Reinforced Adaptive Large Neighborhood Search

The Large Neighborhood Search metaheuristic for solving Constrained Optimization Problems has been proved to be effective on a wide range of problems. This Local Search heuristic has the particularity of using a complete search (such as Constraint Programming) to explore the large neighborhoods obtained by relaxing a fragment of the variables of the current solution. Large Neighborhood Search has three parameters that must be specified (size of the fragment, search limit and fragment selection procedure). Its performances greatly depend on those parameters. Despite the success of the metaheuristic, no generic principle has emerged yet on how to choose the parameters. They are currently set either with domain dependent heuristics or chosen randomly. The objective of this ongoing work is to develop generic heuristics for adaptive selection of the parameters of Large Neighborhood Search. This paper proposes to use a Reinforcement Learning framework in order to adapt the heuristics during the search. Two heuristics are proposed to deal with the first two parameters of the metaheuristic. Three are proposed to adapt the last parameter. Preliminary computational results on the Car Sequencing problem are given. On this problem, only the adaptive selection of the first two parameters is effective.