Principles for the Design of Large Neighborhood Search

Large neighborhood search (LNS) is a combination of constraint programming (CP) and local search (LS) that has proved to be a very effective tool for solving complex optimization problems. However, the practice of applying LNS to real world problems remains an art which requires a great deal of expertise. In this paper, we show how adaptive techniques can be used to create algorithms that adjust their behavior to suit the problem instance being solved. We present three design principles towards this goal: cost-based neighborhood heuristics, growing neighborhood sizes, and the application of learning algorithms to combine portfolios of neighborhood heuristics. Our results show that the application of these principles gives strong performance on a challenging set of job shop scheduling problems. More importantly, we are able to achieve robust solving performance across problem sets and time limits. This material is based upon works supported by the Science Foundation Ireland under Grant No. 00/PI.1/C075, the Embark Initiative of the Irish Research Council of Science Engineering and Technology under Grant PD2002/21, and ILOG, S.A.