A Flexible Job Shop Scheduling Method via a Hybrid Dual Attention Network and Mamba Approach

In recent years, the Flexible Job Shop Scheduling Problem (FJSP) has gained significant attention in smart manufacturing. Deep reinforcement learning (DRL) is gradually supplanting traditional heuristics, enabling broader industrial applicability. Disjunctive graphs facilitate effective state representation, while self-attention models excel in local feature extraction. However, the imbalance between local and global dependencies hinders adaptability to varying data scales, limiting model performance. To mitigate these issues, we propose DAN-Mamba, which is a hybrid framework by integrating a Dual Attention Network (DAN) with the Mamba model to enhance deep feature extraction. By effectively modeling local-global dependencies and complex operation-machine interactions, our method supports adaptive, production-oriented decision-making. Experiments on large-scale synthetic and classical benchmarks demonstrate superior performance, with significant reductions in makespan and strong generalization across diverse scheduling scenarios.