Multiagent reinforcement learning scheduling optimization for airport special vehicles under dynamic uncertainty

Abstract

The scheduling efficiency of airport special vehicles directly determines ground-handling quality and flight punctuality. Conventional methods rely heavily on human experience and static rules, lacking adaptability to dynamic uncertainties involving flights, vehicles, environments, and human-machine interactions. This paper proposes a multi-agent reinforcement learning (MARL) scheduling optimization framework integrating human-machine collaboration. A Markov decision process (MDP) model incorporating human factors is constructed to characterize complex stochastic dynamics. A centralized-training-decentralized-execution (CTDE) architecture is designed to unify global optimization and local autonomy. To address the limited adaptability of QMIX in airport scheduling, four targeted improvements are introduced: graph-attention-based spatiotemporal state encoding, priority-aware conflict resolution for action selection, humanmachine-cooperative reward shaping, and a dynamic-task-processing module. A high-fidelity SimPy simulation environment is built for comparative, ablation, generalization, and stability experiments. Results demonstrate that the proposed method significantly outperforms rule-based scheduling, genetic algorithms, and baseline QMIX in task completion rate, average response time, vehicle utilization, and driver workload (p<<0.01), exhibiting strong dynamic adaptability, scalability, and stability.