Modelling dense crowds with Mean-Field Games (C. Appert-Rolland, CNRS, Univ. Paris Saclay)

Dense crowds raise important security issues, but are more difficult to study experimentally than more diluted pedestrian flows. We have conducted experiments in which a cylindrical intruder crosses a static crowd, in order to measure the resulting perturbation. In contrast with a common belief that, at high densities, mechanical forces dominate - and thus that pedestrian crowds behave like granular matter - our results show a behavior at odds with grains, with lateral displacements which stem from the fact that pedestrians anticipate and can move in directions not aligned with the forces they undergo. Classical agent based models, which mainly focus on collision avoidance, are not able to reproduce the observed behavior. We used a mean-field game approach which was able to reproduce the experimental observations. The reason for this success is that mean-field games can model anticipation much beyond the next collision....