Learning online combinatorial stochastic policies with deep reinforcement

Abstract

We study on-demand delivery services in stochastic environments and propose one representation of the online vehicle routing problem that can be learned with supervised and reinforcement learning. Using a simulation framework with topology projections, we evaluate different models and show that stochastic policies can be learned. Fine-tuning the models through the use of the REINFORCE rule and with a deterministic critic suggests that this approach can lead to promising policies to solve the problem.