Modelling emergent pedestrian evacuation behaviors from intelligent, game-playing agents

Abstract Much work has been done to understand complex crowd dynamics and self-organizing behaviors in high-density crowd situations. But most approaches for modelling pedestrian dynamics in emergencies require complex computations, making it difficult to capture multiple individual behaviors within a single model. This paper describes an agent-based model (ABM) that incorporates Bayesian game theory into pedestrian simulations. It assumes that players (agents) are playing a Bayesian game (i.e. games with incomplete information) and adopt strategies based on the anticipated behaviors of others to achieve a Bayesian Nash Equilibrium (BNE). Here, the model agents make decisions based on the possible positions of neighbors in the next time period to maximize their comfort and efficiently achieve their evacuation goal. A series of simulation experiments were undertaken using corridors, bottlenecks, and intersections in simulated evacuation spaces with the characteristics of mass tramping accidents. BNE provides a realistic and efficient approach for modelling complicated pedestrian dynamics with strong applicability. The BNE-informed ABM performance (evacuation times, routes, and behaviors) demonstrates its ability to realistically simulate emergent patterns of evacuation behaviors. The results indicate that agents using game theory reflect the behaviors of individuals with crowds well: BNE agents evacuate effectively at low densities and low blockages but are confounded in situations with few route choices in highly constricted spaces. The BNE-informed model provides a platform to better understand diverse crowd behaviors (e.g. herding and self-organized queuing, etc.) in varied spatial contexts, contributing to the designs of urban public space, evacuation planning, and crowd management.