Learning to Prune Electric Vehicle Routing Problems

James Fitzpatrick, Deepak Ajwani, Paula Carroll

Abstract: Electric vehicle variants of vehicle routing problems are significantly more difficult and time-consuming to solve than traditional variants. Many solution techniques fall short of the performance that has been achieved for traditional problem variants. Machine learning approaches have been proposed as a general end-to-end heuristic solution technique for routing problems. These techniques have so far proven flexible but don’t compete with traditional approaches on well-studied problem variants. However, developing traditional techniques to solve electric vehicle routing problems is time-consuming. In this work we extend the learning-to-prune framework to the case where exact solution techniques cannot be used to gather labelled training data. We propose a highly-adaptable deep learning heuristic to create high-quality solutions in reasonable computational time. We demonstrate the approach to solve electric vehicle routing with nonlinear charging functions. We incorporate the machine learning heuristics as elements of an exact branch-and-bound matheuristic, and evaluate performance on a benchmark dataset. The results of computational experiments demonstrate the usefulness of our approach from the point of view of variable sparsification.