Propagation Guided Large Neighborhood Search

Laurent Perron, Paul Shaw, Vincent Furnon

Abstract: In this article, we explore how neighborhoods for the Large Neighborhood Search (LNS) framework can be automatically defined by the volume of propagation of our Constraint Programming (CP) solver. Thus we can build non trivial neighborhoods which will not be reduced to zero by propagation and whose size will be close to a parameter of the search. Furthermore, by looking at the history of domain reductions, we are able to deduce even better neighborhoods. This idea is validated by numerous experiments with the car sequencing problem. The result is a powerful and completely automatic method that is able to beat our hand-written neighborhoods both in term of performance and of stability. This is in fact the first time for us that a completely generic code is better than a hand-written one.