A negotiation-based multi-agent reinforcement learning for dynamic scheduling of reconfigurable manufacturing systems

Reconfigurable manufacturing systems demand scheduling decisions that respond in real time to dynamic job arrivals, machine breakdowns, and module-level reconfiguration constraints. This paper proposes a negotiation-driven multi-agent reinforcement learning framework in which enhanced deep Q-network agents are trained centrally and deployed in decentralised execution. Conflicts over machines and shared modules are resolved through an attention-based bidding mechanism incorporating due-date-sensitive features. Experiments in a simulated RMS environment show that the proposed method outperforms the entire multi-agent reinforcement learning family (QMIX, VDN, QTRAN, QPLEX, Weighted QMIX, and MAPPO) by 25% in makespan, matches the strongest single-agent baseline, and attains the lowest reconfiguration overhead of all 15 methods evaluated. Classical metaheuristics achieve marginally better makespan under static conditions, but their reliance on repeated re-optimisation, requiring tens to hundreds of seconds, limits their use in dynamic settings. The proposed framework instead generates decisions within milliseconds and maintains full job completion under a single-machine failure representing a 20% capacity reduction, without retraining. These results establish explicit negotiation as an effective alternative to value decomposition for resolving competitive resource contention, and the framework as a viable solution for real-time scheduling in RMS environments.