Deep Reinforcement Learning for the Capacitated Vehicle Routing Problem with Soft Time Window

The past decade has seen a rapid development of solving travelling salesman problem (TSP) and vehicle routing problem (VRP) with deep reinforcement learning. In order to solve problems that are closer to life, more researchers turn their attention to the variant VRP. In this article, we tackle the capacitated VRP with soft time window (CVRPSTW). In this problem, the vehicles have capacity limit and will be punished if arriving at the customer outside the time window. We use a deep reinforcement learning (DRL) based on the attention mechanism and point network to solve CVRPSTW. In the training part, we use policy gradient with rollout baseline. The experiment shows that the proposed DRL model can effectively solve this variant VRP.