Decomposed Q-Learning for Fixture Selection in Multi-resource-Constrained Flexible Job Shop Scheduling

Pallet automation systems (PASs) are designed for flexible manufacturing, in which fixtures are mounted on pallets to form fixture-pallets (FPs) that move between machines and loading/unloading stations, while jobs are loaded onto and unloaded from the FPs. In such systems, FP selection directly affects loading/unloading time (LUT), resource occupation, and subse quent waiting, creating strong coupling among jobs, machines, and FPs. Existing PAS scheduling studies often simplify these interactions, and reinforcement learning for FP selection remains limited. To address this gap, this paper investigates a multi-resource-constrained flexible job shop scheduling problem with FP and LUT constraints (MRFL), where the objective is to minimize makespan. A decomposed Q-learning framework with a top-k state representation is developed for FP selection. Comparative experiments on both same-scale and unseen larger-scale cases show that the proposed method outperforms other methods in most cases and demonstrates good generalization ability for PAS scheduling.