Computers & Operations Research 161 (2024) 106397

Contents lists available at ScienceDirect

Computers and Operations Research

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

A matheuristic for tri-objective binary integer linear programming

Duleabom An a ,∗, Sophie N. Parragh a , Markus Sinnl a , Fabien Tricoire b
a
Institute of Production and Logistics Management/JKU Business School, Johannes Kepler University Linz, Altenberger Straße 69, 4040 Linz, Austria
b
Institute for Transport and Logistics Management, Vienna University of Economics and Business, Welthandelsplatz 1, 1020 Vienna, Austria

ARTICLE INFO ABSTRACT

Keywords: Many real-world optimisation problems involve multiple objectives. When considered concurrently, they give
Integer programming rise to a set of optimal trade-off solutions, also known as efficient solutions. These solutions have the property
Multi-objective optimisation that neither objective can be improved without deteriorating another objective. Motivated by the success of
Feasibility pump
matheuristics in the single-objective domain, we propose a linear programming-based matheuristic for tri-
Path relinking
objective binary integer linear programming. To achieve a high-quality approximation of the optimal set of
Matheuristic
trade-off solutions, a lower bound set is first obtained using the vector linear programming solver Bensolve.
Then, feasibility pump-based ideas in combination with path relinking are applied in novel ways so as to
obtain a high-quality upper bound set. Our matheuristic is compared to a recently suggested algorithm that
is, to the best of our knowledge, the only existing matheuristic method for tri-objective binary integer linear
programming. In an extensive computational study, we show that our method generates a better approximation
of the true Pareto front than the state-of-the-art matheuristic on a large set of tri-objective benchmark instances.
Since the developed approach starts from a potentially fractional lower bound set, it may also be used as a
primal heuristic in the context of linear relaxation-based multi-objective branch-and-bound algorithms.

1. Introduction and Gandibleux, 2007). For a recent survey on exact methods for multi-
objective integer linear programming we refer to Halffmann et al.
Many real-world optimisation problems involve several and often (2022). However, using the available generic exact methods, only
conflicting objectives related to, e.g. costs, environmental impact and comparably small instances can be solved to optimality within rea-
service quality. In a multi-objective optimisation problem, no single sonable computation times. This limitation motivates the development
feasible solution simultaneously optimises all considered objectives, of a generic matheuristic approach that can obtain a high-quality
giving rise to a set of optimal trade-off solutions known as efficient (or approximation of the true Pareto front.
Pareto optimal) solutions. Each efficient solution is the pre-image of a In branch-and-bound algorithms for single-objective mixed inte-
non-dominated point, when plotted in the space of objective values (the ger programming (MIP) problems, primal heuristics are often used
criterion space). The set of all non-dominated points is called Pareto to provide high-quality feasible solutions at an early stage, thus en-
front (see, e.g. the book of Ehrgott (2005) for an introduction to multi- abling to tighten the bound, resulting in finding an optimal solution
objective optimisation). In this study, we focus on tri-objective binary
more quickly. Hence, they have become an essential ingredient of
integer linear programming problems.
state-of-the-art solvers such as CPLEX and Gurobi. In particular, LP
To tackle a generic multi-objective integer linear programming
relaxation-based primal heuristics have shown to be of value (see
problem, several exact algorithms that find one solution per non-
e.g. Fischetti and Lodi (2003) and Fischetti et al. (2005)). Motivated
dominated point have recently been proposed. Some of these work
by their success, our matheuristic first obtains a set of extreme points
in the criterion space, such as e.g. Boland et al. (2017) and Tamby
and Vanderpooten (2021), while others work in the space of fea- of the LP relaxation referred to as a lower bound (LB) set, using the
sible solutions, the decision space, as initially suggested by Kiziltan vector linear programming solver Bensolve (Löhne and Weißing, 2017).
and Yucaoğlu (1983). The latter relies on branch-and-bound based Depending on the fractionality of the solutions associated with the
ideas. Forget et al. (2022a,b), e.g. have recently developed branch-and- LB set, we then either employ a variant of the feasibility pump (FP)
bound algorithms for problems with more than two objectives, which or a combination of FP and path relinking (PR) to generate feasible
rely on linear programming (LP) relaxation-based bound sets (Ehrgott integer solutions. We note that both FP (Fischetti and Lodi, 2003) and

∗ Corresponding author.
E-mail address: [email protected] (D. An).

https://doi.org/10.1016/j.cor.2023.106397
Received 27 March 2022; Received in revised form 4 August 2023; Accepted 17 August 2023
Available online 22 August 2023
0305-0548/© 2023 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
D. An et al. Computers and Operations Research 161 (2024) 106397

PR (Glover, 1997) are generic heuristic techniques which were initially a certain Hamming distance. Danna et al. (2005) extend the idea
developed for single-objective optimisation. of local branching and propose a relaxation induced neighbourhood
The contributions of our study are as follows: search for single-objective MIP. Relying on the observation that an
incumbent solution and a current LP relaxation solution often have
• We propose a generic LP relaxation-based matheuristic algorithm
common variable values, the method builds a promising neighbourhood
for tri-objective binary integer linear programming, combining
by fixing those variables to the common values. Further, the defined
Bensolve with FP and PR.
sub-problem of the remaining variables is solved. The sub-problem of a
• We propose a PR scheme that generates binary integer solutions
relaxation induced neighbourhood search contains fewer variables than
starting from fractional solutions.
that of local branching and is solved faster. 𝐹 𝑃 has been introduced
• We show that generating additional fractional solutions from
by Fischetti et al. (2005) for single-objective MIP. The key idea of 𝐹 𝑃
available (potentially binary integer) solutions to obtain new
is to build a sequence of roundings that possibly converges to a feasible
high-quality binary integer solutions with FP based ideas has a
solution of the given MIP. The algorithm requires a pair of points: an
positive impact.
LP feasible but possibly fractional solution and its rounding that may
• We propose two new sets of benchmark instances for multi-
be infeasible. At every iteration, the method attempts to minimise the
objective optimisation: One set consists of tri-objective facility
distance between the two points. While closing the distance between
location problems and the other set is based on instances from
them, the algorithm may find a feasible solution. To be specific, once
the well-known MIPLIB benchmark collection for single-objective
the rounded solution is obtained from the LP relaxation solution, the
problems (Gleixner et al., 2021). The instances are made available
closest solution to it is found by solving an LP. If the new LP solution
online.
is a feasible solution to the MIP, the search stops. Otherwise, the LP
• We conduct an extensive computational study showing the ef-
solution is used in the next iteration. Since the algorithm performs well
ficacy of our generic algorithm compared to the only existing
at finding feasible solutions (see e.g. Achterberg and Berthold (2007),
state-of-the-art method.
Fischetti and Salvagnin (2009) and Boland et al. (2014)), we use the
We note that the proposed algorithm is built on the matheuristic principle of 𝐹 𝑃 for our method.
designed for the multi-objective knapsack problem presented at the In-
ternational Conference on Operations Research and Enterprise Systems 2.2. Matheuristics for multi-objective optimisation
(ICORES) 2021 (An et al., 2021). Compared to the work in An et al.
(2021) we have the following new contributions: Soylu (2015) designs a matheuristic algorithm for bi-objective
mixed binary integer linear programming based on a variable neigh-
• The feasibility pump procedure is added to our algorithm and in
bourhood search (Hansen and Mladenović, 2001) and local branch-
the computational study we focus on the effect of combining the
ing (Fischetti and Lodi, 2003). The proposed method finds the ap-
best PR method from An et al. (2021) with the feasiblity pump.
proximating segments of the Pareto front and merges them at the end
• We have extended the computational study by adding another
of the search. Another matheuristic framework for bi-objective binary
performance measure and considering new benchmark problems.
integer linear programming has been proposed by Leitner et al. (2016).
• Moreover, the section on related work and the descriptions of the
The authors exploit that efficient solutions in the same region often
algorithms are expanded considerably.
have common characteristics. Informed by this fact, their algorithm
The remainder of the paper is structured as follows. Section 2 finds feasible solutions by fixing a large number of variables based
reviews related work, while Section 3 provides the basic concepts on existing solutions, similar to the relaxation induced neighbourhood
and background for multi-objective integer linear programming. Our search for single-objective optimisation. They also propose a multi-
matheuristic is described in Section 4, and computational results are objective extension of local branching. Pal and Charkhgard (2019a)
presented in Section 5. Finally, we provide concluding remarks in develop an algorithm for bi-objective binary integer linear program-
Section 6. ming combining several existing algorithms in the literature of both
single and bi-objective optimisation. Their two-stage approach starts
2. Related work with the weighted sum method (Aneja and Nair, 1979), which is an
inner approximation algorithm to obtain an LB set for a bi-objective
A matheuristic is a hybrid approach combining mathematical pro- LP. Then they use the 𝐹 𝑃 algorithm to generate feasible solutions.
gramming with metaheuristics (Boschetti et al., 2009). Although To improve the solutions obtained by the 𝐹 𝑃 , they integrate the 𝐹 𝑃 -
matheuristics have been applied successfully in the single-objective based heuristic with a local search. This method is extended in Pal and
domain (see e.g. Archetti and Speranza (2014)), comparably few contri- Charkhgard (2019b) for multi-objective MIP. The authors develop a
butions exist for multi-objective integer linear programming problems. different weighted sum method for higher-dimensional problems and
In the following, we discuss LP relaxation-based matheuristics in both use a variant of local branching (Requejo and Santos, 2017) to increase
the single and the multi-objective domains. Further, several successful the number of solutions. Since their extended algorithm, to the best of
studies using 𝑃 𝑅, the heuristic on which our method relies, in the our knowledge, is the only generic matheuristic approach that can deal
multi-objective domain are presented. with multi-objective binary integer linear problems with more than two
objectives, we compare it to our matheuristic and refer to it as feasibil-
2.1. Matheuristics for single-objective optimisation ity pump-based heuristic (𝐹 𝑃 𝐵𝐻). Gandibleux et al. (2021) propose
a primal heuristic computing an upper bound set (for minimisation
Local branching, a combination of local search and MIP is firstly problems) that can be embedded into a branch-and-bound algorithm
described by Fischetti and Lodi (2003) for single-objective MIP. In- for multi-objective binary integer linear programming. LP solutions are
spired by local search, they develop a method that can effectively first obtained by employing the 𝜖-constraint method (Haimes, 1971).
explore the neighbourhood of solutions defined by linear cuts named Thereafter, inspired by 𝐹 𝑃 , rounding is applied to the LP solution
local branching constraints. For the given current feasible solution, to find a feasible integer solution that should exist in the restricted
its neighbourhood is built by performing soft fixing that maintains a criterion space defined by two adjacent images and the projection of the
certain number of decision variables in the incumbent solution but does LP solution. The outcome set is not necessarily a close approximation
not specifically fix any of those variables. Secondly, a local branch- of the Pareto front as their goal is to obtain an upper bound set in a
ing constraint is introduced, to find a neighbouring solution within competitive computation time.

2
D. An et al. Computers and Operations Research 161 (2024) 106397

2.3. Path relinking integer solutions) in order to obtain (additional) feasible binary integer
solutions.
PR, initially introduced by Glover (1997) for single-objective prob-
lems, is a way of exploring trajectories between elite solutions in a 3. Preliminaries
neighbourhood space, and while exploring, the algorithm may find new
solutions. Based on the fundamental idea that high-quality solutions Our matheuristic is designed to solve MOBP problems. In the follow-
share common characteristics, we expect a path between solutions ing, we state a MOBP model in its general form, assuming minimisation
might yield new solutions that share significant attributes inherited of the objectives:
from the parent solutions. The procedure of PR is as follows. First an ini-
𝑚𝑖𝑛{𝐶𝑥 ∶ 𝐴𝑥 ≥ 𝑏, 𝑥 ∈ {0, 1}𝑛 }, (MOBP)
tiating and a guiding solution are chosen from a given set of solutions.
These represent a starting point and destination, respectively. Starting where 𝑥𝑗 , 𝑗 = 1, 2, … , 𝑛, is the vector of the decision variables. Further-
from the initiating solution, the algorithm builds a certain number more, we define the feasible set 𝑋 ∶= {𝐴𝑥 ≥ 𝑏, 𝑥 ∈ {0, 1}𝑛 } and its
of neighbouring solutions in each step to reach the guiding solution. image in the criterion space 𝑌 ∶= {𝐶𝑥 ∶ 𝑥 ∈ 𝑋}. The matrix 𝐶 is a 𝑝 × 𝑛
Among the created neighbouring solutions, one is selected as a new objective function matrix where 𝐶𝑘 , (𝑘 = 1, 2, … , 𝑝), is the 𝑘th row of
initiating solution towards which the algorithm moves. As the search 𝐶. In our case, 𝑝 = 3. The matrix 𝐴 is an 𝑚 × 𝑛 constraint matrix and 𝑏
proceeds, the new initiating solution receives more attributes of the is the right-hand-side vector for these constraints.
guiding solution. The search ends when the initiating solution becomes
identical to the guiding solution. As 𝑃 𝑅 requires initial solutions, it can 3.1. Pareto dominance
be intuitively applied to multi-objective optimisation problems where
its LB set is associated with a set of solutions. In multi-objective optimisation, a popular concept for determining
A comparable neighbourhood search technique that employs the the quality of a solution is 𝑃 𝑎𝑟𝑒𝑡𝑜 dominance (Ehrgott, 2005). Let 𝑦
connectedness property in multi-objective combinatorial optimisation and 𝑦′ ∈ R𝑝 we define dominance relations as follows:
has been studied (e.g. Ehrgott and Klamroth (1997) and da Silva et al.
• 𝑦 ≦ 𝑦′ , i.e. 𝑦 weakly dominates 𝑦′ if 𝑦𝑘 ≤ 𝑦′𝑘 for 𝑘 = 1, … , 𝑝,
(2004)). The neighbourhood is constructed by swapping edges in the
graphs of efficient solutions, viewed from a graph theory perspective. • 𝑦 < 𝑦′ , i.e. 𝑦 strictly dominates 𝑦′ if 𝑦𝑘 < 𝑦′𝑘 for 𝑘 = 1, … , 𝑝,
Here, adjacent refers to the ability to transform one efficient solution • 𝑦 ≤ 𝑦′ , i.e. 𝑦 dominates 𝑦′ if 𝑦𝑘 ≤ 𝑦′𝑘 and 𝑦 ≠ 𝑦′ .
into another. Then, a set of efficient solutions is considered to be
A feasible solution 𝑥 ∈ 𝑋 is called efficient (Pareto optimal) if
connected if and only if there exists a path of adjacent efficient solutions
there exists no other solution 𝑥′ ∈ 𝑋 with 𝐶𝑥′ ≤ 𝐶𝑥. A feasible
between any pair of solutions. However, Gorski et al. (2011) have
solution 𝑥 ∈ 𝑋 is called weakly efficient if there exists no 𝑥′ ∈ 𝑋 such
shown that the connectedness property does not hold for general multi-
that 𝐶𝑥′ < 𝐶𝑥. When a feasible solution 𝑥 is (weakly) efficient, 𝑦 is
objective combinatorial optimisation problems. The key distinction
called (weakly) non-dominated. The efficient set 𝑋𝐸 ⊆ 𝑋 is defined as
between our 𝑃 𝑅 approach and other methods is that in 𝑃 𝑅, we also al-
𝑋𝐸 ∶= {𝑥 ∈ 𝑋 ∶ ∄ 𝑥′ ∈ 𝑋 ∶ 𝐶𝑥′ ≤ 𝐶𝑥}, and its image is referred to as
low for infeasible solutions on the path and the neighbouring solutions the non-dominated set or Pareto front 𝑌𝑁 ∶= 𝐶𝑋𝐸 .
gradually inherit more attributes of the guiding solutions in an attempt
to reach the guiding solution within the iteration limit. 𝑃 𝑅 has been 3.2. Weighted sum method and supported solutions
used for diverse applications in combination with other metaheuristics.
For example, Basseur et al. (2005) use 𝑃 𝑅 with a genetic algorithm The weighted sum method (Aneja and Nair, 1979) is a commonly
to solve bi-objective permutation flow-shop problems. Parragh et al. used approach in solving multi-objective optimisation problems to find
(2009) integrate 𝑃 𝑅 with a variable neighbourhood search to tackle an efficient solution by scalarising objective functions with a given
bi-objective dial-a-ride problems. On the other hand, Fernandes et al. vector 𝜆 ∈ R𝑝 as shown below where 𝐶𝑘 is row 𝑘 of 𝐶.
(2021) employ a decomposition method which divides the criterion
space into a set of sub-regions using a predefined set of direction vectors ∑
𝑝
𝑚𝑖𝑛 𝜆 𝑘 𝐶𝑘 𝑥 (WS)
along with 𝑃 𝑅 to solve multi-objective combinatorial optimisation 𝑘=1
problems.
𝑠.𝑡. 𝑥 ∈ 𝑋
2.4. Our matheuristic Each optimal solution obtained through weighted sum scalarisation for
𝜆 ∈ R𝑝> ∶= {𝜆 ∈ R𝑝 ∶ 𝜆 > 0} is efficient for (MOBP). An efficient
Compared to the primal heuristic proposed in Gandibleux et al. solution 𝑥′ ∈ 𝑋𝐸 is called supported iff there exists a weighting vector
(2021) which aims to compute a good upper bound quickly, our 𝜆′ ∈ R𝑝> that allows 𝑥′ to be optimal for (WS) for 𝜆 = 𝜆′ . Otherwise,
matheuristic is designed to obtain a high-quality approximation of 𝑥′ ∈ 𝑋𝐸 is called a non-supported efficient solution.
the true Pareto front. Similar to Pal and Charkhgard (2019b), we
initially construct an LB set by solving the LP relaxation of the multi- 3.3. Lower bound set and LP relaxation
objective binary integer linear programming (MOBP) problem. How-
ever, while (Pal and Charkhgard, 2019b) use a heuristic that generates The notion of a bound set was introduced by Ehrgott and Gandibleux
a subset of the extreme points of the LB set by solving weighted sum (2007). According to the definition established by the authors, an LB
problems, we use Bensolve for that purpose, which relies on an outer set is proposed to bound a subset 𝑌 ′ ⊂ 𝑌 of feasible points. An LB set
approximation algorithm (Löhne and Weißing, 2017). Outer approxima- for 𝑌 ′ is a subset 𝐿 ⊆ R𝑝 such that
tion algorithms generate valid LB sets even if stopped early. This can be
(i) for each 𝑦 ∈ 𝑌 ′ there is some 𝑙 ∈ 𝐿 such that 𝑙 ≤ 𝑦, and
useful in the context of multi-objective branch-and-bound algorithms.
(ii) there is no pair 𝑦 ∈ 𝑌 ′ , 𝑙 ∈ 𝐿 such that 𝑦 dominates 𝑙.
Thus our matheuristic could be more suitable as primal heuristic within
multi-objective branch-and-bound algorithms compared to approaches One common way to obtain an LB set for a minimisation problem
relying on inner approximation methods like the weighted sum. Then is to solve the LP relaxation of the original problem. In this study, for
we apply a generalisation of the idea of the 𝐹 𝑃 framework (Pal and instance, binary variables become real variables bounded by zero and
Charkhgard, 2019a) and a novel 𝑃 𝑅 scheme to the solutions associ- one in the LP relaxation problem. In our case, we use the vector linear
ated with the LB set (which can be fractional solutions and binary programming solver Bensolve to generate the LB set.

3
D. An et al. Computers and Operations Research 161 (2024) 106397

4. LP relaxation-based matheuristic iterations and time, as inputs. To provide a fractional solution for the
𝐹 𝑃 process, we take the mean of two random solutions selected from
We first give a general description of our LP relaxation-based the LB solutions. Thereby the new fractional solution inherits common
matheuristic (𝐿𝑃 𝐵𝑀) before we give details of the individual building variables from its parents and receives fractional variable values (line
blocks in the following subsections. The entire algorithmic framework 8). Then, we round the LP solution 𝑥̃ to make it binary integer (line
is outlined in Algorithm 1. 9). If the rounded solution 𝑥𝑟 is feasible and does not exist in 𝑋, ̃ it
is inserted into it and the search is done for the current 𝑥̃ (lines 10–
4.1. General outline of our matheuristic 12). Otherwise, we check whether 𝑥𝑟 is an infeasible solution found
previously (lines 13–14). If it is, the Flip operation is applied to generate
In the first step of the proposed algorithm, the extreme points of another binary integer solution, which is referred to as 𝑥𝑓 𝑙𝑖𝑝 (line 15).
the LB set are obtained by running Bensolve for at most the given The operation tries to find a new feasible solution by flipping at least
one component of the current binary integer solution (see Appendix B.1
time limit, BensolveTimeLimit. In the following, we call the solutions
for the details of the operation). When no solution is found by the
associated with these extreme points LB solutions. Then, depending on
Flip operation, the algorithm stores the current solution pair and
the fractionality of the LB solutions, we choose a different heuristic to
starts a new iteration with the next candidate solution (lines 16–17).
generate feasible binary integer solutions. The fractionality represents
Otherwise, we examine 𝑥𝑓 𝑙𝑖𝑝 ’s feasibility and whether it does not exist
the ratio between the number of fractional solutions and the total
in the current solution set. If it meets these conditions, it is added to 𝑋̃
number of LB solutions. One heuristic applies a variant of 𝐹 𝑃 while the
and a new iteration begins (lines 19–21). If 𝑥𝑟 does not exist in 𝑇 𝑎𝑏𝑢,
other approach relies on a combination of 𝐹 𝑃 and 𝑃 𝑅. Both 𝐹 𝑃 and
it is added to the list and we find a new LP solution that is closest to
𝑃 𝑅 take input solutions (𝑋) ̃ which are, in the general case, fractional.
𝑥𝑟 by solving the same optimisation problem (line 24) as used by Pal
For problems which have a totally unimodular constraint matrix and Charkhgard (2019a) (see Appendix B.2). If no feasible solution is
(e.g. the multi-objective assignment problem), all LB solutions are found by solving the LP, the search for the current 𝑥̃ terminates (lines
feasible binary integer solutions. They are extreme supported efficient 25–26) and the new iteration starts with the next 𝑥. ̃ Otherwise, the
solutions. Moreover, it can also happen that by chance the fractionality algorithm continues the iteration with the newly found 𝑥̃ (lines 27–28).
of the LB solutions of an instance of a general multi-objective problem The algorithm continues unless all sets of solution pairs are used (lines
is quite low. In this case, the basic 𝐹 𝑃 cannot be used efficiently 4–29) within the given time limit, AlgoTimeLimit. After the completion
as it requires fractional solutions to generate feasible binary integer of the search, the candidate solution set 𝑋̃ may contain fractional,
solutions. Thus, we develop a 𝐹 𝑃 variant, referred to as 𝐹 𝑃 + (see dominated, and non-dominated solutions. Therefore, we remove the
Section 4.2) that creates additional fractional solutions based on the fractional solutions during the postprocessing stage, then evaluate the
current LB solutions from which the algorithm can start. These ad- dominance relation among the solutions in 𝑋̃ (line 30) to exclusively
ditional fractional solutions are then used in 𝐹 𝑃 + to generate new return the non-dominated solutions (line 31).
feasible binary integer solutions. In FPBH, a maximum number of iterations is determined by the
If more than a certain percentage of the LB solutions, referred to number of fractional values in the current LB solution. However, after
as 𝑎𝑙𝑙𝑜𝑤𝑒𝑑𝐹 𝑟𝑎𝑐𝑡𝑖𝑜𝑛𝑎𝑙𝑖𝑡𝑦, are fractional, we use the basic 𝐹 𝑃 idea at initial experiments we fix the number of iterations per candidate solu-
first to provide feasible binary integer solutions together with those LB tion to a small number, this way many LB solutions can be used within
solutions to which 𝐹 𝑃 was not applied (e.g. because of reaching a time a limited time budget. Another difference to FPBH is that we do not
limit) for the 𝑃 𝑅 algorithm. We refer to this method as feasibility pump consider the dominance relationship between the newly found solution
generic path relinking (𝐹 𝑃 𝐺𝑃 𝑅, see Section 4.3). and the solutions stored in the archive during the search. We accept the
Both 𝐹 𝑃 + and 𝐹 𝑃 𝐺𝑃 𝑅 perform a search within a specified time new solution if it is feasible and does not exist in the incumbent solution
limit, denoted as 𝐴𝑙𝑔𝑜𝑇 𝑖𝑚𝑒𝐿𝑖𝑚𝑖𝑡, and produce an approximate set of set while Pal and Charkhgard (2019a,b) only accept non-dominated
efficient solutions 𝑋̃ 𝐸 . We define 𝑋̃ 𝐸 ⊂ 𝑋 as an approximation of the solutions.
set of efficient solutions, comprising solutions that are not necessarily We describe the 𝐹 𝑃 + procedure for a tri-objective assignment prob-
efficient but do not dominate each other. lem with three tasks and agents (see the description of the benchmark
problem in Appendix A). To keep the example simple, we only consider
the feasibility of solutions ignoring the dominance relationship. The list
Algorithm 1: LP relaxation-based matheuristic 𝐿𝑃 𝐵𝑀(𝐿𝑃 ) 𝑋̃ provides the LB solutions of the problem.
1 Run Bensolve for at most BensolveTimeLimit
⎡[1 0 0 0 1 0 0 0 1]⎤
2 𝑋̃ ← Bensolve(𝐿𝑃 ) 𝑋̃ = ⎢[0 1 0 1 0 0 0 0 1]⎥
3 Compute 𝑓 𝑟𝑎𝑐𝑡𝑖𝑜𝑛𝑎𝑙𝑖𝑡𝑦 of 𝑋̃ ⎢ ⎥
⎣[0 0 1 0 1 0 1 0 0]⎦
4 if fractionality of 𝑋̃ ≤ allowedFractionality then
5 𝑋̃ 𝐸 ← 𝐹 𝑃 + (𝑋,
̃ 𝐹 𝑃𝑖𝑡𝑒𝑟 , 𝐴𝑙𝑔𝑜𝑇 𝑖𝑚𝑒𝐿𝑖𝑚𝑖𝑡) In Table 1, the first column represents the set of solutions used for
6 else the iterations. The third column (𝑥)̃ shows the new fractional solutions
7 𝑋̃ 𝐸 ←FPGPR(𝑋, ̃ 𝐹 𝑃𝑖𝑡𝑒𝑟 , 𝐺𝑃 𝑅𝑖𝑡𝑒𝑟 , 𝐴𝑙𝑔𝑜𝑇 𝑖𝑚𝑒𝐿𝑖𝑚𝑖𝑡) obtained by combining the two selected solutions. The last column
(𝑇 𝑎𝑏𝑢) stores the discovered infeasible binary integer solutions. The
8 return 𝑋𝐸 ̃
asterisk marks a feasible binary integer solution.
̃
Suppose 𝑋[1] ̃
and 𝑋[2] are chosen in the first repetition to produce a
fractional solution 𝑥. ̃
̃ The mean value of them is, 0.5× 𝑋[1]+0.5× ̃
𝑋[2] =
4.2. Feasibility pump+ (𝐹 𝑃 + ) (0.5 0.5 0 0.5 0.5 0 0 0 1). We assume its rounded solution 𝑥𝑟 is
infeasible, thus 𝑥𝑟 is stored in 𝑇 𝑎𝑏𝑢 and the LP is solved to find the
The 𝐹 𝑃 algorithm we use is based on the framework suggested closest 𝑥.
̃ Since the operation cannot find a feasible LP solution and
by Pal and Charkhgard (2019a), who extend the 𝐹 𝑃 idea to deal with returns the empty set, we start a new iteration with a new pair of
bi-objective integer linear programming. ̃
solutions. From the two solutions {𝑋[2], ̃
𝑋[3]} we obtain 𝑥̃ = (0 0.5 0.5
Algorithm 2 shows the 𝐹 𝑃 + procedure designed for problems where 0.5 0.5 0 0.5 0 0.5). Assuming its rounded value is infeasible, it is saved
the fractionality of the computed LB solutions is below the given pa- into 𝑇 𝑎𝑏𝑢. Then, we solve the LP and obtain the new 𝑥̃ = (0.8 0.6 0.4 1
rameter 𝑎𝑙𝑙𝑜𝑤𝑒𝑑𝐹 𝑟𝑎𝑐𝑡𝑖𝑜𝑛𝑎𝑙𝑖𝑡𝑦. The algorithm keeps track of used pairs 1 0 0.3 0.2 1) whose rounded value already exists in 𝑇 𝑎𝑏𝑢. Therefore,
of LB solutions {𝑥1 , 𝑥2 } and newly found infeasible solutions during the Flip operation is conducted and we find the feasible solution (1 0 0
the search by maintaining two lists: 𝑆𝑜𝑙𝑃 𝑎𝑖𝑟 and 𝑇 𝑎𝑏𝑢. It takes the 0 0 1 0 1 0). After adding this solution to 𝑋̃ the search continues until
LB solutions 𝑋̃ obtained from Bensolve, the limits for the number of it meets the stopping conditions.

4
D. An et al. Computers and Operations Research 161 (2024) 106397

̃ 𝐹 𝑃𝑖𝑡𝑒𝑟 , 𝐴𝑙𝑔𝑜𝑇 𝑖𝑚𝑒𝐿𝑖𝑚𝑖𝑡)

Algorithm 2: Feasibility pump+ 𝐹 𝑃 + (𝑋, run until 𝑆𝑖 becomes identical to 𝑆𝑔 within the given iteration limit
(lines 10–21). If a fractional solution is used as a guiding solution,
1 𝑆𝑜𝑙𝑃 𝑎𝑖𝑟: set for storing used 𝑥1 -𝑥2 pairs
the algorithm cannot terminate the search. Accordingly, we design
2 𝑇 𝑎𝑏𝑢: list for storing infeasible solutions found during the
the algorithm to start a new iteration when all neighbouring solutions
search
created have already been used as 𝑆𝑖 . To generate the neighbouring
3 𝑇 𝑎𝑏𝑢 = ∅; 𝑆𝑜𝑙𝑃 𝑎𝑖𝑟 = ∅
solutions, we compare two solutions, 𝑆𝑖 and 𝑆𝑔 , and find out those
4 while time ≤ AlgoTimeLimit do
indices where the decision variables differ (line 11). This information is
5 𝑖𝑡𝑒𝑟 = 0
stored in the list 𝑑𝑖𝑓 to be used in the 𝐶𝑟𝑒𝑎𝑡𝑒𝑁𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠 operation. How
6 Choose 𝑥1 , 𝑥2 from 𝑋̃ at random
neighbouring solutions are created is described in Appendix C.1 in de-
7 while {𝑥1 , 𝑥2 } ∉ 𝑆𝑜𝑙𝑃 𝑎𝑖𝑟 & iter ≤ 𝐹 𝑃𝑖𝑡𝑒𝑟 do
tail. If no neighbouring solution is built, the algorithm stops and starts
8 𝑥̃ = 0.5 × 𝑥1 + 0.5 × 𝑥2
another iteration (lines 13–14). Otherwise, it checks the feasibility of
9 𝑥𝑟 ← Round(𝑥) ̃
the neighbouring solutions and archives feasible ones in 𝑐𝑎𝑛𝑑𝑋 (lines
10 if 𝑥𝑟 is feasible & 𝑥𝑟 ∉ 𝑋̃ then
15–18). Among these neighbouring solutions, we choose the next 𝑆𝑖 in
11 𝑋̃ = 𝑋̃ ∪ {𝑥𝑟 }
the 𝑁𝑒𝑥𝑡𝑆𝑖 operation (line 19), which is explained in Appendix C.2.
12 break
Then, the output of 𝑁𝑒𝑥𝑡𝑆𝑖 is inserted into 𝑇 𝑎𝑏𝑢𝑆𝑖 to prevent the
13 else
algorithm from taking the same solution in the neighbourhood space
14 if 𝑥𝑟 ∈ 𝑇 𝑎𝑏𝑢 then
(line 20). At the end of every iteration, the current [𝑆𝑖 𝑖𝑑, 𝑆𝑔 𝑖𝑑] pair is
15 𝑥𝑓 𝑙𝑖𝑝 ← 𝐹 𝑙𝑖𝑝(𝑇 𝑎𝑏𝑢, 𝑥,̃ 𝑥𝑟 )
stored in the list 𝐼𝐺𝑃 𝑎𝑖𝑟 (line 22). Upon termination of the search, the
16 if 𝑥𝑓 𝑙𝑖𝑝 = ∅ then
current 𝑐𝑎𝑛𝑑𝑋, may contain fractional, dominated and non-dominated
17 break
solutions. Therefore, we filter out the fractional solutions and then
18 else
evaluate the dominance relation among the solutions in 𝑐𝑎𝑛𝑑𝑋 (line
19 if 𝑥𝑓 𝑙𝑖𝑝 is feasible & 𝑥𝑓 𝑙𝑖𝑝 ∉ 𝑋̃ then
23), before returning only the non-dominated solutions (line 24).
20 𝑋̃ = 𝑋̃ ∪ {𝑥𝑓 𝑙𝑖𝑝 }
21 break
Algorithm 3: Feasibility pump generic path relinking
22 else ̃ 𝐹 𝑃𝑖𝑡𝑒𝑟 , 𝐺𝑃 𝑅𝑖𝑡𝑒𝑟 , 𝐴𝑙𝑔𝑜𝑇 𝑖𝑚𝑒𝐿𝑖𝑚𝑖𝑡)
𝐹 𝑃 𝐺𝑃 𝑅(𝑋,
23 𝑇 𝑎𝑏𝑢 = 𝑇 𝑎𝑏𝑢 ∪ {𝑥𝑟 }
24 𝑥̃ ← 𝐹 𝑖𝑛𝑑𝑛𝑒𝑤𝐿𝑃 (𝑥𝑟 ) 1 𝑋𝐹 𝑃 = ∅; 𝑇 𝑎𝑏𝑢 = ∅; pc = 0.5
25 if 𝑥̃ = ∅ then 2 while 𝑋̃ ≠ ∅ & time ≤ AlgoTimeLimit×pc do
26 break 3 𝑖𝑡𝑒𝑟 = 0
27 else 4 Select 𝑥̃ from 𝑋̃ at random
28 𝑖𝑡𝑒𝑟 = 𝑖𝑡𝑒𝑟 + 1 5 𝑋𝐹 𝑃 ←FP(𝑥, ̃ 𝑋𝐹 𝑃 , 𝑇 𝑎𝑏𝑢)

6 ̃
𝑐𝑎𝑛𝑑𝑋 = 𝑋𝐹 𝑃 ∪ 𝑋; 𝐼𝐺𝑃 𝑎𝑖𝑟 = ∅; 𝑇 𝑎𝑏𝑢𝑆𝑖 = ∅
29 𝑆𝑜𝑙𝑃 𝑎𝑖𝑟 = 𝑆𝑜𝑙𝑃 𝑎𝑖𝑟 ∪ {{𝑥1 , 𝑥2 }}
7 while time ≤ AlgoTimeLimit×pc do
30 ̃ ̃
𝑋𝐸 ← Postprocessing (𝑋) 8 𝑖𝑡𝑒𝑟 = 0
31 return 𝑋̃ 𝐸 9 Choose 𝑆𝑖 & 𝑆𝑔 from 𝑐𝑎𝑛𝑑𝑋 at random
10 while 𝑆𝑖 ≠ 𝑆𝑔 & [𝑆𝑖 𝑖𝑑,𝑆𝑔 𝑖𝑑] ∉ IGPair & 𝑖𝑡𝑒𝑟 ≤ 𝐺𝑃 𝑅𝑖𝑡𝑒𝑟 do
11 𝑑𝑖𝑓 ← index set of different variable values
12 𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠 ← CreateNeighbours(𝑆𝑖 , 𝑑𝑖𝑓 , 𝑇 𝑎𝑏𝑢𝑆𝑖 )
4.3. Feasibility pump generic path relinking (FPGPR) 13 if |𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠| = ∅ then
14 break
Regarding problems for which more than 𝑎𝑙𝑙𝑜𝑤𝑒𝑑𝐹 𝑟𝑎𝑐𝑡𝑖𝑜𝑛𝑎𝑙𝑖𝑡𝑦 of 15 else
the LB solutions are fractional solutions, we combine the generic frame- 16 for j=1,...,|𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠| do
work of PR with 𝐹 𝑃 and name it 𝐹 𝑃 𝐺𝑃 𝑅, which is described in 17 if neighbours[j] is feasible & not in 𝑐𝑎𝑛𝑑𝑋 then
Algorithm 3. To allocate equal time limits to both the 𝐹 𝑃 -part and the 18 𝑐𝑎𝑛𝑑𝑋 = 𝑐𝑎𝑛𝑑𝑋∪ {𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠[𝑗]}
𝑃 𝑅-part, half of the time limit is assigned to each. This allocation is
accomplished through the utilisation of the parameter 𝑝𝑐 = 0.5. The 𝐹 𝑃 19 𝑆𝑖 ← Next𝑆𝑖 (𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠)
used in 𝐹 𝑃 𝐺𝑃 𝑅 employs a different while loop condition and approach 20 𝑇 𝑎𝑏𝑢𝑆𝑖 = 𝑇 𝑎𝑏𝑢𝑆𝑖 ∪ {𝑆𝑖 }
for archiving new solutions compared to the version integrated into 21 𝑖𝑡𝑒𝑟 = 𝑖𝑡𝑒𝑟 + 1
𝐹 𝑃 + . Therefore, we name it 𝐹 𝑃 and provide the complete pseudocode
22 𝐼𝐺𝑃 𝑎𝑖𝑟 = 𝐼𝐺𝑃 𝑎𝑖𝑟 ∪ {[𝑆𝑖 𝑖𝑑, 𝑆𝑔 𝑖𝑑]}
in Appendix B.3.
At the first stage of 𝐹 𝑃 𝐺𝑃 𝑅 (Algorithm 3), the 𝐹 𝑃 process con- 23 𝑋̃ 𝐸 ← Postprocessing (𝑐𝑎𝑛𝑑𝑋)
tinues until there are no more solutions in the LB solutions set 𝑋̃ or it 24 return 𝑋̃ 𝐸
reaches the given time limit (lines 2–5). The LP solution (𝑥) ̃ is randomly
chosen from 𝑋̃ (line 4) and used for FP (line 5). All newly found binary
integer solutions are stored in the list 𝑋𝐹 𝑃 and passed on to 𝐺𝑃 𝑅 (lines As an illustrative example, we briefly present the 𝐺𝑃 𝑅 procedure for
7–22) in addition to the remaining LB solutions 𝑋, ̃ once the search is a tri-objective knapsack problem (see the description of the benchmark
completed. problem in Appendix A). In this example, we suppose that the initiating
𝐺𝑃 𝑅 requires an initial solution set 𝑐𝑎𝑛𝑑𝑋, which consists of the and guiding solutions are (0 1 0 0.5) and (0 0.2 1 0), respectively, and
generated binary integer solutions and unused LB solutions in the first there is one non-dominated solution in the neighbourhood for the sake
stage of the algorithm (line 6). The algorithm maintains two lists, of simplicity. We have four items to be added into a knapsack whose
namely 𝐼𝐺𝑃 𝑎𝑖𝑟 and 𝑇 𝑎𝑏𝑢𝑆𝑖 , to keep track of the used index pairs of ini- capacity is 15. The corresponding profits and weights are given in the
tiating (𝑆𝑖 ) and guiding (𝑆𝑔 ) solutions ([𝑆𝑖 𝑖𝑑, 𝑆𝑔 𝑖𝑑]) and used initiating matrix 𝑃 and 𝑊 , respectively.
solutions during the search. In the initial steps, we randomly choose 𝑆𝑖 ⎛3 5 7 2⎞
( )
and 𝑆𝑔 from 𝑐𝑎𝑛𝑑𝑋 (line 9). Once 𝑆𝑖 and 𝑆𝑔 are determined, we check 𝑃 = ⎜6 1 8 10⎟ 𝑊 = 8 6 7 4
whether the pair of solutions has been used. If not, 𝐺𝑃 𝑅 continues to ⎜ ⎟
⎝4 7 1 4⎠

5
D. An et al. Computers and Operations Research 161 (2024) 106397

Table 1
𝐹 𝑒𝑎𝑠𝑏𝑖𝑙𝑖𝑡𝑦 𝑃 𝑢𝑚𝑝+ procedure.
Set iter 𝑥̃ 𝑥𝑟 𝑥𝑓 𝑙𝑖𝑝 𝑇 𝑎𝑏𝑢
{1, 2} 1 0.5 0.5 0 0.5 0.5 0 0 0 1 1 1 0 1 1 0 0 0 1 1 1 0 1 1 0 0 0 1
2 ∅ 1 1 0 1 1 0 0 0 1
1 1 0 1 1 0 0 0 1
{2,3} 1 0 0.5 0.5 0.5 0.5 0 0.5 0 0.5 0 1 1 1 1 0 1 0 1
0 1 1 1 1 0 1 0 1
1 1 0 1 1 0 0 0 1
2 0.8 0.6 0.4 1 1 0 0.3 0.2 1 1 1 0 1 1 0 0 0 1 1 0 0 0 0 1 0 1 0*
0 1 1 1 1 0 1 0 1

Table 2
Generic Path Relinking procedure.
Initiating Guiding Neighbours
0 1 0 0.5 0 0.2 1 0 0 0 0 0.5 0 1 1 0.5* 0 1 0 1 0 1 0 0
0 1 1 0.5 0 0.2 1 0 0 0 1 0.5 0 1 1 1* 0 1 1 0
0 1 1 1 0 0.2 1 0 0 0 1 1 0 1 1 0*
0 1 1 0 0 0.2 1 0 0 0 1 0*
0 0 1 0 0 0.2 1 0 0110

In the CreateNeighbours operation, we check the common decision 5.1. Performance indicators
variable values between the two solutions. In this example, the first
item is the only in common, therefore 𝑑𝑖𝑓 =[2,3,4]. Further, by switch- Multi-objective integer linear programming based heuristics pro-
ing one value whose index is stored in 𝑑𝑖𝑓 at a time, we can generate duce an approximation of the true Pareto front. To compare approx-
a new solution. If the variable to be switched is fractional, we create imation sets, several quality indicators have been proposed in the liter-
two neighbouring solutions by changing it to 1 and 0, as the decision ature. We use two metrics to evaluate the quality of our approximation
variables in our problem are binary. Table 2 shows the outcome of the set and compare it to a reference set.
𝑁𝑒𝑥𝑡𝑆𝑖 operation. In the first row, we generate four solutions since
the initiating solution has three different decision variable values to
5.1.1. Hypervolume indicator
the guiding solution and one of them is fractional. According to the
The hypervolume (HV) indicator, introduced by Zitzler and Thiele
profit matrix 𝑃 , the corresponding objective values of each solution are
[1,5,2], [13,14,10], [7,11,11] and [5,1,7]. As all solutions are feasible (1999), measures the volume of the dominated space of all the solutions
with respect to the knapsack constraint, they are archived in 𝑐𝑎𝑛𝑑𝑋. contained in the approximation set. To calculate the dominated space,
The second solution (0 1 1 0.5) that has the most improved objective a reference point must be used. Usually, a reference point is the ‘‘worst
values becomes the new initiating solution and is saved to 𝑇 𝑎𝑏𝑢𝑆𝑖 possible’’ point in the criterion space. The computed figures of HV are
(marked with * in Table 2). In the next iteration, the first and third normalised as follows. Let 𝑌𝑁𝑘 be a set of the 𝑘th objective values of
solutions are passed to 𝑐𝑎𝑛𝑑𝑋 as the second solution is infeasible. Then, the true Pareto front and 𝑦 ∈ R𝑝 be an arbitrary non-dominated point
the second solution is selected as the next initiating solution and saved obtained from a heuristic algorithm. Then, the normalised objective
to 𝑇 𝑎𝑏𝑢𝑆𝑖 . In the third iteration, only the first solution is saved to values of the obtained point are:
𝑐𝑎𝑛𝑑𝑋 as the second solution has already been stored. Between the
𝑦𝑘 − 𝑚𝑖𝑛(𝑌𝑁𝑘 )
two solutions, the second solution has less deteriorated objective values 𝑘 = 1, … , 𝑝.
[12,9,8] than the first [9,18,5], therefore, it takes the initiating solution 𝑚𝑎𝑥(𝑌𝑁𝑘 ) − 𝑚𝑖𝑛(𝑌𝑁𝑘 )
role. One solution is created at the fourth iteration, which has not been All normalised non-dominated point values are between 0 and 1,
explored, therefore, it becomes the next initiating solution. Since the and we set the reference point to (2,2,2). Higher HV values indicate
only created solution at the fifth iteration (0 1 1 0) has been explored
a better approximation. We use the publicly available HV computing
previously, it is discarded from the neighbours and the search stops for
program provided by Fonseca et al. (2006) at http://lopez-ibanez.eu/
the current 𝑆𝑖 − 𝑆𝑔 pair.
hypervolume#intro to obtain the HV values.
5. Computational study
5.1.2. Multiplicative unary epsilon indicator
We evaluate the performance of our matheuristic approach, re- The multiplicative unary epsilon indicator, introduced by Zitzler
ferred to as 𝐿𝑃 𝐵𝑀, focusing on the comparison with 𝐹 𝑃 𝐵𝐻. Bensolve et al. (2003), represents the distance between an approximation set
is provided by Löhne and Weißing (2017) at http://www.bensolve. 𝐴 and a reference set 𝑅. The 𝜖 is defined as the minimum factor
org/. 𝐿𝑃 𝐵𝑀 is implemented in the programming language Julia. necessary to multiply set 𝑅 to make the transformed non-dominated
For the state-of-the-art algorithm, we use the Julia implementation of points associated with set 𝑅 weakly dominated by those of set 𝐴. All the
𝐹 𝑃 𝐵𝐻 (with the default setting) which is publicly available at https: objective values are normalised to [1,2]. With this metric, an indicator
//github.com/aritrasep/FPBH.jl. Since Bensolve employs the 𝐺𝐿𝑃 𝐾 value greater than or equal to 1, and a lower value indicates a better
solver to solve the mathematical models, it is used in all the algorithms. approximation set. To compute the unary epsilon values, we use the
The exact multi-objective integer linear programming solver developed
performance assessment package provided by Bleuler et al. (2003) at
by Kirlik and Sayın (2014) (𝐾𝑆), is also employed in the experiment
https://sop.tik.ee.ethz.ch/pisa/?page=assessment.php.
to obtain the true Pareto front for comparison purposes, i.e. the 𝐾𝑆
results are for reference only. All the experiments are carried out on a
Quad-core X5570 Xeon CPU @2.93 GHz with 48 GB RAM. Throughout 5.2. Benchmark problem instances
the paper, a maximisation objective function is converted into a min-
imisation one by multiplying it by −1. Therefore, every result reported The first two sets of instances, namely, the multi-objective assign-
is for a minimisation problem. ment problem (MOAP) and multi-objective knapsack problem (MOKP),
In the following, we first describe the performance measures and are the same as those on which 𝐹 𝑃 𝐵𝐻 is tested (the respective math-
benchmark problem instances used to conduct the computational ex- ematical models are provided in the Appendix A). Both instance sets
periments. Second, the results of the experiments are presented at the were generated by Kirlik and Sayın (2014) and are publicly available at
end of the section. http://home.ku.edu.tr/~moolibrary/. The other two sets of instances,

6
D. An et al. Computers and Operations Research 161 (2024) 106397

namely TOFLP and TOMIPLIB are derived from existing single and bi-
objective instances and provided for download at https://www.jku.at/
fileadmin/gruppen/132/Test_Cases/LPBM_instances.zip. More details
on each instance set are given below:

• MOAP: The set consists of 10 sub-classes containing 10 instances

(100 instances in total). It is divided into instance classes based
on the number of tasks which varies from 5 to 50 in increments
of 5.
• MOKP: The set consists of 10 sub-classes containing 10 instances
(100 instances in total). It is divided into instance classes based
on the number of items which varies from 10 to 100 in increments
of 10.
• TOFLP: These are newly generated tri-objective facility loca-
tion problem (TOFLP) instances based on the bi-objective sin-
gle source capacitated facility location problem instances intro-
duced by Gadegaard et al. (2019). The total coverage of customer
demand is introduced as the third objective function (see Ap-
pendix A.3). The set consists of 12 sub-classes containing 10
instances each (120 instances in total). It is divided into instance
Fig. 1. Generated solution points for TOFLP-n20_40-ins3.
classes based on the number of facilities which varies from 5 to
60 in increments of 5. The number of customers is double that of
facilities.
• TOMIPLIB: These are newly generated tri-objective binary inte- of 𝐹 𝑃 𝐵𝐻 and 𝐿𝑃 𝐵𝑀 as the reference set. The figures in Tables 3–
ger linear programming instances based on single-objective in- 5 show the average results over 10 test instances. All the results are
stances from the MIP library (MIPLIB 2017, https://miplib.zib. averaged over five runs for each instance.
de/) (Gleixner et al., 2021). Initially, 90 single-objective binary The results on the MOAP are provided in Table 3. Bensolve, denoted
integer linear programming instances categorised as easy (solv- by 𝐵𝐸𝑁, already finds more solutions than 𝐹 𝑃 𝐵𝐻 throughout all sub-
able within 1 h by a single-objective solver in default settings) classes while requiring far less computation time. Specifically, from the
were selected from the library. For these instances, coefficients sub-class with 15 tasks (n ≥ 15), the number of solutions of Bensolve is
for the second and third objective functions are then generated more than twice that of FPBH. Moreover, this difference increases as
by shuffling the coefficients of the original objective function the instances become larger. Both the quantity and the quality of the
(similar to Adelgren and Gupte, 2022). However, initial compu- solutions of Bensolve are better than those of 𝐹 𝑃 𝐵𝐻. For instance, they
tations showed that the resulting tri-objective instance were very reach more than 99% of the maximum HV throughout all the problem
hard. For the set TOMIPLIB used in our computational study, sets. Furthermore, the HV value increases as the problem size rises.
we selected the five instances for which Bensolve can find the By contrast, the highest HV value of 𝐹 𝑃 𝐵𝐻 is 95.56% in the second
complete set of LB solutions within 15 min. smallest problem class (n = 10), and it decreases as the problem size
increases. All the unary epsilon indicator values of Bensolve are better
5.3. Experimental results than those of 𝐹 𝑃 𝐵𝐻. As the solutions produced by Bensolve are all
supported, we aim at generating non-supported solutions using 𝐿𝑃 𝐵𝑀.
Following initial experiments, in our computational experiments, Despite the discovery of only a few non-supported solutions, their
we use BensolveTimeLimit set to 15 min. We note that no configuration contribution to enhancing the solution quality becomes noticeable in
settings are required for using Bensolve. The time limit, denoted by both measures. In fact, a set of supported efficient solutions computed
AlgoTimeLimit, has been specified as two minutes for the instance sets by Bensolve already provides a high-quality approximation of the Pareto
MOAP and MOKP, 10 min for TOFLP, and 1 h for the more challenging front, indicating that the MOAP may not be the most suitable benchmark
instances TOMIPLIB. We note that the CPU time required for Bensolve for LP relaxation-based (MOBP) heuristics.
has been accounted for in 𝐴𝑙𝑔𝑜𝑇 𝑖𝑚𝑒𝐿𝑖𝑚𝑖𝑡. Hence, the pure 𝐿𝑃 𝐵𝑀 𝐿𝑃 𝐵𝑀 results on the MOKP instances are provided in Table 4.
CPU time is given by 𝐴𝑙𝑔𝑜𝑇 𝑖𝑚𝑒𝐿𝑖𝑚𝑖𝑡 minus the CPU time needed for Although 𝐹 𝑃 𝐵𝐻 finishes computing over most of the MOKP instances
Bensolve. The parameters allowedFractionality, 𝐹 𝑃𝑖𝑡𝑒𝑟 and 𝐺𝑃 𝑅𝑖𝑡𝑒𝑟 are set and obtains more solutions for the larger sub-classes than 𝐿𝑃 𝐵𝑀, our
to 20%, 10 and 20, respectively. method finds a higher-quality approximation for all the sub-classes
We report the following results for each algorithm: fr (%) that based on both the HV and the unary epsilon indicators values. This can
represents the ratio between the number of fractional solutions and the also be visually shown in Figs. 1–2, which illustrates that the solution
total number of solutions associated with the LB set in case it varies set of 𝐿𝑃 𝐵𝑀 is relatively well-distributed compared to that of 𝐹 𝑃 𝐵𝐻.
across the instance set, the number of solutions (|𝑌 |), the CPU time (s), Table 5 shows the results on the TOFLP instances and a more
the percentage of the HV indicator value of the reference set (HV (%)), noticeable performance difference among the algorithms. Across all
and the unary epsilon indicator value (𝜖). the sub-classes, Bensolve requires significantly less computational time
To demonstrate the improvement achieved by 𝐿𝑃 𝐵𝑀, we present compared to 𝐹 𝑃 𝐵𝐻. For small instances with 𝑛 ≤ 20, 𝐹 𝑃 𝐵𝐻 takes
the results obtained using 𝐵𝑒𝑛𝑠𝑜𝑙𝑣𝑒 as well. For MOAP and TOFLP less time than 𝐿𝑃 𝐵𝑀, but it finds the fewest solutions in all sub-
problems the solver already generates a high-quality approximation classes. In particular, for the sub-class with 35 items (𝑛 = 35), the
of the Pareto front, providing (a subset of the) supported efficient number of solutions generated by 𝐹 𝑃 𝐵𝐻 starts to decrease and the
solutions. However, for instances with high fractionality (MOKP and gap between the two algorithms increases. A similar phenomenon is
TOMIPLIB), the HV and unary epsilon values are not included due observed for instances with more than 30 items (𝑛 ≥ 30) for both quality
to the limited number of feasible solutions. measures; the performance of 𝐹 𝑃 𝐵𝐻 declines, while that of 𝐿𝑃 𝐵𝑀
For the MOAP, the MOKP and the TOFLP the reference set is pro- steadily improves as the instance size grows. This time, we observe
vided by 𝐾𝑆. As the exact method 𝐾𝑆 cannot solve the TOMIPLIB that using 𝐿𝑃 𝐵𝑀 improves the solution quality in all sub-classes in
instances within a couple of hours, we take the combined solution sets comparison to using 𝐵𝑒𝑛𝑠𝑜𝑙𝑣𝑒 as a stand-alone method.

7
D. An et al. Computers and Operations Research 161 (2024) 106397

Table 3
Comparing the performance of the algorithms on MOAP for 𝑝 = 3. * indicates optimal values, best heuristic values are in bold. TL = 120 s †𝐹 𝑃 + is used as the LB solutions are
binary integer for all instances.
n fr (%) |𝑌 | CPUtime (s) HV (%) 𝜖
𝐾𝑆* 𝐹 𝑃 𝐵𝐻 𝐵𝐸𝑁 𝐿𝑃 𝐵𝑀† 𝐾𝑆 𝐹 𝑃 𝐵𝐻 𝐵𝐸𝑁 𝐿𝑃 𝐵𝑀† 𝐹 𝑃 𝐵𝐻 𝐵𝐸𝑁 𝐿𝑃 𝐵𝑀† 𝐹 𝑃 𝐵𝐻 𝐵𝐸𝑁 𝐿𝑃 𝐵𝑀†
5 0.0 14.1 6.7 7.5 13.0 0.1 0.2 ∼ 0.0 2.0 95.41 99.11 99.96 1.175 1.137 1.054
10 0.0 176.8 20.9 39.0 43.2 10.7 0.3 ∼ 0.0 2.6 95.56 99.36 99.43 1.128 1.065 1.060
15 0.0 674.9 41.0 83.1 85.2 92.5 0.9 ∼ 0.0 3.0 94.95 99.48 99.48 1.120 1.055 1.054
20 0.0 1860.5 63.4 161.3 162.6 359.1 2.5 0.2 5.3 94.42 99.54 99.54 1.118 1.046 1.046
25 0.0 3567.8 89.3 253.1 254.8 872.2 6.0 0.6 7.7 94.46 99.69 99.69 1.111 1.039 1.039
30 0.0 6181.3 140.7 379.4 380.9 1859.7 17.1 1.2 18.4 94.14 99.73 99.75 1.113 1.035 1.035
35 0.0 8972.3 161.6 501.4 504.0 3285.6 32.6 2.1 33.6 94.09 99.77 99.77 1.107 1.029 1.028
40 0.0 14 679.7 241.9 699.1 701.8 6426.0 72.0 3.8 73.7 94.14 99.76 99.76 1.105 1.028 1.028
45 0.0 17 702.2 240.5 838.0 841.2 9239.0 103.3 6.1 105.3 93.88 99.79 99.79 1.107 1.025 1.025
50 0.0 24 916.8 335.5 1034.8 1036.9 15 814.8 TL 9.7 TL 93.92 99.78 99.78 1.109 1.023 1.023

Table 4
Comparing the performance of the algorithms on MOKP for 𝑝 = 3, * indicates optimal values, best heuristic values are in bold. TL = 120 s †𝐹 𝑃 𝐺𝑃 𝑅 is used as the fractionality of
the LB set is above 20% for all instances.
n fr (%) |𝑌 | CPUtime (s) HV (%) 𝜖
𝐾𝑆* 𝐹 𝑃 𝐵𝐻 𝐵𝐸𝑁 𝐿𝑃 𝐵𝑀† 𝐾𝑆 𝐹 𝑃 𝐵𝐻 𝐵𝐸𝑁 𝐿𝑃 𝐵𝑀† 𝐹 𝑃 𝐵𝐻 𝐿𝑃 𝐵𝑀† 𝐹 𝑃 𝐵𝐻 𝐿𝑃 𝐵𝑀†
10 100.0 9.8 5.1 0.0 7.6 0.1 0.2 ∼ 0.0 1.2 95.49 97.52 1.161 1.116
20 100.0 38.0 18.2 0.0 26.7 1.0 0.3 ∼ 0.0 1.4 96.20 98.45 1.124 1.074
30 100.0 115.8 43.3 0.0 55.9 5.5 0.6 ∼ 0.0 1.7 96.12 98.21 1.095 1.064
40 99.9 311.2 96.5 0.2 102.2 23.2 1.5 0.1 2.3 96.23 98.54 1.083 1.052
50 100.0 444.2 112 0.0 117.5 40.1 2.7 0.5 3.8 96.92 98.48 1.071 1.049
60 100.0 917.1 195.5 0.1 200.8 116.0 6.4 0.1 7.3 96.91 98.65 1.083 1.041
70 99.9 1643.4 346.6 0.5 250.1 283.5 15.7 0.1 15.3 97.07 98.62 1.071 1.040
80 100.0 2295.8 441.5 0.2 315.8 440.0 32.0 0.1 31.3 97.63 98.80 1.053 1.035
90 99.9 3207.8 503.6 0.5 348.6 833.9 52.3 0.2 50.3 97.33 98.79 1.053 1.032
100 99.9 5849.0 894.2 0.8 518.3 2478.4 101.4 0.2 119.1 97.18 98.86 1.051 1.031

Table 5
Comparing the performance of the algorithms on TOFLP, * indicates optimal values, best heuristic values are in bold. TL = 600 s †𝐹 𝑃 + is used as the fractionality of the LB set
is below 20% for all instances.
n fr (%) |𝑌 | CPUtime (s) HV (%) 𝜖
𝐾𝑆* 𝐹 𝑃 𝐵𝐻 𝐵𝐸𝑁 𝐿𝑃 𝐵𝑀† 𝐾𝑆 𝐹 𝑃 𝐵𝐻 𝐵𝐸𝑁 𝐿𝑃 𝐵𝑀† 𝐹 𝑃 𝐵𝐻 𝐵𝐸𝑁 𝐿𝑃 𝐵𝑀† 𝐹 𝑃 𝐵𝐻 𝐵𝐸𝑁 𝐿𝑃 𝐵𝑀†
5 1.6 120.3 19.5 30.7 49.3 3.4 0.4 ∼ 0.0 3.3 97.59 99.573 99.730 1.130 1.058 1.053
10 0.5 1360.9 58.7 113.2 145.1 180.0 6.6 0.2 8.8 97.86 99.740 99.773 1.165 1.052 1.047
15 2.1 3617.7 92.8 232.4 325.5 788.9 53.8 1.1 55.7 97.74 99.833 99.858 1.179 1.038 1.035
20 4.0 8299.5 152.9 380.9 561.4 2752.0 272.1 3.6 273.2 98.26 99.865 99.887 1.174 1.038 1.034
25 5.2 17 502.8 182.1 592.5 832.4 8566.6 599.9 9.3 569.1 98.09 99.886 99.899 1.191 1.033 1.031
30 4.9 25 585.9 263.6 835.1 1102.2 16 891.7 TL 20.1 579.7 98.19 99.914 99.923 1.187 1.030 1.028
35 4.8 38 791.5 325.1 1068.8 1295.7 38 859.4 TL 39.1 588.8 97.80 99.918 99.924 1.199 1.028 1.027
40 5.1 54 791.4 312.1 1326.2 1552.9 88 433.7 TL 62.4 588.9 97.37 99.935 99.940 1.206 1.025 1.023
45 6.0 70 778.5 298.9 1596.5 1791.0 161 139.3 TL 105.9 590.4 96.22 99.930 99.933 1.219 1.024 1.023
50 7.7 100 187.1 257.8 1872.3 2057.6 226 236.8 TL 165.2 594.4 95.98 99.942 99.946 1.231 1.021 1.021
55 8.7 120 342.7 185.0 2089.4 2229.9 465 427.9 TL 246.0 598.4 95.94 99.946 99.949 1.233 1.023 1.020
60 7.9 148 529.8 151.8 2426.7 2512.3 804 132.5 TL 380.2 599.7 95.57 99.950 99.951 1.251 1.021 1.021

Table 6 shows the results on the TOMIPLIB instances. 𝐹 𝑃 𝐵𝐻 employs the vector linear programming solver, Bensolve, to obtain a
and 𝐿𝑃 𝐵𝑀 are competitive in every aspect for set packing instances lower bound set. Further, depending on the fractionality of the solu-
(cvs08r139,cvs16r70). For the remaining instances, 𝐹 𝑃 𝐵𝐻 finds con- tions associated with the lower bound set, different combinations of
siderably fewer solutions than 𝐿𝑃 𝐵𝑀. For neos instances in which feasibility pump-based ideas and path relinking are used to generate
𝐹 𝑃 𝐵𝐻 takes less computation time, we conducted extra experiments binary integer solutions. Our extensive computational study shows the
with 𝐿𝑃 𝐵𝑀 with a shorter time limit comparable to that of 𝐹 𝑃 𝐵𝐻: efficacy of our algorithm which outperforms the benchmark method,
1200 and 1000 s, respectively. This was done to confirm whether our 𝐹 𝑃 𝐵𝐻, generating a high-quality approximation of the true Pareto
algorithm continues to outperform 𝐹 𝑃 𝐵𝐻 when given a similar time front in the vast majority of the instances.
limit as the time required by 𝐹 𝑃 𝐵𝐻. The additional time spent is due The proposed algorithm is designed to generate integer solutions
starting from solutions associated with the lower bound set that might
to the algorithm’s completion of the current iteration within the loop.
contain fractional values. It can be applied to any multi-objective
In the experiments, we find that our method still outperforms 𝐹 𝑃 𝐵𝐻.
binary integer linear programming problem. In order to address general
These results can be found in Table 7. An overview of the solution
multi-objective integer linear programming problems, several compo-
quality of 𝐹 𝑃 𝐵𝐻 and 𝐿𝑃 𝐵𝑀 is given in Figs. 3 and 4. The comparison
nents would have to be adapted, both in the feasibility pump-based
is made by calculating the indicator values on all the instances of each procedure as well as in the path relinking approach. These involve,
problem class. e.g. the 𝐹 𝑙𝑖𝑝 operation and 𝐹 𝑖𝑛𝑑𝑛𝑒𝑤𝐿𝑃 as well as the neighbourhood
definition in the generation of the path between initiating and guiding
6. Conclusion solution. Bertacco et al. (2007) study the extension of the feasibility
pump-idea to the general integer case in the context of single objective
We have developed a linear programming-based matheuristic for tri- problems. The ideas presented there could be an interesting starting
objective binary integer linear programming. The proposed algorithm point for further research.

8
D. An et al. Computers and Operations Research 161 (2024) 106397

Table 6
Comparing the performance of the algorithms on TOMIPLIB, best heuristic values are in bold. TL = 3600 s †𝐹 𝑃 𝐺𝑃 𝑅 is used as the LB solutions are fractional for all instances.
Instance fr (%) 𝐹 𝑃 𝐵𝐻 |𝑌 | CPUtime (s) HV (%) 𝜖
𝐵𝐸𝑁 𝐿𝑃 𝐵𝑀† 𝐵𝐸𝑁 𝐹 𝑃 𝐵𝐻 𝐿𝑃 𝐵𝑀† 𝐹 𝑃 𝐵𝐻 𝐿𝑃 𝐵𝑀† 𝐹 𝑃 𝐵𝐻 𝐿𝑃 𝐵𝑀†
cvs08r139 99.9 19.8 5 19.8 472.5 3477.9 TL 94.75 91.22 1.171 1.182
cvs16r70 100.0 8.5 0 9.0 889.7 3476.6 TL 95.37 96.65 1.145 1.119
n2seq36f 99.9 33.5 2 132.8 303.5 TL TL 93.35 99.56 1.119 1.048
neos1516309 100.0 17.5 0 306.2 21.1 1272.6 TL 88.98 96.09 1.210 1.085
neos1599274 100.0 24.0 0 108.8 35.6 1032.6 TL 94.82 99.40 1.141 1.063

Table 7
Comparing the performance of the algorithms on the TOMIPLIB, best heuristic values are in bold. †𝐹 𝑃 𝐺𝑃 𝑅 is used as the LB solutions are
fractional for all instances (fr = 100%).
Instance |𝑌 | CPUtime (s) HV (%) 𝜖
𝐹 𝑃 𝐵𝐻 𝐿𝑃 𝐵𝑀† 𝐹 𝑃 𝐵𝐻 𝐿𝑃 𝐵𝑀† 𝐹 𝑃 𝐵𝐻 𝐿𝑃 𝐵𝑀† 𝐹 𝑃 𝐵𝐻 𝐿𝑃 𝐵𝑀†
neos1516309 17.5 67.5 1272.6 1220.7 88.98 94.92 1.210 1.089
neos1599274 24.0 26.0 1032.6 1035.3 94.82 97.69 1.141 1.093

Appendix A. The mathematical models of our benchmark prob-

lems

The following problems are used to compare the performance of the

algorithms.

A.1. Multi-objective assignment problem

In the well-known assignment problem, a certain number of tasks

𝑙 ∈ {1, … , 𝑛} and agents 𝑟 ∈ {1, … , 𝑛} are given. The decision variable
𝑥𝑟𝑙 takes the value of 1 if task 𝑙 is assigned to the agent and 0, otherwise.
The costs 𝑐𝑟𝑙1 , … , 𝑐𝑟𝑙𝑝 are incurred if a task is allocated to an agent. The
MOAP can be stated as follows:
∑𝑛 ∑ 𝑛
min 𝑐𝑟𝑙𝑗 𝑥𝑟𝑙 𝑗 = 1, … , 𝑝 (A.1)
𝑟=1 𝑙=1
∑𝑛
s.t. 𝑥𝑟𝑙 = 1 𝑟 = 1, … , 𝑛 (A.2)
𝑙=1
∑𝑛
𝑥𝑟𝑙 = 1 𝑙 = 1, … , 𝑛 (A.3)
𝑟=1
Fig. 2. Generated solution points for MOKP-n70-ins3. 𝑥𝑟𝑙 ∈ {0, 1} 𝑟, 𝑙 = 1, … , 𝑛. (A.4)
The goal of the MOAP is to find an optimal assignment of all the
tasks to agents while minimising the 𝑝 cost functions (A.1). Eq. (A.2)
proposes a limit whereby each agent is assigned to only one task.
For future work, we plan to tailor the algorithm to real-world Eq. (A.3) ensures that each task is assigned to one agent only.
applications such as in supply chain network design. Since the size of As the constraint matrix of AP is totally unimodular, every vertex
real-world problems is much larger, one major task could be reducing of the LP relaxation is a binary integer vector. Thus, we can naturally
the computation time by employing a learning algorithm. obtain binary integer solutions of the AP by solving the LP relaxation.

A.2. Multi-objective knapsack problem

CRediT authorship contribution statement
In the MOKP, a set of items with a certain weight 𝑤𝑟 and profit 𝑣𝑟
Duleabom An: Formal analysis, Methodology, Software, Visual- is given. A decision maker must select a subset of these items such
ization, Writing – original draft, Writing – review & editing. Sophie that the total weight does not exceed a given capacity 𝐶𝑎𝑝. The binary
N. Parragh: Data curation, Funding acquisition, Methodology, Re- decision variable 𝑥𝑟 has a value of 1, when item 𝑟 is selected for the
sources, Supervision, Writing – original draft, Writing – review & knapsack. Otherwise, 𝑥𝑟 is 0. Here, 𝐶𝑎𝑝, 𝑣𝑟 , and 𝑤𝑟 are non-negative
editing. Markus Sinnl: Data curation, Resources, Writing – original integer values. The MOKP model is stated as follows:
draft, Writing – review & editing. Fabien Tricoire: Resources, Writing ∑
𝑛

– original draft. max 𝑣𝑗𝑟 𝑥𝑟 𝑗 = 1, … , 𝑝 (A.5)

𝑟=1
∑
𝑛
Acknowledgements s.t. 𝑤𝑟 𝑥𝑟 ≤ 𝐶𝑎𝑝 (A.6)
𝑟=1

This research was funded in whole, or in part, by the Austrian 𝑥𝑟 ∈ {0, 1} 𝑟 = 1, … , 𝑛 (A.7)
Science Fund (FWF) [P 31366-NBL] and [P 35160-N]. For the purpose Eq. (A.5) represents the objective functions maximising the 𝑝 total
of open access, the author has applied a CC BY public copyright profits of selected items. Eq. (A.6) denotes the capacity constraint. The
licence to any Author Accepted Manuscript version arising from this total weight of the items stored in the knapsack cannot exceed the given
submission. capacity.

9
D. An et al. Computers and Operations Research 161 (2024) 106397

Fig. 3. Comparing the HV values by instance set (larger is better).

Fig. 4. Comparing the unary 𝜖 values by instance set (smaller is better).

A.3. Tri-objective facility location problem 𝑥𝑖𝑗 , 𝑦𝑖 , 𝑧𝑗 ∈ {0, 1}, ∀𝑖 ∈ 𝐼, 𝑗 ∈ 𝐽 (A.11)

Eq. (A.8) defines three objectives: total servicing cost, total fixed
A set 𝐼 of potential facility sites and a set 𝐽 of customers are
opening cost, and total customer demand coverage. Eq. (A.9) associates
given. When a facility is opened, a fixed opening cost 𝑓𝑖 is incurred. A
serviced customers to demand coverage. Eq. (A.10) imposes that a
servicing cost 𝑐𝑖𝑗 arises when the facility 𝑖 services a customer 𝑗 whose
customer should only be serviced by an opened facility.
demand is represented by 𝑑𝑗 . The bi-objective single source capacitated
facility location problem BO-SSCFLP minimises the total servicing
cost and fixed opening cost under the condition of each facility’s Appendix B. Feasibility pump components
capacity and customer demand. In the extended model, the TOFLP,
we consider maximising the service coverage. Hence, the capacity limit B.1. Flip operation
of facilities is eliminated from BO-SSCFLP and the total coverage of
customer demand is introduced as the third objective function to the Algorithm 4 describes the 𝐹 𝑙𝑖𝑝 operation used in 𝐿𝑃 𝐵𝑀. The opera-
TOFLP. tion tries to find a new feasible solution by flipping some components of
∑∑ ∑ ∑ the current binary integer solution that has already been found during
min( 𝑐𝑖𝑗 𝑥𝑖𝑗 , 𝑓𝑖 𝑦𝑖 , − 𝑑𝑗 𝑧𝑗 ) (A.8) the search.
𝑖∈𝐼 𝑗∈𝐽 𝑖∈𝐼 𝑗∈𝐽
∑ Algorithm 4 consists of deterministic (lines 4–12) and stochas-
s.t. 𝑥𝑖𝑗 = 𝑧𝑗 , ∀𝑗 ∈ 𝐽 (A.9) tic (lines 13–27) attempts. In the stochastic attempt, the number of
𝑖∈𝐼 components to be flipped is decided first by the 𝑅𝑎𝑛𝑑𝑜𝑚𝐵𝑒𝑡𝑤𝑒𝑒𝑛 oper-
𝑥𝑖𝑗 ≤ 𝑦𝑖 , ∀𝑖 ∈ 𝐼, 𝑗 ∈ 𝐽 (A.10) ation. After, the 𝑅𝑎𝑛𝑑𝑜𝑚𝑙𝑦𝐶ℎ𝑜𝑜𝑠𝑒 operation decides which components

10
D. An et al. Computers and Operations Research 161 (2024) 106397

are flipped. We refer to Pal and Charkhgard (2019a) for detailed Algorithm 5: 𝐹 𝑃 (𝑥,
̃ 𝑋𝐹 𝑃 , 𝑇 𝑎𝑏𝑢)
explanations.
1 while 𝑖𝑡𝑒𝑟 ≤ 𝐹 𝑃𝑖𝑡𝑒𝑟 do
Algorithm 4: 𝐹 𝑙𝑖𝑝(𝑇 𝑎𝑏𝑢, 𝑥,
̃ 𝑥𝑟 ) 2 𝑥𝑟 ← Round(𝑥) ̃
1 N = dimension of 𝑥̃ 3 if 𝑥𝑟 is feasible & 𝑥𝑟 ∉ 𝑋𝐹 𝑃 then
2 {𝑒1 , … , 𝑒𝑁 } ← sort the set 𝑗 = {1, … , 𝑁} based on the value of 4 𝑋𝐹 𝑃 = 𝑋𝐹 𝑃 ∪ {𝑥𝑟 }
|𝑥̃ 𝑗 − 𝑥𝑗𝑟 | in descending order 5 Remove 𝑥̃ from 𝑋̃
3 𝑥𝐼 = 𝑁𝑢𝑙𝑙; 𝑥̂ 𝑟 ← 𝑥𝑟 ; 𝑖 = 1 6 return 𝑋𝐹 𝑃
4 while 𝑖 ≤ 𝑁 and 𝑥𝐼 = 𝑁𝑢𝑙𝑙 do 7 else
5 if 𝑥̂ 𝑟 [𝑒𝑖 ]=1 then 8 if 𝑥𝑟 ∈ 𝑇 𝑎𝑏𝑢 then
6 𝑥̂ 𝑟 [𝑒𝑖 ] ← 0 9 ̃ 𝑥𝑟 )
𝑥𝑓 𝑙𝑖𝑝 ← 𝐹 𝑙𝑖𝑝(𝑇 𝑎𝑏𝑢, 𝑥,
7 else 10 if 𝑥𝑓 𝑙𝑖𝑝 = ∅ then
8 𝑥̂ 𝑟 [𝑒𝑖 ] ← 1 11 Remove 𝑥̃ from 𝑋̃
12 return 𝑋𝐹 𝑃
9 if 𝑥̂ 𝑟 ∉ 𝑇 𝑎𝑏𝑢 then
10 𝑥𝐼 ← 𝑥̂ 𝑟 13 else
11 else 14 if 𝑥𝑓 𝑙𝑖𝑝 is feasible & 𝑥𝑓 𝑙𝑖𝑝 ∉ 𝑋𝐹 𝑃 then
12 𝑖=𝑖+1 15 𝑋𝐹 𝑃 = 𝑋𝐹 𝑃 ∪ {𝑥𝑓 𝑙𝑖𝑝 }
16 Remove 𝑥̃ from 𝑋̃
13 if 𝑥𝐼 =𝑁𝑢𝑙𝑙 then 17 return 𝑋𝐹 𝑃
14 𝑖=1
15 while 𝑖 ≤ 𝑁 do 18 else
16 𝑥̂ 𝑟 ← 𝑥𝑟 19 𝑇 𝑎𝑏𝑢 = 𝑇 𝑎𝑏𝑢 ∪ {𝑥𝑟 }
17 𝑁𝑢𝑚 ← 𝑅𝑎𝑛𝑑𝑜𝑚𝐵𝑒𝑡𝑤𝑒𝑒𝑛(⌈𝑁∕2⌉, 𝑁 − 1) 20 𝑥̃ ← 𝐹 𝑖𝑛𝑑𝑛𝑒𝑤𝐿𝑃 (𝑥𝑟 )
18 𝐿 ← 𝑅𝑎𝑛𝑑𝑜𝑚𝑙𝑦𝐶ℎ𝑜𝑜𝑠𝑒(𝑁𝑢𝑚, {1, … , 𝑁}) 21 if 𝑥̃ = ∅ then
19 for 𝑙 ∈ 𝐿 do 22 Remove 𝑥̃ from 𝑋̃
20 if 𝑥̂ 𝑟 [𝑒𝑙 ]=1 then 23 return 𝑋𝐹 𝑃
21 𝑥̂ 𝑟 [𝑒𝑙 ] ← 0 24 else
22 else 25 go to line 26
23 𝑥̂ 𝑟 [𝑒𝑙 ] ← 1
26 𝑖𝑡𝑒𝑟 = 𝑖𝑡𝑒𝑟 + 1
24 if 𝑥̂ 𝑟 ∉ 𝑇 𝑎𝑏𝑢 then
27 return 𝑋𝐹 𝑃
25 𝑥𝐼 ← 𝑥̂ 𝑟
26 else
27 𝑖=𝑖+1
integers, then the number of neighbouring solutions is the same as the
28 return 𝑥𝐼
number of different decision variable values between 𝑆𝑖 and 𝑆𝑔 . For
each index in 𝑑𝑖𝑓 in which the indices of the differing variable values
B.2. Optimisation problem for 𝐹 𝑖𝑛𝑑𝑛𝑒𝑤𝐿𝑃 are stored, the corresponding decision variable value of 𝑆̂𝑖 is switched
one at a time. Specifically, if the current value is 1, it changes to 0,
𝐹 𝑖𝑛𝑑𝑛𝑒𝑤𝐿𝑃 updates 𝑥̃ using 𝑥𝑟 by solving the following optimisa-
tion problem, and vice versa (lines 5–8). When the variable value of the respective
index of 𝑆̂𝑖 is fractional, the value is changed to 0 and 1 such that we
⎧ 𝑛 ⎫
⎪ ∑ ∑ 𝑛
⎪ remove the fractional value and eventually generate binary solutions.
𝑥̃ = arg min ⎨ xj + (1 − x j ) ∶ x ∈ XLP ⎬ ,
⎪𝑗=1∶𝑥𝑗𝑟 =0 𝑗 ⎪ Both generated neighbouring solutions are archived in 𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠 (lines
𝑗=1∶𝑥𝑟 =1
⎩ ⎭
11–12). Among the 𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠, we remove solutions stored in 𝑇 𝑎𝑏𝑢𝑆𝑖
where 𝑛 is the dimension of 𝑥𝑟 and XLP is the feasible set of LP that have been used as 𝑆𝑖 (line 13).
relaxation solutions of MOBP. Further, x j , 𝑗 = 1, 2, … , 𝑛, is the vector
of the decision variables. We aim to find the LP solution that is closest
to the current 𝑥𝑟 by solving the optimisation problem. Algorithm 6: 𝐶𝑟𝑒𝑎𝑡𝑒𝑁𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠(𝑆𝑖 , 𝑑𝑖𝑓 , 𝑇 𝑎𝑏𝑢𝑆𝑖 )
1 𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠 = ∅
B.3. FP 2 for i in dif do
3 𝑆̂𝑖 ← 𝑆𝑖
We use this basic version of the 𝐹 𝑃 procedure in combination 4 if 𝑆̂𝑖 [𝑖] is binary then
with 𝐺𝑃 𝑅 in Section 4.3. When the algorithm encounters a return 5 if 𝑆̂𝑖 [𝑖] = 1 then
statement, which can occur on lines 6, 12, 17 and 23, it continues the 6 𝑆̂𝑖 [𝑖] ← 0
execution from line 3 in Algorithm 3. 7 else
8 𝑆̂𝑖 [𝑖] ← 1
Appendix C. Generic path relinking components
9 𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠 = 𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠 ∪ {𝑆̂𝑖 };
C.1. CreateNeighbours 10 else
11 𝑆̂𝑖 [𝑖] ← 1; 𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠 = 𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠 ∪ {𝑆̂𝑖 };
The CreateNeighbours operation is introduced to generate neighbour- 12 𝑆̂𝑖 [𝑖] ← 0; 𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠 = 𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠 ∪ {𝑆̂𝑖 };
ing solutions of the current 𝑆𝑖 . Algorithm 6 shows how we construct the 13 𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠 = 𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠 − 𝑇 𝑎𝑏𝑢𝑆𝑖
neighbourhood. Here, 𝑆𝑖 is duplicated to 𝑆̂𝑖 to maintain the original 14 return 𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠
values for subsequent iterations. First, the number of neighbouring
solutions to be generated is decided. If all the values in 𝑆𝑖 are binary

11
D. An et al. Computers and Operations Research 161 (2024) 106397

C.2. 𝑁𝑒𝑥𝑡𝑆𝑖 Ehrgott, M., Klamroth, K., 1997. Connectedness of efficient solutions in multiple criteria
combinatorial optimization. European J. Oper. Res. 97 (1), 159–166.
Fernandes, I.F., Goldbarg, E.F., Maia, S.M., Goldbarg, M.C., 2021. Multi-and many-
We note that we had designed and compared different selection
objective path-relinking: A taxonomy and decomposition approach. Comput. Oper.
rules to decide the next initiating solution in our previous work (An Res. 105370.
et al., 2021), and our findings indicate that the method of selecting Fischetti, M., Glover, F., Lodi, A., 2005. The feasibility pump. Math. Program. 104 (1),
the most improved (or least deteriorated) solution yields the best 91–104.
performance. To select the next initiating solution 𝑆𝑖 in the 𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠, Fischetti, M., Lodi, A., 2003. Local branching. Math. Program. 98 (1), 23–47.
Fischetti, M., Salvagnin, D., 2009. Feasibility pump 2.0. Math. Program. Comput. 1 (2),
we compare the improvement in the objective values of the 𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠. 201–222.
For this purpose, 𝑟𝑎𝑛𝑘_𝑡𝑎𝑏𝑙𝑒 is introduced in Algorithm 7. We rank the Fonseca, C.M., Paquete, L., López-Ibánez, M., 2006. An improved dimension-sweep
objective values of each neighbour in descending order and assign these algorithm for the hypervolume indicator. In: 2006 IEEE International Conference
rankings to the 𝑟𝑎𝑛𝑘_𝑡𝑎𝑏𝑙𝑒 (lines 4–6). The rankings of each neighbour on Evolutionary Computation. IEEE, pp. 1157–1163.
Forget, N., Gadegaard, S.L., Klamroth, K., Nielsen, L.R., Przybylski, A., 2022a. Branch-
are summed and passed to the 𝑛𝑏_𝑑𝑒𝑔𝑟𝑒𝑒 list (line 8). The neighbour
and-bound and objective branching with three or more objectives. Comput. Oper.
that has the highest degree is selected as the next 𝑆𝑖 (line 9). Res. 148, 106012.
Forget, N., Gadegaard, S.L., Nielsen, L.R., 2022b. Warm-starting lower bound set
computations for branch-and-bound algorithms for multi objective integer linear
Algorithm 7: 𝑁𝑒𝑥𝑡𝑆𝑖 (𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠) programs. European J. Oper. Res. 302 (3), 909–924.
1 rank_table: |𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠| × 𝑝 matrix Gadegaard, S.L., Nielsen, L.R., Ehrgott, M., 2019. Bi-objective branch-and-cut algorithms
based on LP relaxation and bound sets. INFORMS J. Comput. 31 (4), 790–804.
2 nb_degree : list of size |𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠|
Gandibleux, X., Gasnier, G., Hanafi, S., 2021. A primal heuristic to compute an upper
3 objnb : corresponding objective values of neighbours bound set for multi-objective 0-1 linear optimisation problems. In: Hayes, C.F.,
4 for i=1,...,|𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠| do Mannion, P., Vamplew, P. (Eds.), Proc. of the 1st Multi-Objective Decision Making
5 for j=1,...,𝑝 do Workshop (MODeM 2021).
Gleixner, A., Hendel, G., Gamrath, G., Achterberg, T., Bastubbe, M., Berthold, T.,
6 rank_table[𝑖][𝑗] = rank of 𝑜𝑏𝑗𝑛𝑏[𝑖][𝑗] in 𝑗 𝑡ℎ column
Christophel, P., Jarck, K., Koch, T., Linderoth, J., et al., 2021. MIPLIB 2017: data-
driven compilation of the 6th mixed-integer programming library. Math. Program.
7 for i=1,...,|𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠| do Comput. 13 (3), 443–490.
8 nb_degree[𝑖] = sum(rank_table 𝑖𝑡ℎ row) Glover, F., 1997. Tabu search and adaptive memory programming—advances, applica-
tions and challenges. In: Interfaces in Computer Science and Operations Research.
9 𝑆𝑖 ← highest nb_degree solution in 𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠 Springer, pp. 1–75.
10 return 𝑆𝑖 Gorski, J., Klamroth, K., Ruzika, S., 2011. Connectedness of efficient solutions in
multiple objective combinatorial optimization. J. Optim. Theory Appl. 150 (3).
Haimes, Y., 1971. On a bicriterion formulation of the problems of integrated system
References identification and system optimization. IEEE Trans. Syst. Man Cybern. 1 (3),
296–297.
Halffmann, P., Schäfer, L.E., Dächert, K., Klamroth, K., Ruzika, S., 2022. Exact
Achterberg, T., Berthold, T., 2007. Improving the feasibility pump. Discrete Optim. 4
algorithms for multiobjective linear optimization problems with integer variables:
(1), 77–86.
A state of the art survey. J. Multi-Criteria Decis. Anal. 29 (5–6), 341–363.
Adelgren, N., Gupte, A., 2022. Branch-and-bound for biobjective mixed-integer linear
Hansen, P., Mladenović, N., 2001. Variable neighborhood search: Principles and
programming. INFORMS J. Comput. 34 (2), 909–933.
applications. European J. Oper. Res. 130 (3), 449–467.
An, D., Parragh, S., Sinnl, M., Tricoire, F., 2021. A LP relaxation based matheuristic
Kirlik, G., Sayın, S., 2014. A new algorithm for generating all nondominated solutions
for multi-objective integer programming. In: Proceedings of the 10th International
of multiobjective discrete optimization problems. European J. Oper. Res. 232 (3),
Conference on Operations Research and Enterprise Systems (ICORES 2021). pp.
479–488.
88–98.
Kiziltan, G., Yucaoğlu, E., 1983. An algorithm for multiobjective zero-one linear
Aneja, Y.P., Nair, K.P., 1979. Bicriteria transportation problem. Manage. Sci. 25 (1),
programming. Manage. Sci. 29 (12), 1444–1453.
73–78.
Leitner, M., Ljubić, I., Sinnl, M., Werner, A., 2016. ILP heuristics and a new exact
Archetti, C., Speranza, M.G., 2014. A survey on matheuristics for routing problems.
method for bi-objective 0/1 ILPs: Application to FTTx-network design. Comput.
EURO J. Comput. Optim. 2 (4), 223–246.
Oper. Res. 72, 128–146.
Basseur, M., Seynhaeve, F., Talbi, E.-G., 2005. Path relinking in pareto multi-objective
Löhne, A., Weißing, B., 2017. The vector linear program solver Bensolve–notes on
genetic algorithms. In: International Conference on Evolutionary Multi-Criterion
theoretical background. European J. Oper. Res. 260 (3), 807–813.
Optimization. Springer, pp. 120–134.
Pal, A., Charkhgard, H., 2019a. A feasibility pump and local search based heuristic
Bertacco, L., Fischetti, M., Lodi, A., 2007. A feasibility pump heuristic for general
for bi-objective pure integer linear programming. INFORMS J. Comput. 31 (1),
mixed-integer problems. Discrete Optim. 4 (1), 63–76.
115–133.
Bleuler, S., Laumanns, M., Thiele, L., Zitzler, E., 2003. PISA — A platform and pro-
Pal, A., Charkhgard, H., 2019b. FPBH: A feasibility pump based heuristic for
gramming language independent interface for search algorithms. In: Fonseca, C.M.,
multi-objective mixed integer linear programming. Comput. Oper. Res. 112,
Fleming, P.J., Zitzler, E., Deb, K., Thiele, L. (Eds.), Evolutionary Multi-Criterion
104760.
Optimization (EMO 2003). In: Lecture Notes in Computer Science, Springer, Berlin,
Parragh, S.N., Doerner, K.F., Hartl, R.F., Gandibleux, X., 2009. A heuristic two-phase
pp. 494–508.
solution approach for the multi-objective dial-a-ride problem. Networks 54 (4),
Boland, N., Charkhgard, H., Savelsbergh, M., 2017. The quadrant shrinking method: A
227–242.
simple and efficient algorithm for solving tri-objective integer programs. European
Requejo, C., Santos, E., 2017. A feasibility pump and a local branching heuristics for the
J. Oper. Res. 260 (3), 873–885.
weight-constrained minimum spanning tree problem. In: International Conference
Boland, N.L., Eberhard, A.C., Engineer, F.G., Fischetti, M., Savelsbergh, M.W.,
on Computational Science and Its Applications. Springer, pp. 669–683.
Tsoukalas, A., 2014. Boosting the feasibility pump. Math. Program. Comput. 6 (3),
da Silva, C.G., Clímaco, J., Figueira, J., 2004. Geometrical Configuration of the Pareto
255–279.
Frontier of Bi-Criteria {0, 1}-Knapsack Problems. Technical Report Tech. Rep.
Boschetti, M.A., Maniezzo, V., Roffilli, M., Röhler, A.B., 2009. Matheuristics: Optimiza-
16/2004, INESC, Coimbra, Portugal.
tion, simulation and control. In: International Workshop on Hybrid Metaheuristics.
Soylu, B., 2015. Heuristic approaches for biobjective mixed 0–1 integer linear
Springer, pp. 171–177.
programming problems. European J. Oper. Res. 245 (3), 690–703.
Danna, E., Rothberg, E., Le Pape, C., 2005. Exploring relaxation induced neighborhoods
Tamby, S., Vanderpooten, D., 2021. Enumeration of the nondominated set of
to improve MIP solutions. Math. Program. 102 (1), 71–90.
multiobjective discrete optimization problems. INFORMS J. Comput. 33 (1), 72–85.
Ehrgott, M., 2005. Multicriteria Optimization, Vol. 491. Springer Science & Business
Zitzler, E., Thiele, L., 1999. Multiobjective evolutionary algorithms: a comparative case
Media.
study and the strength Pareto approach. IEEE Trans. Evol. Comput. 3 (4), 257–271.
Ehrgott, M., Gandibleux, X., 2007. Bound sets for biobjective combinatorial
Zitzler, E., Thiele, L., Laumanns, M., Fonseca, C.M., Da Fonseca, V.G., 2003. Perfor-
optimization problems. Comput. Oper. Res. 34 (9), 2674–2694.
mance assessment of multiobjective optimizers: An analysis and review. IEEE Trans.
Evol. Comput. 7 (2), 117–132.

12