Deep Reinforcement Learning for inventory optimization with non-stationary uncertain demand

Abstract

We consider here a single-item lot sizing problem with fixed costs, lead time, and both backorders and lost sales, and we show that, after an appropriate training in randomly generated environments, Deep Reinforcement Learning (DRL) agents can interpolate in real-time near-optimal dynamic policies on instances with a rolling-horizon, provided a previously unseen demand forecast and without the need to periodically resolve the problem. Extensive computational experiments show that the policies provided by these agents compete, and in some circumstances even outperform by several percentage points of gap, those provided by heuristics based on dynamic programming. These results confirm the importance of DRL in the context of inventory control problems and support its use in solving practical instances featuring realistic assumptions.