Generic Adaptive Heuristics for Large Neighborhood Search

The Large Neighborhood Search (LNS) approach for solving Constrained Optimization Problems has been proved to be effective on a wide range of problems. LNS is a local search metaheuristic that uses a complete search method (such as CP) to explore a large neighborhood. At each step, the neighborhood is defined by relaxing a subset, called fragment, of the variables in the current solution. Despite the success of LNS, no general principle has emerged on how to choose the fragment from one restart to the other. Practitioners often prefer to relax randomly the solution to favor diversification. This work focuses on the design of generic adaptive heuristics for choosing automatically the fragment in LNS, improving the standard random choice. The defined heuristics are tested on the Car Sequencing problem for which we introduce a new original relaxation. From those experiments, we conclude that all our heuristics except one are performing better than a random fragment selection. We also show that our mean dynamic impact proximity and min/max dynamic impact proximity heuristics are significantly better than all the others.