Multi-Objective Large Neighborhood Search

摘要

Large neighborhood search (LNS) [25] is a framework that combines the expressiveness of constraint programming with the efficiency of local search to solve combinatorial optimization problems. This paper introduces an extension of LNS, called multi-objective LNS (MO-LNS), to solve multi-objective combinatorial optimization problems ubiquitous in practice. The idea of MO-LNS is to maintain a set of nondominated solutions rather than just one best-so-far solution. At each iteration, one of these solutions is selected, relaxed and optimized in order to strictly improve the hypervolume of the maintained set of nondominated solutions. We introduce modeling abstractions into the OscaR solver for MO-LNS and show experimentally the efficiency of this approach on various multi-objective combinatorial optimization problems.