Using reinforcement learning to solve vehicle routing problems with dynamic customers

ICLR 2026 Conference Submission19856 Authors

Keywords: Reinforcement Learning, Dynamic Vehicle Routing Problems, Graph Neural Networks, Vehicle Routing Problems

TL;DR: In this paper, we apply reinforcement learning based constructive heuristics to dynamic vehicle routing problems with dynamic customers.

Abstract: Deep Machine Learning methods have been proven to be effective in solving the travelling salesman problem and general vehicle routing problems. In this paper, we use reinforcement learning to learn effective and fast constructive heuristics for solving a problem with dynamic customer requests, the partially dynamic travelling repairman problem, and variants with customer demands and time windows. We perform an ablation study on the policy network that map between state and action spaces of arbitrary size to investigate which features from previous literature translate well to this new domain and introduce a recurrent neural network component to the decoder that tracks arrivals of dynamic customers improving performance on problems with time windows. The encoder-decoder network can map between state and action spaces of arbitrary dimension, so we investigate how it generalizes to problems of different sizes. The performance of the construction heuristic is compared with several baselines on real world examples and different spatio-temporal customer request distributions of different sizes.

Primary Area: reinforcement learning

Submission Number: 19856