A Deep Reinforcement Learning Approach to solve the Vehicle Routing Problem with Resource Constraints

Abstract

This paper introduces a deep reinforcement learning approach to solve the team orienteering problem, an important variant of the vehicle routing problem with resource constraints. Real-world problems like the team orienteering problem occur whenever there are insufficient resources to visit all the mission nodes. Classical methods to routing problems find optimal solutions with theoretical guarantees but often take an impractical amount of time and are thus intractable for large-scale problems. Recently, deep reinforcement learning approaches have been developed to solve various routing problems to near-optimality. These data-driven methods train a policy neural network to learn heuristics and autonomously construct trajectories with good quality in a short time. The efficacy of the deep reinforcement learning method is demonstrated through a test case study. The performance of the trained model is shown to improve after using advanced inference strategies, getting high-quality results close to the optimal.