Learning a Stopping Criterion for Local Search

Local search is a very effective technique to tackle combinatorial problems in multiple areas ranging from telecommunications to transportations, and VLSI circuit design. A local search algorithm typically explores the space of solutions until a given stopping criterion is met. Ideally, the algorithm is executed until a target solution is reached (e.g., optimal or near-optimal). However, in many real-world problems such a target is unknown. In this work, our objective is to study the application of machine learning techniques to carefully craft a stopping criterion. More precisely, we exploit instance features to predict the expected quality of the solution for a given algorithm to solve a given problem instance, we then run the local search algorithm until the expected quality is reached. Our experiments indicate that the suggested method is able to reduce the average runtime up to 80% for real-world instances and up to 97% for randomly generated instances with a minor impact in the quality of the solutions.