A novel graph reinforcement learning framework for multilevel flexible scheduling in fuzzy production environments

Existing intelligent optimisation algorithms for fuzzy scheduling often rely on multidimensional, feature-based evaluations of fuzzy numbers, making repeated solution evaluation computationally expensive. This study presents an end-to-end graph reinforcement learning framework for a multilevel flexible manufacturing scheduling problem with fuzzy processing times. The framework jointly models the problem using a multi-attribute hierarchical fuzzy heterogeneous scheduling graph and a fuzzy-aware Markov decision process, with triangular fuzzy numbers represented as node and edge attributes. A fuzzy-aware heterogeneous graph attention network incorporates fuzzy processing times into attention weights through edge feature-enhanced attention and exploits the size-invariant message-passing structure of graph neural networks for cross-scale generalisation. A hierarchically decoupled heterogeneous actor-critic architecture decomposes each scheduling action into stage, entity and machine selection, while adaptive load-balancing guidance and constraint-aware masking support resource allocation and feasibility. Proximal policy optimisation is used as the standard policy optimiser. Experiments are performed to evaluate solution quality, scalability and computational efficiency across multiple instance groups. The proposed framework produces competitive schedules and is approximately 18.6 times faster.