Comparison of Machine Learning Algorithms for Vehicle Routing Problems

Abstract

Vehicle routing is a classical problem in combinatorial optimization. A large number of exact and heuristic solution methods have been developed in the past. In the last few years, machine learning algorithms have been applied to such problems with some success. This paper investigates three recent machine learning algorithms: reinforcement learning, the dynamic attention model and neural large neighborhood search. These algorithms are compared on a variety of benchmark problems from the literature. It is found that the neural large neighborhood approach gave the best quality solutions. The dynamic attention model was found to require the largest amount of memory and was not able to be trained for larger instances. Reinforcement learning provided a good compromise between runtime and solution quality.