Learning Optimal Crowd Evacuation from Scratch Through Self‐Play

A key goal in evacuation management is to quickly and safely remove panicking crowds from buildings, festivals, or airplanes while preventing crush fatalities. Recently, there has been much progress in realistically modeling crowds in complex environments, based on social force models, cellular automata, and machine learning. However, current models assume specific social interactions and do not allow to systematically explore how to optimize crowd cooperation and evacuation. In contrast, the present work focuses on the question, how an ideal crowd of superintelligent agents, comprising humans, robots, or smart active particles, would cooperate to optimize evacuation. A method is developed that uses multiagent reinforcement learning combined with self‐play to learn optimal crowd behavior from scratch. Crucially, the agents in this approach are pressure‐aware and autonomously learn collision and crushing avoidance. After training, they adopt interpretable evacuation strategies like queuing and zipper merging and outperform traditional evacuation models in terms of fatality avoidance and evacuation rate. Our method can be used to enhance guidelines for mass evacuation, potentially saving lives.