Multi-Agent Reinforcement Learning for Optimizing Path Planning of Transporters Considering the Congestion

Abstract

This paper studied a path planning method by utilizing the framework of Multi-Agent Reinforcement Learning (MARL) and other useful learning techniques to ensure agile and stable solution for the mega Distribution Centers (mega DCs) supporting smart logistics operations. We established a different types of reward and other techniques to enhance overall system performance. And, we trained MARL networks to optimize the paths of a limited number of transportation devices, considering system efficiency in smaller-scale environments. Finally, a strategy by using the parameters obtained from the initial training has been developed for operating a larger-scale system. Several mechanisms based on the graph theory has been also utilized to deal with collisions, congestion, and potential deadlock issues among transporters. From the extensive numerical analyses, the performance of the proposed mechanism has been validated for its practical applications in real cases.