Generative agent-based modelling in evacuation simulation

Abstract

In this study, a novel Agent-Based Model (ABM) was developed that integrates Generative Artifical Intelligence (GenAI) derived behavioural attributes with the Social Force Model (SFM) to simulate evacuation scenarios. GenAI was integrated with Agent-Based Modelling (ABM) to enhance evacuation simulations. Using Meta's LLaMA 3 model, a behavioural grid based on personality traits, demographics, and environmental factors was developed to simulate diverse human responses during emergencies. Three models were compared: a standard social force baseline, a personality-driven impatience model, and a social cohesion group model. The impatience model achieved a 167% higher initial evacuation rate than the baseline, while the group model completed evacuation fastest overall. Random Forest analysis revealed personality traits as primary determinants of behaviour, with neuroticism and agreeableness having the strongest influence. This approach is shown to balance psychological realism with computational efficiency, enabling large-scale simulations of behaviourally diverse crowds without requiring real-time AI inference, and offering valuable insights for emergency management and evacuation planning.