Multi-Echelon Inventory optimizations for Divergent Networks by Combining Deep Reinforcement Learning and Heuristics Improvement

Mathematical solutions of obtaining an optimal policy of multi-echelon inventory are often computationally intractable for supply chain systems of realistic sizes. To deal with supply chain environment dynamics and the very large space of planning variables, this paper proposes a novel simulation-optimization approach for dynamic optimization of a continuous review (s, Q) multi-echelon inventory policy by combining deep reinforcement learning approach with local search metaheuristics. We compare the performance of a scatter search, a deep Q-learning and our proposed combined approach. The simulation outcomes show that the combined approach improves decision performances evidently in terms of profitability, adaptability and solution time efficiency.