Abstract
Multi-product kitting delivery poses significant challenges for real-time scheduling in hybrid manufacturing systems that integrate processing and assembly, as dynamic order arrivals alter supply dependencies and feasible job-machine assignments. This paper proposes a Sliding-Window Reinforcement Learning (SWRL) framework for flexible assembly flow shop scheduling under complex kitting constraints.
The problem is formulated as a heterogeneous graph-based Markov decision process, capturing the dual-layer kitting structure and tail-product bottleneck dynamics that create a sparse reward landscape. To address these challenges, SWRL integrates:
- Sliding-window filtering mechanism: Filters inactive nodes and prioritizes kitting-critical operations.
- Spatiotemporal graph encoding network: Tracks bottleneck shifts across consecutive decision states.
- Dynamic action mapping module: Adapts to changing action spaces under variable topologies using a constrained waiting strategy.
Experiments on real-world instances from a home appliance manufacturer demonstrate that SWRL consistently reduces tardiness compared to classical dispatching rules and existing deep reinforcement learning methods, showing robust performance across varying resource configurations, order loads, and arrival concentrations.
Blogger's Review: The proposed sliding-window reinforcement learning method effectively addresses the complexities of dynamic scheduling, significantly improving scheduling efficiency through innovative graph modeling and filtering mechanisms. Its robustness offers valuable insights for future scheduling in manufacturing systems.