Exact and heuristic algorithms for the interval data robust assignment problem
Introduction
The Assignment Problem is a classical combinatorial optimization problem whose study can be traced back to the 1930s and even earlier, see [28]. Its standard version deals with choosing an optimal way to assign n items to n other items, e.g., n tasks to n workers in the most well-known interpretation, given the costs for all possible task-worker pairs. The problem is polynomially solvable, and serves as a model in numerous application settings. We refer the interested reader to the new book by Burkard et al. [11] for further details on the problem, its algorithmic solution methods, and its variants.
We consider the case where the costs (objective function coefficients) may be uncertain. To handle uncertainty, in this paper we use the minmax regret (also known as robust deviation) approach, where it is required to find a feasible solution that is for any possible scenario (realization of the vector of the objective function coefficients), with as small as possible. Minmax regret combinatorial optimization problems have been studied extensively over the past two decades (see, e.g., [2], [4], [5], [18], [19], [21], [24], [25] and the references therein). The motivation for using the minmax regret approach as a modeling tool is thoroughly discussed in the book [21]; the book [18] and the survey [2] provide reviews of recent results.
In interval data minmax regret (IDMR) models, it is assumed that for each objective function coefficient (the cost of a task-worker pair in the context of the assignment problem), an uncertainty interval is specified, and the set of scenarios is the Cartesian product of all uncertainty intervals. In the robust optimization literature, there have been extensive studies of minmax regret versions of many classical combinatorial optimization problems (see, e.g., [18]). It has been observed that typically polynomially solvable classical problems with the minisum type of objective function become NP-hard in the IDMR version [5], [1], although there are some exceptions, see [4], [3]. The same holds for the IDMR assignment problem which was proven to be weakly NP-hard in [19] and, later, strongly NP-hard in [1]. Currently, significant research efforts are devoted to developing practically efficient computational methods for solving IDMR versions of classical combinatorial optimization problems, and there has been successful work in this direction [23], [24], [25]. However, computational work and experience for the IDMR Assignment Problem have been limited to testing a mixed-integer linear programming (MILP) formulation with the CPLEX solver, and a straightforward scenario-based heuristic, for relatively small problem sizes [19], [18].
The main purpose of this paper is to fill this gap and conduct a computational study of the IDMR Assignment Problem, with the goal of comparing several solution approaches for different problem sizes. We compare two exact solution methods: solving the MILP formulation from [19] with ILOG CPLEX MILP solver, and Benders decomposition which was used effectively for other robust combinatorial optimization problems [23], [25]. For larger problem sizes, we also test several heuristics, including a simple scenario-based heuristic, a regular local search, a variable depth neighborhood local search, and two hybrid population-based heuristics that use an optimization subproblem within the framework of a Genetic Algorithm, thus being consistent with the concept of a Memetic Algorithm. We conduct experiments with several families of problem instances generated by different methods.
We note that our hybrid population-based heuristics, to the best of our knowledge, are the first application of a population-based algorithm to a robust combinatorial optimization problem (along with [26] where a population-based approach is applied to the robust set covering problem). One of the two hybrid heuristics developed in this paper, according to our experiments, shows the best performance for large problems among all tested approaches.
The paper is organized as follows. The IDMR Assignment Problem is formally defined in Section 2, along with the main notation and known main properties. Two mathematical programming formulations, one from [19] and the other used as a basis for Benders decomposition, are stated in Section 3. In Section 4, the details of the Benders decomposition implementation are given, and in Section 5 the results of the computational experiments with the exact methods are discussed. Section 6 is devoted to the development of heuristic methods to solve the problem, and in Section 7 the results of computational experience with heuristic methods are discussed. In Section 8, conclusions are presented.
Section snippets
Notation and problem statement
We use notation and terminology developed in earlier work on minmax regret optimization [4], [5], [26]. Let be a bipartite network with the set of nodes, , and the set of edges . Suppose that for every edge , two nonnegative real numbers cuv−, cuv+ are given, . The cost of edge (u,v) can take on any real value from its uncertainty interval [cuv−, cuv+], regardless of the values taken by the costs of other edges. The Cartesian product of
Formulation 1
In the remainder of the paper, will denote not only a perfect matching in G, but also the n2-dimensional characteristic vector of this matching with components Vectors will be denoted with upper-case letters, and their components with the same lower-case letters. Then, Problem ROB.ASSIGN can be reformulated as follows:or, using Lemma 1, Introducing a free variable and
Implementation of the exact algorithms
In Section 3, two different mathematical programming formulations have been given. Current Mixed-Integer Linear Programming packages are capable of solving formulations such as (9), (10), (11), (12), (13), (14) with relative ease [10], but only up to a certain instance size. The results of applying the CPLEX solver to this formulation are discussed in Section 5.
To tackle the formulation (4), (5), (6), (7), (8) which has an exponential number of constraints (5), we use Benders decomposition [9],
Implementation details
All the algorithms in this paper have been coded in ANSI C++. Two external sources of code were used. The first one is a routine to solve any assignment problem. For this purpose, the recent C version of the code described in [12] was chosen. This code is available at http://www.assignmentproblems.com/APC_APS.htm and is an implementation of the Hungarian method for dense cost matrices, the case considered in this paper. Second, the ILOG CPLEX library, version 11.2, is used to solve linear
Heuristic procedures
The simplest heuristic for the Robust Assignment Problem is to solve Problem OPT.ASSIGN(s) for some natural scenario and present the obtained solution as an approximate solution. This takes just solving two regular Assignment Problems (one for obtaining a solution to Problem OPT.ASSIGN(s) and one for evaluating its value Z(X)). Natural choices for scenario s are s= and s+. The scenario-based heuristic that consists in solving Problem OPT.ASSIGN(s) for s=s= and s=s+ and choosing the best of the
Computational experience with heuristic methods
Section 6 outlined different heuristic approaches to the problem. In this section, we discuss the results of computational experiments with these heuristics, and the comparison with the solutions found by exact algorithms. As mentioned in Section 5.1, all algorithms have been programmed in C++ and run on the same computer. Again, each algorithm has been given a maximum computation time of 600 s as a time limit.
Table 3, Table 4, Table 5, Table 6 report the results for each family of instances
Conclusions, recommendations, and additional comments
For the interval data minmax regret assignment problem, we compared several exact and heuristic methods. Our main recommendations and findings are as follows.
- 1.
The difficulty of the problem depends not only on the instance size but also on the degree of uncertainty.
- 2.
The scenario-based heuristic obtains good solutions in a very short time (typically within a small fraction of a second for the tested problem sizes). It should be used when a good solution must be obtained very quickly, or to obtain
Acknowledgements
The research of the first author was partially supported by the Spanish CICYT Grant DPI2007-63026. The research of the second author was supported by a discovery Grant from the Natural Sciences and Engineering Research Council (NSERC) of Canada.
References (28)
- et al.
Complexity of the min-max and min-max regret assignment problems
Operations Research Letters
(2005) - et al.
Min-max and min-max regret versions of combinatorial optimization problems: a survey
European Journal of Operational Research
(2009) - et al.
Complexity of the min-max (regret) versions of min-cut problems
Discrete Optimization
(2008) On the complexity of a class of combinatorial optimization problems with uncertainty
Mathematical Programming
(2001)- et al.
Interval data minmax regret network optimization problems
Discrete Applied Mathematics
(2004) Linear programming on Cray supercomputers
Journal of the Operational Research Society
(1990)OR-Library: distributing test problems by electronic mail
Journal of the Operational Research Society
(1990)- et al.
A genetic algorithm for the set covering problem
European Journal of Operational Research
(1996) Partitioning procedures for solving mixed-variables programming problems
Numerische Mathematik
(1962)- et al.
Mixed-integer programming: a progress report
Assignment Problems
Algorithms and codes for the assignment problem
Annals of Operations Research
Generalized benders decomposition
Journal of Optimization Theory and Applications
Tabu Search
Cited by (47)
Investigating the recoverable robust single machine scheduling problem under interval uncertainty
2022, Discrete Applied MathematicsCitation Excerpt :For a general survey on robust discrete optimisation problems, we refer the reader to [14]. As a brief comment on other papers that consider robust assignment problems, Deǐneko et al. [6] study the problem in the context of discrete scenarios and Pereira and Averbakh [24] consider interval uncertainty in combination with the regret criterion. The single machine scheduling problem with the objective of minimising the (weighted) sum of completion times under uncertain job processing times is well-studied, particularly for the case of discrete uncertain scenarios.
Maximum excess dominance: Identifying impractical solutions in linear problems with interval coefficients
2020, European Journal of Operational ResearchA two-stage robust optimisation for terminal traffic flow problem
2020, Applied Soft Computing JournalGenerating hard instances for robust combinatorial optimization
2020, European Journal of Operational ResearchCitation Excerpt :In fact, where NP-hardness proofs typically rely on constructing problem instances with specific properties, nothing is known about hardness of randomly generated instances, or smoothed analysis, in robust optimization. Where the related min-max regret problem has sparked research into specialized solution algorithms (see, e.g., Catanzaro, Labbé, & Salazar-Neumann, 2011; Kasperski, Makuchowski, & Zieliński, 2012; Pereira & Averbakh, 2011), little such research exists for the min-max problem, as simply using an off-the-shelf mixed-integer programming solver, such as CPLEX, can already lead to satisfactory results. Looking into other fields of optimization problems, instance libraries have been a main driver of algorithm development (Müller-Hannemann & Schirra, 2010).
Exact and heuristic algorithms for the interval min-max regret generalized assignment problem
2018, Computers and Industrial EngineeringAlgorithms for the Minmax Regret Path problem with interval data
2018, Information Sciences