Cooperative Scheduling Under Synchronous Decision and Asynchronous Execution for Flexible Job Shop Scheduling Problems

Flexible job shop scheduling requires coordinated allocation among parallel machines under continuous-time asynchronous execution. Existing learning-based approaches often serialize decisions at shared contention moments, leading to a mismatch between decision abstraction and execution dynamics. To address this issue, this article introduces a synchronous-decision–asynchronous-execution paradigm, which models each decision step as an executable joint allocation at shared decision points, thereby aligning the learning interface with parallel shop-floor dynamics. Based on this paradigm, we develop a learning-based realization, termed graph-enhanced dual-network QMIX, which captures local coordination and distinguishes scheduling and execution value semantics. Experiments on synthetic instances and standard benchmarks demonstrate improved makespan, training stability, and cross-scale generalization compared with priority dispatching rules and representative deep reinforcement learning baselines.