2008). These tests were applied in a sample
composed of the results found for all instances with k ¼ 2, i.e.,
for all the results shown in Appendix A. The result of these statis-
Table 8
A brief counting of results obtained by model (MODk ) in relation to model (MODKojic ). tical tests is presented in Table 9. Table 9 shows that, for the set of
instances used in this paper, the UB values found by (MODk ) are not
Better Worse Equal
less than the UB values of (MODKojic ); the LB values found by (MODk )
UB 103 93 14 are not greater than the LB values of (MODKojic ), and there is a sig-
LB 4 6 200
nificant difference between the runtime of these two models. The

Table 9
Paired Wilcoxon signed rank test with continuity correction applied to the results of
pleted by associating a value k 2 f2; 3; 4g, representing the parti- (MODk ) and (MODKojic ).
tion size. Hence, the name of the instances has the form of k n mr.
Measure W p-value H1
A single core is used for each run. The time limit given to the
solver for each test instance is 1800 seconds and the maximum UB 8613 0.0955 D<0
LB 26 0.5808 D>0
memory used for testing is 5 GB. After each running, the algorithm
Time 7207 0.0002 D–0
writes a line with the following attributes of the solution found:
6
Alexandre Frias Faria, Sérgio Ricardo de Souza and Elisangela Martins de Sá Computers and Operations Research 127 (2021) 105133

Table 10
Summary of the results using Memetic Algorithm from Pop and Matei (2013) for k ¼ 2.

n 50 100 200 300 400 500

Minimum 5.02 4.24 1.93 1.66 1.6 1.01
1 st Quartile 307.52 266.19 136.95 189.27 215.1 152.53
Median 2655.39 2717.69 2738.47 2326.14 2641.9 2241.56
Mean 14590.08 14158.73 13990.34 14205.23 10921.5 11737.91
3 rd Quartile 25874.43 29184.93 30403.83 30361.72 19352.5 21558.03
Maximum 52876.44 51822.65 49272.91 50829.11 44902.7 43763.10

difference is mainly concerned with the optimal solution found for Table 12
the instance 2 50 20c, with the runtime of 1271:12 seconds. Paired Wilcoxon signed rank test. The critical value for W at the sample size equal to
14 and a ¼ 0:05 is 25.
The results obtained by the model (MODk ) must be also com-
pared with the results for k 2 f2; 3; 4g found by the Memetic Algo- k W p-value
rithm introduced in Pop and Matei (2013). It is worth highlighting 3 51 0.4758
that this comparison is not fair because this metaheuristic was not 4 49 0.4276
reproduced in the current article. Additionally, the maximum run-
time considered in Pop and Matei (2013) is not fixed in 1800 sec-
onds, as is done in the experimental environment in Kojić (2010),
and, also, in Pop and Matei (2013), some results were obtained 6. Conclusions
with the runtime greater than 4000 seconds. Table 10 shows a
summary of the results found by the metaheuristic proposed in This paper proposed a mixed-integer linear programming for-
Pop and Matei (2013), referring to all instances for k ¼ 2. For the mulation for solving the Multidimensional Multi-Way Number
cases k 2 f3; 4g, however, the analysis must be made separately, Partitioning Problem (MDMWNPP). The model (MODk ) showed
once the results were available only for the instances with this formulation. This problem, introduced by Pop and Matei
n 2 f50; 100g, and all values of m. (2013), is a generalization of the Multidimensional Two-Way
Table 11 shows the comparison of UB values from the Memetic Number Partitioning Problem (MDTWNPP) for any number of
Algorithm and the ones from the model (MODk ) for k ¼ 2 using the partitions. Originally, MDTWNPP is defined for only two parti-
Paired Wilcoxon signed-rank test for H1 : D < 0. For each tions (k ¼ 2), and a mathematical model was originally proposed
n 2 f50; 100; 200; 300; 400; 500g, one test was performed. All the by Kojić (2010). MDMWNPP, on the other hand, is defined for
samples have a size of 35, and the critical value for this sample size k P 2, and the mathematical model proposed in the current
is 213. The results show that there is no significant difference paper is the first one introduced in the literature, according to
between the two compared methods for the group of instances the knowledge of the authors. Therefore, this is the main contri-
associated with n 2 f50; 100; 200; 300g. However, it is important bution of this paper.
to highlight that the Memetic Algorithm returns better values of To validate the proposed model, its version for k ¼ 2 is applied
objective function than the model (MODk ) for k ¼ 2 and for solving MDTWNPP, considering the set of instances introduced
n 2 f400; 500g. by Kojić (2010). The results, obtained from CPLEX solver, demon-
The results for k 2 f3; 4g found with the model (MODk ) are com- strate that the given runtime is not sufficient for the convergence
pared with the results obtained by Pop and Matei (2013) for of the mathematical model in almost every case. Only one instance
instances type a, n 2 f50; 100g, and m 2 f2; 3; 4; 5; 10; 15; 20g. reaches the optimal solution in less than 1800 seconds. On the
Table 12 shows the comparison of the UB values found using the other hand, according to Table 8, the proposed model showed com-
Paired Wilcoxon signed-rank test for H1 : D < 0. Analyzing this petitive with the model for MDTWNPP proposed by Kojić (2010)
Table, it is possible to verify that there is no statistical difference since, in 117 out of 210 instances tested, the found values are equal
between the results found for (MODk ) and the Memetic Algorithm or better than those obtained with the model (MODKojic ) from Kojić
results for these values of k. It is important to highlight that Pop (2010).
and Matei (2013) is the only article, according to the knowledge This difference in the quality of the results is because the
of the authors of this article, to address MDMWNPP by solving model (MODKojic ) is specifically built to represent MDTWNPP.
instances for k 2 f3; 4g, however, only for the instances type a. However, as Proposition 1 shown, the model (MODk ) proposed
On the other hand, our article presents, for the first time, according here has greater generality, since the model (MODKojic ) has two
to our knowledge, results for this set of instances for k 2 f3; 4g, fewer constraints for instances such that v il > 0, in contrast to
n 2 f50; 100; 200; 300; 400; 500g, m 2 f2; 3; 4; 5; 10; 15; 20g and model (MODk ), that works for v il 2 Rm . Finally, it is important
r 2 fa; b; c; d; eg, i.e., for all instances of this set. to note that the model (MODk ) has a symmetry of solutions
related to the order of the parts that must be eliminated. Addi-
Pk1
tionally, it has n variables, xki ¼ 1  j¼1 xji ; 8i 2 In , more than
Table 11 necessary to solve correctly MDMWNPP. Reducing symmetries
Paired Wilcoxon signed rank test comparing Memetic Algorithm and the MILP and variables of this model is one of the future works, as well
proposed for each n.
as the addition of the ‘‘number of coordinate digits” factor to
n W p-value the instance characterization. Thus, the increase in complexity
50 294 0.3701 caused by the number of digits could be measured. One impor-
100 429 0.9699 tant open question is how to generate new instances with more
200 392 0.8965 diverse characteristics to exploit MDMWNPP. To reproduce
300 319 0.5290
these results, all material needed to realize this work can be
400 171 0.0087
500 171 0.0087 found in the repository https://github.com/AlexandreFrias/
AlgorithmsCOR2020.
7
Alexandre Frias Faria, Sérgio Ricardo de Souza and Elisangela Martins de Sá Computers and Operations Research 127 (2021) 105133

7. Compliance and ethical standards Acknowledgements

Authors Alexandre Frias Faria, Sérgio Ricardo de Souza, and The authors would like to thank the Coordination for the
Elisangela Martins de Sá declare that they have no conflict of inter- Improvement of Higher Education Personnel (CAPES), the Minas
est. This article does not contain any studies with human partici- Gerais State Research Foundation (FAPEMIG), the National Council
pants or animals performed by any of the authors. of Technological and Scientific Development (CNPq), and the Fed-
eral Center of Technological Education of Minas Gerais (CEFET-
MG) for supporting the development of the present study. This
Declaration of Competing Interest study was financed in part by the Coordination for the Improve-
ment of Higher Education Personnel (CAPES) – Brazil – Finance
The authors declare that they have no known competing finan- Code 001.
cial interests or personal relationships that could have appeared to
influence the work reported in this paper. Appendix A. Tables with results comparing model (MODKojic )
and model (MODk ) using k ¼ 2

Table A.13
Comparison between the model (MODKojic ) and the model (MODk ) using k ¼ 2: instances n ¼ 50

Model (MODKojic ) Model (MODk ) using k ¼ 2

Instance Status UB LB time(s) Status UB LB time(s)

2_50_2a Feasible 7.504 0 1796.98 Feasible 4.448 0 1796.72
2_50_2b Feasible 112.559 0 1799.79 Feasible 111.611 0 1799.94
2_50_2c Feasible 4.467 0 1798.6 Feasible 7.513 0 1799.95
2_50_2d Feasible 0.774971 0 1799.95 Feasible 222.954 0 1798.97
2_50_2e Feasible 6.947 0 1799.95 Feasible 9.015 0 1799.94
2_50_3a Feasible 185.82 0 1799.92 Feasible 181.706 0 1799.4
2_50_3b Feasible 137.887 0 1799.27 Feasible 137.887 0 1799.95
2_50_3c Feasible 258.492 0 1799.41 Feasible 258.494 0 1799.95
2_50_3d Feasible 88.831 0 1799.95 Feasible 87.563 0 1799.17
2_50_3e Feasible 253.919 0 1799.96 Feasible 258.692 0 1799.94
2_50_4a Feasible 1208.73 0 1799.93 Feasible 1220.93 0 1799.38
2_50_4b Feasible 1279.75 0 1799.89 Feasible 1279.75 0 1799.9
2_50_4c Feasible 946.785 0 1799.97 Feasible 943.731 0 1799.96
2_50_4d Feasible 967.16 0 1799.96 Feasible 967.16 0 1799.36
2_50_4e Feasible 843.079 0 1799.96 Feasible 852.229 0 1799.96
2_50_5a Feasible 2523.86 0 1799.94 Feasible 2523.85 0 1799.5
2_50_5b Feasible 1789.99 0 1799.95 Feasible 1789.99 0 1799.94
2_50_5c Feasible 2561.65 0 1799.96 Feasible 2755.23 0 1799.95
2_50_5d Feasible 2163.26 0 1799.97 Feasible 2560.62 0 1799.5
2_50_5e Feasible 2021.72 0 1799.96 Feasible 2030.88 0 1799.93
2_50_10a Feasible 17399.4 0 1799.59 Feasible 17393.3 0 1799.03
2_50_10b Feasible 18550.1 0 1799.67 Feasible 18556.2 0 1799.64
2_50_10c Feasible 17827.7 0 1799.74 Feasible 17833.9 0 1799.74
2_50_10d Feasible 17055.1 0 1799.74 Feasible 17055.1 0 1799.19
2_50_10e Feasible 15108.7 0 1799.73 Feasible 15114.8 0 1799.74
2_50_15a Feasible 30311 10668.8 1799.44 Feasible 30311 12420.7 1798.83
2_50_15b Feasible 33499 7874.82 1799.32 Feasible 32625.3 10886.2 1799.55
2_50_15c Feasible 27787.8 11994.7 1799.65 Feasible 27787.8 12703.5 1799.61
2_50_15d Feasible 33823.9 14160.6 1799.44 Feasible 33823.9 14168.3 1798.67
2_50_15e Feasible 31718.8 14641.3 1799.4 Feasible 31718.8 13597.8 1799.58
2_50_20a Feasible 52826.3 35826.2 1799.58 Feasible 52826.3 36452.6 1799.21
2_50_20b Feasible 51917.9 40903.7 1799.77 Feasible 51917.9 39533.9 1799.51
2_50_20c Optimal 50560.9 50555.8 1271.12 Feasible 50560.9 45183.7 1799.75
2_50_20d Feasible 53956 40208.7 1799.63 Feasible 53956 40775.8 1798.73
2_50_20e Feasible 48281.5 38447.7 1799.7 Feasible 48281.5 36424.2 1799.74

8
Alexandre Frias Faria, Sérgio Ricardo de Souza and Elisangela Martins de Sá Computers and Operations Research 127 (2021) 105133

Table A.14
Comparison between the model (MODKojic ) and the model (MODk ) using k ¼ 2: instances n ¼ 100

Model (MODKojic ) Model (MODk ) using k ¼ 2

Instance Status UB LB time(s) Status UB LB time(s)

2_100_2a Feasible 3.669 0 1799.92 Feasible 4.159 0 1799.4
2_100_2b Feasible 0.925 0 1799.93 Feasible 1.079 0 1799.92
2_100_2c Feasible 2.176 0 1799.94 Feasible 6.442 0 1799.9
2_100_2d Feasible 1.343 0 1799.94 Feasible 0.509277 0 1799.3
2_100_2e Feasible 4.51 0 1799.91 Feasible 3.204 0 1799.9
2_100_3a Feasible 151.339 0 1799.93 Feasible 142.179 0 1799.48
2_100_3b Feasible 169.791 0 1799.94 Feasible 143.062 0 1799.96
2_100_3c Feasible 6.103 0 1799.95 Feasible 154.879 0 1799.93
2_100_3d Feasible 3.051 0 1799.96 Feasible 6.101 0 1799.24
2_100_3e Feasible 108.622 0 1799.93 Feasible 105.57 0 1799.94
2_100_4a Feasible 939.942 0 1799.84 Feasible 916.488 0 1799.67
2_100_4b Feasible 1330.55 0 1799.89 Feasible 1348.86 0 1799.92
2_100_4c Feasible 785.547 0 1799.93 Feasible 782.501 0 1799.92
2_100_4d Feasible 928.469 0 1799.88 Feasible 952.881 0 1799.17
2_100_4e Feasible 1241.95 0 1799.88 Feasible 1248.05 0 1799.88
2_100_5a Feasible 2844.34 0 1799.72 Feasible 2847.39 0 1799.66
2_100_5b Feasible 1908.29 0 1799.83 Feasible 1914.38 0 1799.9
2_100_5c Feasible 1282.8 0 1799.88 Feasible 1285.85 0 1799.96
2_100_5d Feasible 926.167 0 1799.92 Feasible 913.957 0 1799.21
2_100_5e Feasible 1983.71 0 1799.84 Feasible 1983.7 0 1799.91
2_100_10a Feasible 14184.2 0 1798.87 Feasible 14181.2 0 1799.36
2_100_10b Feasible 15382.8 0 1799.34 Feasible 15379.7 0 1799.59
2_100_10c Feasible 16068.8 0 1799.56 Feasible 16062.7 0 1799.6
2_100_10d Feasible 17493.3 0 1799.64 Feasible 17493.2 0 1798.27
2_100_10e Feasible 15761.7 0 1799.64 Feasible 15749.5 0 1799.57
2_100_15a Feasible 30478.6 0 1789.9 Feasible 30487.8 0 1789.28
2_100_15b Feasible 29768.4 0 1790.81 Feasible 25609.7 0 1790.78
2_100_15c Feasible 22243.7 0 1793.04 Feasible 22255.9 0 1792.58
2_100_15d Feasible 30072.6 0 1789.66 Feasible 31065.6 0 1787.15
2_100_15e Feasible 27579.4 0 1790.88 Feasible 24487.2 0 1792.21
2_100_20a Feasible 49216.5 0 1789.26 Feasible 50336.6 0 1788.78
2_100_20b Feasible 46688.4 0 1789.4 Feasible 47534.3 0 1789.4
2_100_20c Feasible 49807.5 0 1790.69 Feasible 47662.6 0 1789.38
2_100_20d Feasible 43780.6 0 1789 Feasible 43786.7 0 1788.32
2_100_20e Feasible 44792.4 0 1788.96 Feasible 43926.3 0 1789.2

9
Alexandre Frias Faria, Sérgio Ricardo de Souza and Elisangela Martins de Sá Computers and Operations Research 127 (2021) 105133

Table A.15
Comparison between the model (MODKojic ) and the model (MODk ) using k ¼ 2: instances n ¼ 200

Model (MODKojic ) Model (MODk ) using k ¼ 2

Instance Status UB LB time(s) Status UB LB time(s)

2_200_2a Feasible 1.373 0 1799.91 Feasible 5.547 0 1799.67
2_200_2b Feasible 1.368 0 1799.92 Feasible 3.214 0 1799.91
2_200_2c Feasible 0.587074 0 1799.92 Feasible 1.236 0 1799.94
2_200_2d Feasible 0.251879 0 1799.93 Feasible 1.686 0 1799
2_200_2e Feasible 0.502961 0 1799.91 Feasible 1.238 0 1799.96
2_200_3a Feasible 116.614 0 1799.95 Feasible 81.695 0 1799.72
2_200_3b Feasible 105.565 0 1799.94 Feasible 158.341 0 1799.9
2_200_3c Feasible 100.995 0 1799.93 Feasible 77.263 0 1799.9
2_200_3d Feasible 125.962 0 1799.92 Feasible 100.452 0 1798.96
2_200_3e Feasible 119.666 0 1799.89 Feasible 92.379 0 1799.89
2_200_4a Feasible 898.189 0 1799.8 Feasible 907.337 0 1799.59
2_200_4b Feasible 995.062 0 1799.72 Feasible 992.02 0 1799.91
2_200_4c Feasible 1072.39 0 1799.83 Feasible 1087.64 0 1799.76
2_200_4d Feasible 1082.54 0 1799.88 Feasible 1010.67 0 1799.19
2_200_4e Feasible 18.31 0 1799.85 Feasible 1042.52 0 1799.88
2_200_5a Feasible 1863.36 0 1799.46 Feasible 1832.84 0 1799.49
2_200_5b Feasible 1483.64 0 1799.8 Feasible 1474.48 0 1799.69
2_200_5c Feasible 1758.8 0 1799.7 Feasible 1763.96 0 1799.82
2_200_5d Feasible 1489.74 0 1799.81 Feasible 2268.29 0 1799
2_200_5e Feasible 2268.28 0 1799.79 Feasible 2265.23 0 1799.63
2_200_10a Feasible 18863.4 0 1798.98 Feasible 14272.3 0 1798.61
2_200_10b Feasible 16706.9 0 1798.89 Feasible 16694.7 0 1798.88
2_200_10c Feasible 12766.2 0 1799 Feasible 17290.2 0 1798.76
2_200_10d Feasible 12791.1 0 1799.07 Feasible 11745.5 0 1798.44
2_200_10e Feasible 16297.2 0 1798.84 Feasible 14621.6 0 1799.11
2_200_15a Feasible 28911.2 0 1791.2 Feasible 27746.6 0 1790.81
2_200_15b Feasible 29275.1 0 1789.35 Feasible 31319.6 0 1791
2_200_15c Feasible 27604.1 0 1792.91 Feasible 30788.3 0 1790.34
2_200_15d Feasible 32500.1 0 1792.13 Feasible 30035.2 0 1788.33
2_200_15e Feasible 32784.7 0 1793.49 Feasible 28660.8 0 1787.76
2_200_20a Feasible 49890 0 1795.33 Feasible 46416.1 0 1794.83
2_200_20b Feasible 47532.1 0 1794.31 Feasible 51462.9 0 1794.82
2_200_20c Feasible 50830.7 0 1795.35 Feasible 50139 0 1792.71
2_200_20d Feasible 46651.6 0 1793.38 Feasible 45235.4 0 1796.2
2_200_20e Feasible 37156.1 0 1797.55 Feasible 49838.6 0 1793.98

10
Alexandre Frias Faria, Sérgio Ricardo de Souza and Elisangela Martins de Sá Computers and Operations Research 127 (2021) 105133

Table A.16
Comparison between the model (MODKojic ) and the model (MODk ) using k ¼ 2: instances n ¼ 300

Model (MODKojic ) Model (MODk ) using k ¼ 2

Instance Status UB LB time(s) Status UB LB time(s)

2_300_2a Feasible 0.455 0 1799.93 Feasible 0.132022 0 1799.72
2_300_2b Feasible 4.578 0 1799.93 Feasible 3.67 0 1799.92
2_300_2c Feasible 2.646 0 1799.94 Feasible 2.382 0 1799.93
2_300_2d Feasible 2.131 0 1799.91 Feasible 3.508 0 1799.24
2_300_2e Feasible 0.746 0 1799.91 Feasible 2.456 0 1799.95
2_300_3a Feasible 3.049 0 1799.94 Feasible 75.294 0 1799.65
2_300_3b Feasible 90.855 0 1799.9 Feasible 164.825 0 1799.93
2_300_3c Feasible 112.914 0 1799.92 Feasible 115.962 0 1799.92
2_300_3d Feasible 125.236 0 1799.93 Feasible 88.248 0 1799.41
2_300_3e Feasible 53.453 0 1799.94 Feasible 68.117 0 1799.92
2_300_4a Feasible 1159.3 0 1799.72 Feasible 509.55 0 1799.62
2_300_4b Feasible 4.572 0 1799.97 Feasible 1040.59 0 1799.63
2_300_4c Feasible 900.614 0 1799.82 Feasible 816.459 0 1799.8
2_300_4d Feasible 13.731 0 1799.91 Feasible 960.002 0 1799.11
2_300_4e Feasible 1034.2 0 1799.88 Feasible 1104.97 0 1799.8
2_300_5a Feasible 2805.02 0 1799.46 Feasible 1909.82 0 1799.68
2_300_5b Feasible 1866.07 0 1799.48 Feasible 2047.72 0 1799.24
2_300_5c Feasible 2744.77 0 1799.49 Feasible 1918.96 0 1799.58
2_300_5d Feasible 1844.71 0 1799.4 Feasible 2226.06 0 1799.13
2_300_5e Feasible 7.629 0 1799.81 Feasible 2230.65 0 1799.45
2_300_10a Feasible 11863.5 0 1781.75 Feasible 11842.1 0 1799.01
2_300_10b Feasible 19635.1 0 1798.83 Feasible 17988.5 0 1799.02
2_300_10c Feasible 21692.7 0 1796.43 Feasible 20207.4 0 1761.17
2_300_10d Feasible 21040.9 0 1799.02 Feasible 18249.7 0 1798.48
2_300_10e Feasible 23732.6 0 1798.87 Feasible 13091.3 0 1798.71
2_300_15a Feasible 33756.7 0 1787.96 Feasible 30466.8 0 1790.37
2_300_15b Feasible 30245.3 0 1788.28 Feasible 37664.2 0 1779.76
2_300_15c Feasible 27543.1 0 1788.7 Feasible 36755 0 1783.11
2_300_15d Feasible 29571.5 0 1789.36 Feasible 35387.5 0 1779.54
2_300_15e Feasible 36320.2 0 1788.11 Feasible 30198.6 0 1780.96
2_300_20a Feasible 47270 0 1795.26 Feasible 46991.3 0 1799.12
2_300_20b Feasible 55996.3 0 1797.31 Feasible 56535.8 0 1799.08
2_300_20c Feasible 47315.5 0 1796.37 Feasible 43529.6 0 1791.61
2_300_20d Feasible 58791.4 0 1799.21 Feasible 49340 0 1795.37
2_300_20e Feasible 55596.8 0 1799.18 Feasible 50129 0 1798.97

11
Alexandre Frias Faria, Sérgio Ricardo de Souza and Elisangela Martins de Sá Computers and Operations Research 127 (2021) 105133

Table A.17
Comparison between the model (MODKojic ) and the model (MODk ) using k ¼ 2: instances n ¼ 400

Model (MODKojic ) Model (MODk ) using k ¼ 2

Instance Status UB LB time(s) Status UB LB time(s)

2_400_2a Feasible 1.972 0 1799.94 Feasible 116.347 0 1799.93
2_400_2b Feasible 3.415 0 1799.93 Feasible 2.225 0 1799.93
2_400_2c Feasible 1.571 0 1799.89 Feasible 1.133 0 1799.7
2_400_2d Feasible 1.367 0 1799.9 Feasible 1.787 0 1799.82
2_400_2e Feasible 0.715 0 1799.93 Feasible 2.886 0 1799.91
2_400_3a Feasible 118.141 0 1799.9 Feasible 75.556 0 1799.93
2_400_3b Feasible 135.906 0 1799.88 Feasible 136.046 0 1799.9
2_400_3c Feasible 6.104 0 1799.89 Feasible 91.76 0 1799.68
2_400_3d Feasible 135.912 0 1799.87 Feasible 115.328 0 1799.91
2_400_3e Feasible 149.167 0 1799.88 Feasible 78.609 0 1799.91
2_400_4a Feasible 881.653 0 1799.73 Feasible 880.218 0 1799.73
2_400_4b Feasible 891.301 0 1799.54 Feasible 912.663 0 1799.6
2_400_4c Feasible 610.67 0 1799.81 Feasible 743.438 0 1799.63
2_400_4d Feasible 1480.13 0 1799.58 Feasible 1252.97 0 1799.81
2_400_4e Feasible 514.393 0 1799.78 Feasible 819.661 0 1799.71
2_400_5a Feasible 10.686 0 1799.9 Feasible 1890.48 0 1799.42
2_400_5b Feasible 2301.44 0 1798.37 Feasible 1915.99 0 1799.27
2_400_5c Feasible 2544.21 0 1799.34 Feasible 2539.38 0 1799.12
2_400_5d Feasible 1486.69 0 1799.11 Feasible 2033.83 0 1799.32
2_400_5e Feasible 1971.49 0 1799.65 Feasible 2534.14 0 1799.35
2_400_10a Feasible 22542.5 0 1798.86 Feasible 19035.6 0 1798.91
2_400_10b Feasible 20077.6 0 1797.43 Feasible 20364.7 0 1798.84
2_400_10c Feasible 14403.6 0 1798.84 Feasible 14412.8 0 1798.28
2_400_10d Feasible 19027 0 1798.91 Feasible 18021.9 0 1798.64
2_400_10e Feasible 20773.1 0 1798.73 Feasible 20445.2 0 1798.72
2_400_15a Feasible 28656.7 0 1788.97 Feasible 29555.5 0 1788.65
2_400_15b Feasible 30357.8 0 1788.05 Feasible 32100.9 0 1789.32
2_400_15c Feasible 36823 0 1788.23 Feasible 31755.1 0 1788.83
2_400_15d Feasible 29978.8 0 1789.11 Feasible 32766.1 0 1788.23
2_400_15e Feasible 32964.4 0 1788.81 Feasible 29317.5 0 1788.55
2_400_20a Feasible 47452.9 0 1792.29 Feasible 51048.1 0 1797.88
2_400_20b Feasible 50420 0 1791.65 Feasible 47661.2 0 1797.78
2_400_20c Feasible 49353.1 0 1792.05 Feasible 56310.1 0 1798.62
2_400_20d Feasible 44643.2 0 1792.76 Feasible 52217.2 0 1798.29
2_400_20e Feasible 53079.7 0 1793.01 Feasible 45866 0 1792.36

12
Alexandre Frias Faria, Sérgio Ricardo de Souza and Elisangela Martins de Sá Computers and Operations Research 127 (2021) 105133

Table A.18
Comparison between the model (MODKojic ) and the model (MODk ) using k ¼ 2: instances n ¼ 500

Model (MODKojic ) Model (MODk ) using k ¼ 2

Instance Status UB LB time(s) Status UB LB time(s)

2_500_2a Feasible 2.59 0 1799.94 Feasible 5.909 0 1799.9
2_500_2b Feasible 2.902 0 1798.2 Feasible 4.342 0 1799.89
2_500_2c Feasible 4.435 0 1799.85 Feasible 1.201 0 1799.67
2_500_2d Feasible 2.36 0 1799.92 Feasible 403.755 0 1799.93
2_500_2e Feasible 0.567 0 1799.94 Feasible 0.663 0 1799.92
2_500_3a Feasible 107.081 0 1799.88 Feasible 52.991 0 1799.93
2_500_3b Feasible 152.894 0 1798.1 Feasible 86.54 0 1799.9
2_500_3c Feasible 78.459 0 1799.9 Feasible 162.631 0 1799.48
2_500_3d Feasible 116.253 0 1799.91 Feasible 97.261 0 1799.92
2_500_3e Feasible 97.935 0 1799.88 Feasible 62.009 0 1799.92
2_500_4a Feasible 482.635 0 1799.77 Feasible 751.607 0 1799.53
2_500_4b Feasible 654.78 0 1798.37 Feasible 749.372 0 1799.73
2_500_4c Feasible 570.741 0 1799.58 Feasible 1104.98 0 1798.57
2_500_4d Feasible 1088.59 0 1799.84 Feasible 606.985 0 1799.78
2_500_4e Feasible 1225.18 0 1799.78 Feasible 701.553 0 1799.33
2_500_5a Feasible 1709.02 0 1799.75 Feasible 887.853 0 1799.3
2_500_5b Feasible 2083.46 0 1796.74 Feasible 1235.63 0 1799.62
2_500_5c Feasible 1614.87 0 1799.4 Feasible 2522.59 0 1798.42
2_500_5d Feasible 2232.3 0 1799.45 Feasible 1690.47 0 1798.9
2_500_5e Feasible 1890.48 0 1799.41 Feasible 1752.8 0 1799.26
2_500_10a Feasible 17624.1 0 1764.56 Feasible 19184.9 0 1765.61
2_500_10b Feasible 17966 0 1760.21 Feasible 19208.4 0 1768.18
2_500_10c Feasible 20602.2 0 1763.96 Feasible 19555.8 0 1763.64
2_500_10d Feasible 22524.7 0 1798.4 Feasible 19944.4 0 1797.38
2_500_10e Feasible 23812.6 0 1796.19 Feasible 15217.2 0 1798.56
2_500_15a Feasible 33077.8 0 1784.89 Feasible 31842.2 0 1789.5
2_500_15b Feasible 33210.2 0 1787.21 Feasible 32756.9 0 1788.97
2_500_15c Feasible 33929.8 0 1788.97 Feasible 32709.3 0 1788.01
2_500_15d Feasible 35875.1 0 1789.24 Feasible 34283.4 0 1789.04
2_500_15e Feasible 35861.5 0 1788.95 Feasible 26418.7 0 1788.86
2_500_20a Feasible 54786.9 0 1793.32 Feasible 47784.8 0 1792.52
2_500_20b Feasible 44374.9 0 1792.31 Feasible 53358.3 0 1793.78
2_500_20c Feasible 39986.4 0 1793.36 Feasible 49051.1 0 1791.26
2_500_20d Feasible 52962.7 0 1792.8 Feasible 52449.4 0 1792.34
2_500_20e Feasible 53636.3 0 1792.87 Feasible 47057 0 1792.47

13
Alexandre Frias Faria, Sérgio Ricardo de Souza and Elisangela Martins de Sá Computers and Operations Research 127 (2021) 105133

Appendix B. Tables with results from model (MODk ) using k‰f3; 4g

Table B.19
Experimental results using instances of Kojić (2010) and model (MODk ) for k 2 f3; 4g and n ¼ 50.

k¼3 k¼4
Instance Status UB LB Time(s) Status UB LB Time(s)
3_50_2a Feasible 108.376 0 1798.15 Feasible 759.295 0 1797.98
3_50_2b Feasible 233.444 0 1799.9 Feasible 1016.41 0 1798.67
3_50_2c Feasible 153.758 0 1799.98 Feasible 326.054 0 1799.67
3_50_2d Feasible 275.211 0 1799.92 Feasible 817.197 0 1799.72
3_50_2e Feasible 199.015 0 1799.75 Feasible 788.127 0 1799.58
3_50_3a Feasible 1727.11 0 1799.88 Feasible 4561.99 0 1799.56
3_50_3b Feasible 1619.99 0 1799.93 Feasible 4671.89 0 1799.44
3_50_3c Feasible 1637.95 0 1799.76 Feasible 3653.37 0 1799.2
3_50_3d Feasible 1531.66 0 1799.64 Feasible 3810.06 0 1799.6
3_50_3e Feasible 1174.1 0 1799.63 Feasible 6097.69 0 1799.54
3_50_4a Feasible 4843.93 0 1799.8 Feasible 7864.45 0 1799.58
3_50_4b Feasible 4811.94 0 1799.6 Feasible 12383.4 0 1798.68
3_50_4c Feasible 4512.36 0 1799.6 Feasible 10605.7 0 1798.49
3_50_4d Feasible 3918.02 0 1799.8 Feasible 12398.4 0 1799.51
3_50_4e Feasible 6685.06 0 1799.33 Feasible 10761 0 1799.62
3_50_5a Feasible 10242.3 0 1799.61 Feasible 14442.1 0 1799.64
3_50_5b Feasible 11453.3 0 1799.65 Feasible 17474.4 0 1798.78
3_50_5c Feasible 10644.5 0 1799.68 Feasible 18586.2 0 1798.66
3_50_5d Feasible 9918.96 0 1799.71 Feasible 20639 0 1799.57
3_50_5e Feasible 10939.4 0 1799.23 Feasible 19519.9 0 1799.62
3_50_10a Feasible 37497.1 0 1799.74 Feasible 52094.9 0 1799.03
3_50_10b Feasible 38845.7 0 1799.7 Feasible 55590.9 0 1797.83
3_50_10c Feasible 33154.1 0 1799.81 Feasible 47480.6 0 1797.65
3_50_10d Feasible 31332.9 0 1799.71 Feasible 56615.4 0 1798.61
3_50_10e Feasible 33244.7 0 1799.29 Feasible 58775 0 1799
3_50_15a Feasible 65740.3 0 1799 Feasible 82603.6 5409.36 1799.13
3_50_15b Feasible 67606.9 0 1798.83 Feasible 81396.8 4316.24 1798.48
3_50_15c Feasible 63972.2 0 1799.16 Feasible 87756 3380.3 1798.03
3_50_15d Feasible 65884.7 0 1799.27 Feasible 88641.8 4416 1799.01
3_50_15e Feasible 53260 0 1798.58 Feasible 84652.7 5736.9 1799.09
3_50_20a Feasible 86148.1 15876.5 1799.5 Feasible 97629.8 17292.4 1799.42
3_50_20b Feasible 81418.2 17304.7 1799.3 Feasible 102400 4438.93 1798.47
3_50_20c Feasible 88259.7 17472.9 1799.35 Feasible 108617 14259.2 1798.15
3_50_20d Feasible 86192.5 18418 1799.26 Feasible 102946 19955 1799.35
3_50_20e Feasible 87280.8 10118.4 1798.72 Feasible 104535 13843.4 1799.46

14
Alexandre Frias Faria, Sérgio Ricardo de Souza and Elisangela Martins de Sá Computers and Operations Research 127 (2021) 105133

Table B.20
Experimental results using instances of Kojić (2010) and model (MODk ) for k 2 f3; 4g and n ¼ 100.

k¼3 k¼4
Instance Status UB LB Time(s) Status UB LB Time(s)
3_100_2a Feasible 230.358 0 1727.21 Feasible 594.436 0 1799.27
3_100_2b Feasible 188.676 0 1722.81 Feasible 744.755 0 1798.59
3_100_2c Feasible 138.306 0 1723.01 Feasible 586.926 0 1797.22
3_100_2d Feasible 229.948 0 1716.48 Feasible 638.946 0 1799.32
3_100_2e Feasible 259.944 0 1749.81 Feasible 924.913 0 1760.27
3_100_3a Feasible 2121.11 0 1772.62 Feasible 3929.83 0 1799.34
3_100_3b Feasible 1671.87 0 1761.11 Feasible 4402.19 0 1798.91
3_100_3c Feasible 1202.61 0 1780.96 Feasible 3722.65 0 1797.82
3_100_3d Feasible 1171.16 0 1772.71 Feasible 3738.01 0 1799.45
3_100_3e Feasible 1980.93 0 1749.35 Feasible 3898.88 0 1799.32
3_100_4a Feasible 3493.16 0 1799.28 Feasible 9997.65 0 1799.45
3_100_4b Feasible 5293.17 0 1799.25 Feasible 7934.6 0 1799.15
3_100_4c Feasible 4503.36 0 1799.29 Feasible 9129.63 0 1797.8
3_100_4d Feasible 5347.41 0 1798.39 Feasible 7630.58 0 1799.55
3_100_4e Feasible 4933.83 0 1798.75 Feasible 8781.43 0 1799.47
3_100_5a Feasible 7234.48 0 1799.5 Feasible 15472.7 0 1799.49
3_100_5b Feasible 9195 0 1799.33 Feasible 18923.1 0 1799.23
3_100_5c Feasible 8792.01 0 1799.52 Feasible 18709 0 1798.66
3_100_5d Feasible 10214.8 0 1799.34 Feasible 15904.6 0 1799.53
3_100_5e Feasible 8241.72 0 1798.76 Feasible 17489.9 0 1798.94
3_100_10a Feasible 38760.5 0 1799.7 Feasible 51816.8 0 1798.95
3_100_10b Feasible 34028.2 0 1799.75 Feasible 56949.7 0 1798.99
3_100_10c Feasible 37067.5 0 1799.75 Feasible 48903.4 0 1798.84
3_100_10d Feasible 36873.9 0 1799.79 Feasible 44704.3 0 1799.03
3_100_10e Feasible 37968.2 0 1799.17 Feasible 51129.9 0 1799.73
3_100_15a Feasible 58854 0 1799.8 Feasible 78459 0 1799.13
3_100_15b Feasible 62369.2 0 1799.84 Feasible 74455 0 1798.89
3_100_15c Feasible 72196.2 0 1799.77 Feasible 77818.4 0 1798.38
3_100_15d Feasible 57447.8 0 1799.73 Feasible 84866.8 0 1799.09
3_100_15e Feasible 61952.9 0 1799.46 Feasible 84472.4 0 1799.07
3_100_20a Feasible 81959.9 0 1799.22 Feasible 95612.7 0 1799.46
3_100_20b Feasible 89503.4 0 1799.31 Feasible 91895.1 0 1799.15
3_100_20c Feasible 79363.2 0 1799.24 Feasible 107789 0 1798.77
3_100_20d Feasible 84352.8 0 1799.43 Feasible 107969 0 1799.34
3_100_20e Feasible 85619.1 0 1799.07 Feasible 98138.3 0 1799.45

15
Alexandre Frias Faria, Sérgio Ricardo de Souza and Elisangela Martins de Sá Computers and Operations Research 127 (2021) 105133

Table B.21
Experimental results using instances of Kojić (2010) and model (MODk ) for k 2 f3; 4g and n ¼ 200.

k¼3 k¼4
Instance Status UB LB Time(s) Status UB LB Time(s)
3_200_2b Feasible 249.313 0 1714.96 Feasible 629.785 0 1773.3
3_200_2c Feasible 169.712 0 1712.18 Feasible 900.571 0 1768.18
3_200_2d Feasible 72.815 0 1716.7 Feasible 650.07 0 1796.61
3_200_2e Feasible 69.019 0 1714.56 Feasible 351.272 0 1795.5
3_200_3a Feasible 1408.54 0 1736.42 Feasible 2032.06 0 1799.11
3_200_3b Feasible 1409.48 0 1736.63 Feasible 5030.79 0 1794.29
3_200_3c Feasible 1839.59 0 1736.47 Feasible 3206.77 0 1798.66
3_200_3d Feasible 1865.9 0 1750.44 Feasible 3086.72 0 1799.36
3_200_3e Feasible 2925.48 0 1735.87 Feasible 2980.31 0 1799.28
3_200_4a Feasible 3680.86 0 1789.42 Feasible 9780.88 0 1798.97
3_200_4b Feasible 5473.03 0 1767.65 Feasible 8490.88 0 1798.76
3_200_4c Feasible 5273.1 0 1771.79 Feasible 7903.95 0 1798.84
3_200_4d Feasible 2875.36 0 1770.78 Feasible 5749.82 0 1799.36
3_200_4e Feasible 3988.26 0 1769.89 Feasible 9247.18 0 1799.48
3_200_5a Feasible 8274.47 0 1799.13 Feasible 14249.7 0 1799.27
3_200_5b Feasible 8489.2 0 1799.23 Feasible 14030.2 0 1798.04
3_200_5c Feasible 10119.1 0 1799.21 Feasible 10720.3 0 1799.14
3_200_5d Feasible 6974.41 0 1799.28 Feasible 14026.7 0 1799.16
3_200_5e Feasible 10487.1 0 1798.89 Feasible 14228.1 0 1799.65
3_200_10a Feasible 34901.8 0 1799.7 Feasible 55588.9 0 1799.06
3_200_10b Feasible 31203.6 0 1799.69 Feasible 55113.2 0 1799.34
3_200_10c Feasible 37618.4 0 1799.71 Feasible 48135.4 0 1798.47
3_200_10d Feasible 33559 0 1799.69 Feasible 49832.5 0 1799.73
3_200_10e Feasible 40131 0 1799.36 Feasible 56188.4 0 1799.74
3_200_15a Feasible 59119.6 0 1799.76 Feasible 80413.3 0 1799.52
3_200_15b Feasible 61926.6 0 1799.69 Feasible 69755.6 0 1799.69
3_200_15c Feasible 63285.8 0 1799.75 Feasible 79558.3 0 1799.21
3_200_15d Feasible 64840.9 0 1799.83 Feasible 75775.8 0 1799.13
3_200_15e Feasible 62643.9 0 1799.58 Feasible 77088.9 0 1799.74
3_200_20a Feasible 80097 0 1799.37 Feasible 100031 0 1799.6
3_200_20b Feasible 82983.8 0 1799.89 Feasible 100805 0 1799.48
3_200_20c Feasible 79399.4 0 1799.88 Feasible 106623 0 1799.21
3_200_20d Feasible 84969.7 0 1799.87 Feasible 103685 0 1799.45
3_200_20e Feasible 82492.7 0 1799.62 Feasible 102098 0 1799.7

16
Alexandre Frias Faria, Sérgio Ricardo de Souza and Elisangela Martins de Sá Computers and Operations Research 127 (2021) 105133

Table B.22
Experimental results using instances of Kojić (2010) and model (MODk ) for k 2 f3; 4g and n ¼ 300.

k¼3 k¼4
Instance Status UB LB Time(s) Status UB LB Time(s)
3_300_2a Feasible 103.03 0 1799.47 Feasible 313.148 0 1799.11
3_300_2b Feasible 317.844 0 1715.54 Feasible 666.654 0 1798.94
3_300_2c Feasible 138.287 0 1706.92 Feasible 351.477 0 1756.37
3_300_2d Feasible 153.483 0 1706.58 Feasible 823.647 0 1762.54
3_300_2e Feasible 138.752 0 1714.41 Feasible 269.92 0 1799.27
3_300_3a Feasible 1471.47 0 1735.59 Feasible 2440.1 0 1799.04
3_300_3b Feasible 1444.62 0 1735.26 Feasible 3096.27 0 1798.88
3_300_3c Feasible 1739.72 0 1737.47 Feasible 1688.65 0 1798.65
3_300_3d Feasible 1276.1 0 1733.72 Feasible 3893.89 0 1799.17
3_300_3e Feasible 1645.78 0 1729.11 Feasible 2992.1 0 1799.23
3_300_4a Feasible 3279.57 0 1774.14 Feasible 10171.6 0 1798.83
3_300_4b Feasible 4905.09 0 1765.96 Feasible 7831.64 0 1798.69
3_300_4c Feasible 4440.1 0 1771.05 Feasible 11128.2 0 1798.67
3_300_4d Feasible 4221.08 0 1769.64 Feasible 6688.7 0 1799.11
3_300_4e Feasible 5834.08 0 1764.85 Feasible 8471.32 0 1799.42
3_300_5a Feasible 9014.91 0 1799.04 Feasible 15542.9 0 1798.65
3_300_5b Feasible 8270.16 0 1799.19 Feasible 20697.6 0 1798.73
3_300_5c Feasible 9919.95 0 1797.16 Feasible 16959.8 0 1798.64
3_300_5d Feasible 10297.6 0 1799.32 Feasible 16461.2 0 1799.04
3_300_5e Feasible 9100.96 0 1797.07 Feasible 16141.3 0 1798.95
3_300_10a Feasible 37028.6 0 1799.61 Feasible 49298.5 0 1799.32
3_300_10b Feasible 33925.4 0 1799.74 Feasible 58197.1 0 1799.57
3_300_10c Feasible 43682.6 0 1799.68 Feasible 52429.6 0 1799.65
3_300_10d Feasible 35065.5 0 1799.67 Feasible 44749.9 0 1799.25
3_300_10e Feasible 37841.1 0 1799.47 Feasible 51797.3 0 1799.8
3_300_15a Feasible 61105.6 0 1799.75 Feasible 79954 0 1798.38
3_300_15b Feasible 58358.4 0 1799.61 Feasible 87728.4 0 1799.7
3_300_15c Feasible 64863.9 0 1799.77 Feasible 79070.6 0 1799.25
3_300_15d Feasible 67186.6 0 1799.71 Feasible 67247.7 0 1799.49
3_300_15e Feasible 53469.3 0 1799.64 Feasible 73494.7 0 1799.91
3_300_20a Feasible 89533.1 0 1799.83 Feasible 113213 0 1799.03
3_300_20b Feasible 96176.2 0 1799.82 Feasible 97957.2 0 1799.36
3_300_20c Feasible 83349.4 0 1799.8 Feasible 95144.3 0 1799.59
3_300_20d Feasible 79232 0 1799.28 Feasible 113188 0 1799.56
3_300_20e Feasible 88667.9 0 1799.43 Feasible 106096 0 1799.55

17
Alexandre Frias Faria, Sérgio Ricardo de Souza and Elisangela Martins de Sá Computers and Operations Research 127 (2021) 105133

Table B.23
Experimental results using instances of Kojić (2010) and model (MODk ) for k 2 f3; 4g and n ¼ 400.

k¼3 k¼4
Instance Status UB LB Time(s) Status UB LB Time(s)
3_400_2a Feasible 93.905 0 1799.23 Feasible 601.744 0 1798.69
3_400_2b Feasible 100.099 0 1799.21 Feasible 690.419 0 1799.01
3_400_2c Feasible 158.381 0 1799.21 Feasible 254.95 0 1799.11
3_400_2d Feasible 182.881 0 1799.39 Feasible 476.441 0 1799.11
3_400_2e Feasible 327.772 0 1799.19 Feasible 666.861 0 1799.08
3_400_3a Feasible 2062.86 0 1798.86 Feasible 3472.03 0 1798.38
3_400_3b Feasible 1747.87 0 1799.29 Feasible 4373.07 0 1798.71
3_400_3c Feasible 1755.22 0 1799.33 Feasible 2495.82 0 1799.15
3_400_3d Feasible 1153.8 0 1799.14 Feasible 2883.38 0 1799.18
3_400_3e Feasible 1545.31 0 1799.12 Feasible 3577.91 0 1799.03
3_400_4a Feasible 6917.87 0 1799.11 Feasible 8596.1 0 1798.42
3_400_4b Feasible 5651.3 0 1799.2 Feasible 10012.5 0 1798.45
3_400_4c Feasible 3366 0 1799.19 Feasible 7912.11 0 1798.93
3_400_4d Feasible 5661.21 0 1799.07 Feasible 9518.91 0 1799.06
3_400_4e Feasible 5973.58 0 1799.02 Feasible 10377.7 0 1799.08
3_400_5a Feasible 9103.47 0 1799.13 Feasible 14301.2 0 1797.72
3_400_5b Feasible 8814.49 0 1799.11 Feasible 15513.2 0 1798.98
3_400_5c Feasible 9366.14 0 1799.16 Feasible 16039.5 0 1798.87
3_400_5d Feasible 9344.93 0 1798.64 Feasible 14145.1 0 1798.72
3_400_5e Feasible 8186.98 0 1799.06 Feasible 15633.4 0 1798.83
3_400_10a Feasible 32611.9 0 1799.63 Feasible 53364 0 1798.87
3_400_10b Feasible 33280.4 0 1799.54 Feasible 51815.4 0 1799.15
3_400_10c Feasible 37574 0 1799.51 Feasible 43721.2 0 1799.78
3_400_10d Feasible 35355.6 0 1799.28 Feasible 44755.5 0 1799.77
3_400_10e Feasible 38396.5 0 1799.37 Feasible 54871.1 0 1799.13
3_400_15a Feasible 65681.9 0 1799.8 Feasible 77390.6 0 1798.72
3_400_15b Feasible 62965.4 0 1799.42 Feasible 86242.4 0 1799.96
3_400_15c Feasible 58118.1 0 1799.79 Feasible 82750.8 0 1799.32
3_400_15d Feasible 68745.8 0 1799.64 Feasible 79647.6 0 1799.64
3_400_15e Feasible 63347.5 0 1799.62 Feasible 84181.6 0 1800.01
3_400_20a Feasible 84340.9 0 1799.92 Feasible 111878 0 1798.73
3_400_20b Feasible 85859.4 0 1799.52 Feasible 99060.8 0 1799.79
3_400_20c Feasible 87625.7 0 1799.73 Feasible 112145 0 1800
3_400_20d Feasible 72477.2 0 1799.74 Feasible 97210.7 0 1800.05
3_400_20e Feasible 91552.3 0 1799.39 Feasible 106559 0 1799.77

18
Alexandre Frias Faria, Sérgio Ricardo de Souza and Elisangela Martins de Sá Computers and Operations Research 127 (2021) 105133

Table B.24
Experimental results using instances of Kojić (2010) and model (MODk ) for k 2 f3; 4g and n ¼ 500.

k¼3 k¼4
Instance Status UB LB Time(s) Status UB LB Time(s)
3_500_2a Feasible 116.497 0 1799.14 Feasible 425.876 0 1797.96
3_500_2b Feasible 160.295 0 1799.19 Feasible 329.433 0 1799.21
3_500_2c Feasible 224.893 0 1799.18 Feasible 367.294 0 1799.02
3_500_2d Feasible 61.047 0 1798.9 Feasible 393.62 0 1799.05
3_500_2e Feasible 39.714 0 1799.06 Feasible 440.448 0 1799.02
3_500_3a Feasible 1429.18 0 1799.28 Feasible 4278.56 0 1797.39
3_500_3b Feasible 1688.88 0 1799.23 Feasible 3192.41 0 1799.11
3_500_3c Feasible 1020.23 0 1799.09 Feasible 3115.03 0 1799.01
3_500_3d Feasible 1233 0 1798.93 Feasible 4075.95 0 1799.15
3_500_3e Feasible 706.12 0 1798.84 Feasible 4192.18 0 1799.07
3_500_4a Feasible 5928.38 0 1799.18 Feasible 10749.4 0 1797.29
3_500_4b Feasible 2930.39 0 1799.15 Feasible 8320.57 0 1799.1
3_500_4c Feasible 3650.3 0 1799.16 Feasible 10950 0 1799.06
3_500_4d Feasible 3266.17 0 1798.89 Feasible 12999.8 0 1798.89
3_500_4e Feasible 4125.47 0 1798.91 Feasible 8850.88 0 1799
3_500_5a Feasible 5566.55 0 1799.06 Feasible 12879.6 0 1797.09
3_500_5b Feasible 7005.89 0 1799.07 Feasible 13007.9 0 1799.04
3_500_5c Feasible 8932.97 0 1799 Feasible 17142.6 0 1799.15
3_500_5d Feasible 7963.28 0 1798.95 Feasible 18104.2 0 1799
3_500_5e Feasible 11284 0 1798.87 Feasible 13586.2 0 1798.84
3_500_10a Feasible 44880.2 0 1799.54 Feasible 44425.6 0 1798.84
3_500_10b Feasible 38800.9 0 1799.66 Feasible 45546.1 0 1799.98
3_500_10c Feasible 36549.5 0 1798.98 Feasible 43597.3 0 1799.9
3_500_10d Feasible 32786.4 0 1799.41 Feasible 50462.5 0 1799.15
3_500_10e Feasible 42432.6 0 1799.32 Feasible 51362.8 0 1799.94
3_500_15a Feasible 70379.1 0 1799.88 Feasible 72442.5 0 1799.35
3_500_15b Feasible 65216.2 0 1799.81 Feasible 77359.9 0 1800.06
3_500_15c Feasible 66522.9 0 1799.81 Feasible 89428.7 0 1800.11
3_500_15d Feasible 67032.5 0 1799.6 Feasible 81164.6 0 1799.46
3_500_15e Feasible 59802.2 0 1799.81 Feasible 81246.5 0 1800.08
3_500_20a Feasible 76546.9 0 1799.95 Feasible 109933 0 1799.14
3_500_20b Feasible 77116.9 0 1799.49 Feasible 96878.6 0 1799.77
3_500_20c Feasible 83956 0 1799.96 Feasible 111243 0 1800
3_500_20d Feasible 83892.2 0 1799.65 Feasible 109301 0 1800
3_500_20e Feasible 89721.6 0 1800.02 Feasible 111790 0 1800.15

References Korf, R.E., Schreiber, E.L., Moffitt, M.D., 2014. Optimal sequential multi-way number
partitioning. In: International Symposium on Artificial Intelligence and
Mathematics, ISAIM 2014, Fort Lauderdale, FL, USA, January 6–8, 2014. URL:
Dalgaard, P., 2008. Introductory Statistics with R. Springer, New York, NY, USA.
http://www.cs.uic.edu/pub/Isaim2014/WebPreferences/ISAIM2014_Korf_etal.
Faria, A.F., de Souza, S.R., Souza, M.J.F., Silva, C.A., Coelho, V.N., 2019. A variable
pdf..
neighborhood search approach for solving the multidimensional multi-way
Kratica, J., Kojić, J., Savić, A., 2014. Two metaheuristic approaches for solving
number partitioning problem. In: Sifaleras, A., Salhi, S., Brimberg, J. (Eds.),
multidimensional two-way number partitioning problem. Comput. Oper. Res.
Variable Neighborhood Search: Proceedings of the 6th International Conference
46, 59–68. https://doi.org/10.1016/j.cor.2014.01.003.
on Variable Neighborhood Search (ICVNS 2018), Sithonia, Greece, October 4–7,
Moffitt, M.D., 2013. Search strategies for optimal multi-way number partitioning.
2018, Springer Volume 11328 of Lecture Notes in Computer Science, pp. 243–
In: Proceedings of the Twenty-Third International Joint Conference on Artificial
258..
Intelligence. AAAI Press, pp. 623–629.
Fuksz, L., Pop, P., Zelina, I., 2013. Heuristic algorithms for solving the bi-dimensional
Pop, P.C., Matei, O., 2013. A genetic algorithm approach for the multidimensional
two-way number partitioning problem. Studia Universitatis Babesß-Bolyai Serie
two-way number partitioning problem. In: Nicosia, G., Pardalos, P. (Eds.),
Informatica 58, 17–28.
Learning and Intelligent Optimization. Springer, Berlin, Heidelberg, pp. 81–86.
Gent, I.P., Walsh, T., 1998. Analysis of heuristics for number partitioning. Comput.
Pop, P.C., Matei, O., 2013. A memetic algorithm approach for solving the
Intell. 14, 430–451.
multidimensional multi-way number partitioning problem. Appl. Math.
Graham, R.L., 1969. Bounds on multiprocessing timing anomalies. SIAM J. Appl.
Model. 37, 9191–9202.
Math. 17, 416–429.
Rodriguez, F.J., Glover, F., García-Martíez, C., Martí, R., Lozano, M., 2017. Grasp with
Hacibeyoglu, M., Alaykiran, K., Acilar, A.M., Tongur, V., Ulker, E., 2018. A
exterior path-relinking and restricted local search for the multidimensional
comparative analysis of metaheuristic approaches for multidimensional two-
two-way number partitioning problem. Comput. Oper. Res. 78, 243–254.
way number partitioning problem. Arab. J. Sci. Eng. 43, 7499–7520. https://doi.
https://doi.org/10.1016/j.cor.2016.09.005.
org/10.1007/s13369-018-3155-9.
Santucci, V., Baioletti, M., Di Bari, G., Milani, A., 2019. A binary algebraic differential
Horowitz, E., Sahni, S., 1974. Computing partitions with applications to the
evolution for the multidimensional two-way number partitioning problem. In:
knapsack problem. J. ACM (JACM) 21, 277–292.
Liefooghe, A., Paquete, L. (Eds.), Proceedings of the 19th European Conference
Karmarkar, N., Karp, R.M., 1982. The Differencing Method of Set Partition. Report
on the Evolutionary Computation in Combinatorial Optimization, Springer
UCB/CSD 81/113 Computer Science Division. University of California Berkeley,
Volume 11452 of Lecture Notes in Computer Science, Cham, pp. 17–32..
CA.
Schreiber, E.L., 2014. Optimal Multi-Way Number Partitioning. Ph.D. thesis.
Karp, R.M., 1972. Reducibility among combinatorial problems. In: Miller, R.E.,
University of California Los Angeles.
Thatcher, J.W., Bohlinger, J.D. (Eds.), Proceedings of a Symposium on the
Schreiber, E.L., Korf, R.E., 2014. Cached iterative weakening for optimal multi-way
Complexity of Computer Computations. Springer US, pp. 85–103.
number partitioning. In: Proceedings of the Twenty-Eighth Annual Conference
Kojić, J., 2010. Integer linear programming model for multidimensional two-way
number partitioning problem. Comput. Math. Appl. 60, 2302–2308. 
on Artificial Intelligence (AAAI-14) Quebec City, Canada.
 
Schroeppel, R., Shamir, A., 1981. A T ¼ O 2n=2 , S ¼ O 2n=4 algorithm for certain
Korf, R.E., 1998. A complete anytime algorithm for number partitioning. Artif. Intell.
NP-complete problems. SIAM J. Comput. 10, 456–464.
106, 181–203.
Korf, R.E., 2009. Multi-way number partitioning. In: Boutilier, C. (Ed.), IJCAI 2009,
Proceedings of the 21st International Joint Conference on Artificial Intelligence
July 11–17, 2009 Pasadena, California, USA, pp. 538–543.

19