Multi-echelon inventory optimization using deep reinforcement learning

This paper studies the applicability of a deep reinforcement learning approach to three different multi-echelon inventory systems, with the objective of minimizing the holding and backorder costs. First, we conduct an extensive literature review to map the current applications of reinforcement learning in multi-echelon inventory systems. Next, we apply our deep reinforcement learning method to three cases with different network structures (linear, divergent, and general structures). The linear and divergent cases are derived from literature, whereas the general case is based on a real-life manufacturer. We apply the proximal policy optimization (PPO) algorithm, with a continuous action space, and show that it consistently outperforms the benchmark solution. It achieves an average improvement of 16.4% for the linear case, 11.3% for the divergent case, and 6.6% for the general case. We explain the limitations of our approach and propose avenues for future research.