Reinforcement learning-based dynamic evacuation guidance for fire emergencies: Toward safety digital twins

초록

This study proposes a reinforcement learning-based dynamic evacuation guidance system for fire emergencies in multi-story buildings. The proposed framework integrates a cellular automata-based fire simulation with an agent-based evacuation simulation to model the coupled dynamics of fire spread and occupant movement. The fire simulation represents the spatiotemporal propagation of flames and smoke, including inter-floor spread through vertical shafts and the stack effect. The evacuation simulation models occupant behavior by incorporating walking speed reductions caused by smoke and crowd density, as well as probabilistic route choice using a mixed logit model to capture heterogeneous decision-making. Within this environment, a reinforcement learning agent dynamically controls directional restrictions at corridor junctions and exits. The agent observes crowd flow, crowd density, fire hazard scores, and exit information, and learns an evacuation guidance policy that minimizes both evacuation time and fatalities through interaction with the simulation environment. To ensure the relevance of the learned policy under realistic deployment conditions, the agent is additionally trained under a sensor uncertainty wrapper that models hazard-sensor noise, CCTV pedestrian detection misses, and measurement latency. Experiments were conducted on a three-story model under 26 fire scenarios with varying ignition locations, with all reinforcement learning agents evaluated under the full sensor uncertainty environment and compared against two deterministic baselines-a rule-based dynamic guidance and a strengthened shortest-path-based baseline that incorporates global path information. The proposed uncertainty-aware policy achieves average reductions of approximately 32% in total evacuation time and 58% in fatalities compared with the rule-based dynamic guidance, and approximately 20% and 42% in the same metrics compared with the strengthened shortest-path-based baseline, with statistical significance confirmed by paired Wilcoxon signed-rank tests with Holm-Bonferroni correction (Holm-adjusted p < 0.001). The results demonstrate that reinforcement learning can effectively learn adaptive evacuation guidance policies that consider both fire dynamics and crowd behavior under realistic sensor conditions. These findings support the use of simulation-driven AI decision support as one component of Safety Digital Twin-based safety management for complex buildings.

키워드