Deep Reinforcement Learning For Inventory Control

This research investigates the potential of Deep Reinforcement Learning (DRL) to optimize spare parts inventory control by incorporating real-time degradation data. While DRL has shown promise in various inventory control applications, its application to spare parts control leveraging degradation data remains relatively unexplored. We propose a framework to develop intelligent spare parts ordering policies for systems with multiple machines and stochastic lead times. Three state-of-the-art DRL algorithms – DQN, PPO, and DCL – are employed to learn optimal ordering strategies directly from the system dynamics and degradation data. To evaluate the performance of the DRL-based policies, we compare them against the Proactive Base Stock Policy (ProBSP), a high performing spare parts policy that leverages degradation data. Our numerical experiments analyze the impact of various problem parameters, such as degradation rates and lead time variability, on the performance of the DRL and ProBSP policies. The results demonstrate that DCL consistently outperforms DQN and PPO, and in certain scenarios, surpasses ProBSP. This research contributes to the advancement of data-driven decision-making in spare parts inventory control and provides valuable insights for practitioners in industries with critical equipment.