Research on boarding strategy optimization using multi-agent reinforcement learning based on spatio-temporal graph attention

Aircraft boarding time directly influences the departure schedule and can potentially cause cascading delays. Existing boarding strategies often rely on pre-set sequences or manual rules, which degrade significantly under disruptive conditions. Moreover, current models tend to oversimplify passenger heterogeneity and spatial resource constraints, often evaluating performance based on aggregate efficiency metrics that lack interpretability. This results in insufficient strategy explainability and difficulty in achieving proactive adjustments. To address these limitations, we formulate the boarding process as a multi-agent sequential decision-making system under shared resource constraints and propose a closed-loop learning-based boarding optimization strategy framework driven by process state feedback. Adjustable parameters are used to cover typical heterogeneous passengers and spatial resources are modeled. A spatio-temporal graph attention mechanism is introduced to capture process information such as congestion generation locations, evolution and propagation, and key individuals. Multi-agent proximal policy optimization (MAPPO) is adopted as the solution algorithm, and an evaluation index considering both efficiency and passenger experience is constructed. Numerical simulations demonstrate that the proposed method achieves superior performance in boarding time, congestion severity and safety violation events, improving both efficiency and passenger experience. Additionally, the method identifies congested spatial locations, peak evolution and critical individuals contributing to congestion, enhancing the interpretability of the strategy. It also demonstrates better convergence stability and robustness across diverse scenarios. Overall, this approach provides managers with an interpretable, generalizable and deployable solution for real-world boarding optimization.