Ruminative Reinforcement Learning: Improve Intelligent Inventory Control by Ruminating on the Past

Reinforcement Learning (RL) can solve practical sequential decision problems, even when structures of the problems are less understood. However, some sequential decision problems intrinsically have structural parts that are easily to formulate and distinguish from less understood parts. Exploiting this knowledge may help improve performance of RL. This study proposed and investigated an approach to exploit the knowledge of structural parts of inventory management problems in the context of RL. The proposed method is motivated by human behavior of ruminating on what has happened and what would happen if alternative choices would have been taken. Our investigation provides an insight into RL mechanism and our experimental results show viability of the approach.