The Use of Continuous Action Representations to Scale Deep Reinforcement Learning : An Application to Inventory Control

We propose a new policy architecture to apply deep reinforcement learning (DRL) with a continuous action representation. This contrasts most current DRL implementations that make use of a discrete action space. The latter are not scalable to large problems. To obtain feasible discrete actions, we add a mapping function to the policy network that maps the continuous outputs of the policy network to a feasible integer solution. We demonstrate our approach on a (capacitated) joint replenishment problem for various settings. We show how the use of a continuous action representation enables solving larger problem instances, while requiring less training time and matching or outperforming state-of-the-art benchmarks. This promising research avenue might pave the way for application of deep reinforcement learning at scale and in practice