From one to many: Simulating groups of agents with reinforcement learning controllers

Abstract

Simulation of crowd behavior has been approached through many different methodologies, but the problem of mimicking human decisions and reactions remains a challenge for all. We propose an alternative model for simulation of pedestrian movements using Reinforcement Learning. Taking the approach of microscopic models, we train an agent to move towards a goal while avoiding obstacles. Once one agent has learned, its knowledge is transferred to the rest of the members of the group by sharing the resulting Q-Table. This results in individual behavior leading to emergent group behavior. We present a framework with states, actions and reward functions general enough to easily adapt to different environment configurations.