Reinforcement Learning for Optimizing Can-Order Policy with the Rolling Horizon Method

This study presents a novel approach to a mixed-integer linear programming (MILP) model for periodic inventory management that combines reinforcement learning algorithms. The rolling horizon method (RHM) is a multi-period optimization approach that is applied to handle new information in updated markets. The RHM faces a limitation in easily determining a prediction horizon; to overcome this, a dynamic RHM is developed in which RL algorithms optimize the prediction horizon of the RHM. The state vector consisted of the order-up-to-level, real demand, total cost, holding cost, and backorder cost, whereas the action included the prediction horizon and forecasting demand for the next time step. The performance of the proposed model was validated through two experiments conducted in cases with stable and uncertain demand patterns. The results showed the effectiveness of the proposed approach in inventory management, particularly when the proximal policy optimization (PPO) algorithm was used for training compared with other reinforcement learning algorithms. This study signifies important advancements in both the theoretical and practical aspects of multi-item inventory management.