Automatic Parameter Tuning via Reinforcement Learning for Crowd Simulation with Social Distancing

Abstract

Reinforcement learning (RL) has been applied to a variety of fields such as gaming and robot navigation. We study the application of RL in crowd simulation by proposing an automatic parameter tuning system based on Proximal Policy Optimization (PPO). The system can be used with any crowd simulation software to improve the quality of the simulation by automatically assigning parameters to each agent during the simulation. Our experiments indicate that the automatic parameter tuning system can reduce unexpected congestions in counterflow scenarios. In addition, by utilizing the improved commonly used crowd simulation algorithms and our parame-ter tunning system, we can represent social distancing behavior of pedestrians under COVID-19, where pedestrians comply to the suggested social distance when they have enough space to move while they reduce their social distances to others when there is limited space.