Learning to Solve Multi-AGV Scheduling Problem with Pod Repositioning Optimization in RMFS

This paper studies a novel non-deterministic polynomial-hard (NP-hard) warehouse optimization problem, the multi-automated guided vehicle (multi-AGV) scheduling problem with pod repositioning optimization (MAS-PR), in a robotic mobile fulfillment system (RMFS). A major challenge in MAS-PR is that the storage assignment of pods and the task allocation decision of AGVs are closely coupled. To address this problem, an end-to-end Problem-Specific Mask-based Deep Reinforcement Learning (PSMDRL) method is proposed in this paper. In PSMDRL, we transform the MAS-PR into a Markov Decision Process (MDP) model, and then use an attention-based network model to learn the efficient allocation and retrieval policy. A mask strategy with problem-specific knowledge and a model structure with decoder embedding are designed to further improve the efficiency and effectiveness. Experimental results demonstrate that the proposed PSMDRL is more effective and efficient than the representative algorithms in solving the MAS-PR problem.