New RL Framework Optimizes Dynamic Assembly Flow Shop Scheduling.

This paper introduces a novel approach to tackle the complex problem of dynamic assembly flow shop scheduling, particularly challenging due to multi-product kitting delivery and fluctuating order arrivals in hybrid manufacturing systems. These factors simultaneously alter supply dependencies and feasible job-machine assignments, demanding real-time scheduling solutions. The proposed solution is a sliding-window-based reinforcement learning (SWRL) framework designed for end-to-end online scheduling. It models the problem as a heterogeneous graph-based Markov decision process, capturing intricate kitting structures and bottleneck dynamics that often lead to sparse reward landscapes. SWRL integrates a sliding-window filtering mechanism to prioritize critical operations, a spatiotemporal graph encoding network to track bottleneck shifts, and a dynamic action mapping module with a constrained waiting strategy. Experiments using real-world data from a home appliance manufacturer demonstrated that SWRL consistently reduced tardiness compared to classical dispatching rules and other deep reinforcement learning methods, proving robust across various resource configurations and order loads.