Inventory Planning in Capacitated High-Tech Assembly Systems Under Non-Stationary Demand

We study inventory planning in high-tech manufacturing supply chains driven by the semiconductor market. Although time correlation and trends in demand fundamentally impact optimal control in capacitated assembly systems, there is a lack of methods that rigorously address these intricacies. We formulate the problem as a Markov decision process and introduce a uniquely tractable demand model capturing pivotal elements of non-stationary uncertainty observed in practice. This formulation facilitates a comparison between the widely acknowledged class of base-stock policies and deep reinforcement learning (DRL) as well as meticulous benchmarking against the optimal policy for smaller problems. Integrating and extending key results from the literature, we devise novel base-stock policies tailored to this complex setting. Specifically, we enhance the inventory balancing equations for parallel assembly lines and develop the base-stock level computation by incorporating the multi-echelon shortfall distribution to account for the cascading effects of capacity constraints throughout the network. A comprehensive computational study demonstrates that, while standard base-stock policies typically perform poorly, our tailored base-stock policy attains adequate performance in stationary environments, where it naturally scales well due to its simplicity. However, even when equipped with dynamic base-stock levels based on updated demand information, performance regresses under non-stationarity as the policy fails to adequately anticipate future capacity shortages. Integration with DRL proves to be highly effective with optimality gaps consistently below 1%. An extensive case, inspired by the setting at our industry partner ASML, shows that the DRL policy remains to significantly outperform the best base-stock policy, and benefits higher service levels by investing additional inventory, particularly allocated downstream in the supply chain.