Reinforcement learning based material requirement planning for customised assemble-to-order systems with incomplete bill-of-materials information and stochastic lead times

We study a material requirement planning (MRP) problem for customised assemble-to-order (ATO) systems in which order-specific bills of materials (BOMs) are unknown when orders are placed and are determined during design lead times, while component replenishment lead times are stochastic. MRP for general ATO systems has long been considered analytically intractable; the combination of uncertain BOMs, delayed BOM revelation, and stochastic replenishment lead times in a dynamic customisation environment further enlarges the state space and strengthens the coupling between replenishment and allocation. We propose a structure-aware deep reinforcement learning framework. Our approach utilises a graph-based encoder and a GRU (Gated Recurrent Unit) module to effectively extract features from the product-component requirements and stochastic pipeline dynamics. Leveraging these features, we employ a dual-policy head that simultaneously outputs state-dependent parameters for base-stock replenishment and allocation strategies, thereby preserving the interpretability of classical inventory rules. We further introduce structure-based regularisation terms, inspired by structural patterns suggested by simplified ATO systems, to encourage economically meaningful behaviour and improve robustness. Numerical experiments show lower cost and competitive on-time fulfillment in the base case and in stress regimes with high BOM uncertainty, long replenishment lead times, tight delivery windows, and high backlog penalties.