Deep Reinforcement Learning has shown great potential to solve complex sequential decision-making problems and has been applied successfully to different intractable in- ventory control problems. Therefore, this paper models a multi-product, multi-echelon inventory control problem as a Markov Decision Process and applies Proximal Pol- icy Optimization to solve it. We consider a realistic supply chain network of general structure provided in the scope of the ASML inventory challenge. Different application approaches of PPO are evaluated regarding their capability to handle the large com- binatorial complexity inherent to the problem. We propose a continuous action space to improve scalability and a Combinatorial Optimization Layer to transform continu- ous actions into feasible, discrete ones. Further, Graph Neural Networks are leveraged to extract structured feature embeddings and to predict continuous actions in a DRL framework using a supervised pre-training approach. Our results show, that DRL and GNN’s have the potential to match a benchmark heuristic, but have problems when being faced with many inter-dependencies of components.     «

Deep Reinforcement Learning has shown great potential to solve complex sequential decision-making problems and has been applied successfully to different intractable in- ventory control problems. Therefore, this paper models a multi-product, multi-echelon inventory control problem as a Markov Decision Process and applies Proximal Pol- icy Optimization to solve it. We consider a realistic supply chain network of general structure provided in the scope of the ASML inventory challenge. Differe...     »