Optimizing evacuation strategies using a multi-agent model with behavior-mode transitions in fire environments

Fire remains a persistent threat to human life and property. However, the specific effects of visibility on pedestrian movement and health conditions—particularly the role of visibility thresholds—are still poorly understood. This study develops a coupled evacuation model integrating an Agent-Based Model (ABM) with the Fire Dynamics Simulator (FDS) to examine how visibility shapes pedestrian behavior during fires. The model analyzes how response time, visual thresholds, and movement mode changes affect total evacuation time, casualty probability, and trampling risk. An XGBoost algorithm predicts evacuation risks, while the NSGA-III multi-objective optimization identifies Pareto-optimal strategies, further evaluated using the entropy weight method to determine the global optimum. Results indicate that, within the investigated scenario, adaptive behaviors (such as bending or crawling) slightly prolong evacuation time while substantially reducing casualties and trampling risk. A higher proportion of bending postures significantly lowers overall risk. The integrated XGBoost–NSGA-III framework identifies optimal visibility ranges and movement ratios, demonstrating that visibility-aware evacuation strategies can substantially improve occupant safety and support intelligent, risk-informed fire evacuation planning.