DRAGON: A Dynamic Risk-Aware Graph Optimization Network for Adaptive Building Evacuation Using Graph Convolutional Network and Q-Learning

Efficient evacuation route planning in dynamic environments remains a major challenge, particularly when environmental risks and accessibility conditions change rapidly during emergencies. Traditional shortest-path algorithms, while effective in static graphs, often fail to adapt to evolving conditions, leading to suboptimal or unsafe evacuation guidance. This study aims to develop an adaptive and intelligent evacuation routing system capable of dynamically responding to real-time environmental changes. We propose DRAGON (Dynamic Risk-Aware Graph Optimization Network), an integrated framework that combines reinforcement learning for adaptive decision-making with Graph Convolutional Networks (GCNs) for spatial representation learning. The GCN encodes the building topology and risk levels into node embeddings, while the RL agent learns optimal evacuation policies that adapt to dynamic conditions. In simulations across multiple building layouts, DRAGON achieves success rates of 93–95%, with average path lengths between 11.9 and 14.2 units and average risk values of 0.15–0.18, outperforming classical methods such as Dijkstra, A*, Bidirectional A*, and SARSA in both risk mitigation and evacuation efficiency. The system adapts in real-time to evolving hazards while maintaining computational feasibility, with observed runtimes increasing by only 18–25% over A* in large building graphs. Experimental results demonstrate that DRAGON achieves a significantly lower average cost and minimizes risk exposure, particularly in high-risk scenarios, while maintaining comparable time efficiency to traditional methods.