Research on Simulation and Path Optimization of University Emergency Evacuation Based on Multi-Agent Reinforcement Learning

To address the spatial misalignment between local observations and global bottleneck consequences in university emergency evacuation, as well as the difficulty of sparse team rewards in guiding diversion behaviors, the Topological Awareness Hierarchical State Encoder (TAHSE) and Bottleneck Sensitive Reward Decomposition Mechanism (BSRD) are proposed. These two mechanisms are coupled into a unified architecture within a centralized training-decentralized execution framework. TAHSE aggregates the features of the local neighborhood graph using attention and uses learnable bottleneck tokens as the keys and values for the second layer's cross-domain attention, enabling agents to obtain the compressed prior of the global bottleneck solely based on their own observations and pre-stored bottleneck graphs, reducing the computational complexity from the square level of the full graph attention to the linear level. BSRD decomposes a single negative time team reward into three components: progress, bottleneck pressure, and coordination completion. It triggers pressure penalties with a warning threshold lower than the safety upper limit and uses an adaptive coefficient to weigh the safety constraints as the difficulty of the scenario increases. Twenty seed repetition experiments were conducted in three-tier campus scenarios (500/1500/3000 agents) and structured perturbation protocols: The gap ratio of the proposed method relative to the upper bound reference in the three scenarios was controlled within 15%, while the MAPPO baseline increased from 18.4% to 31.5% as the scale expanded; the peak density effect under topological perturbation reached 1.42 (p<0.001), and the triggering threshold for diversion shifted systematically towards the bottleneck nodes with reduced capacity. The results show that embedding the bottleneck as a portable prior in the decentralized strategy and guiding diversion with reward signals ahead of the safety upper limit can achieve joint improvement in efficiency and safety at a controllable computational cost and is more reliable under structural variation conditions.