Learning to guide local search optimisation for routing problems

Nasrin Sultana, Jeffrey Chan, Babak Abbasi, Tabinda Sarwar, Alex Kai Qin

Abstract: Machine learning has shown promises in tackling routing problems yet falls short of state-of-the-art solutions achieved by stand-alone operations research algorithms. This paper introduces “Learning to Guide Local Search” (L2GLS), a novel approach that leverages Local Search operators' strengths and reinforcement learning to adjust search efforts adaptively. The results of comparing L2GLS with the existing cutting-edge approaches indicate that L2GLS attains new levels of state-of-the-art performance, particularly excelling in handling large instances that continue to challenge existing algorithms.