Multi-Agent Reinforcement Learning with Shared Resource in Inventory Management

Mingxiao Feng, Guozi Liu, Li Zhao, Lei Song, Jiang Bian, Tao Qin, Wengang Zhou, Houqiang Li, Tie-Yan Liu

Keywords: Multi-Agent Reinforcement Learning, Inventory Management, Shared Resource, Decentralized Training Paradigm, Model-based RL

Abstract: We consider inventory management (IM) problem for a single store with a large number of SKUs (stock keeping units) in this paper, where we need to make replenishment decisions for each SKU to balance its supply and demand. Each SKU should cooperate with each other to maximize profits, as well as compete for shared resources e.g., warehouse spaces, budget etc. Co-existence of cooperation and competition behaviors makes IM a complicate game, hence IM can be naturally modelled as a multi-agent reinforcement learning (MARL) problem. In IM problem, we find that agents only interact indirectly with each other through some shared resources, e.g., warehouse spaces. To formally model MARL problems with above structure, we propose shared resource stochastic game along with an efficient algorithm to learn policies particularly for a large number of agents. By leveraging shared-resource structure, our method can greatly reduce model complexity and accelerate learning procedure compared with standard MARL algorithms, as shown by extensive experiments.

One-sentence Summary: We propose shared-resource stochastic game to capture the problem structure in IM and propose a novel algorithm that leverages the shared-resource structure to solve IM problem efficiently.