A Reinforcement Learning Method for Inventory Control Under State-based Stochastic Demand

With the development of intelligent manufacturing, managers increasingly realize that to improve the efficiency of production, we should not only improve the level of manufacturing technology, but also enhance the ability of supply chain management. In supply chain management, inventory control is always essential but troubling. Because the formulation of the optimal strategy in different inventory scenarios is very complicated, the model in the literature often makes compromise when compared to the reality. This paper deals with the non-stationary demand considering the impact of promotion and seasonal factors. Specifically, the paper constructs an open-AI gym environment for inventory control problems with state-based demand, and proposes a complete modeling method. A reinforcement learning method based on SARSA is developed to solve the problem. We conduct a series of numerical experiments, which provide guidance for the selection of the hyper-parameter
$\varepsilon$
, and confirm the ability of reinforcement learning to capture the unknown potential demand changes in the environment. Finally, some suggestions are drawn on how to deploy reinforcement learning in real inventory scenarios.