From one to many: Simulating groups of agents with reinforcement learning controllers
Document typeConference lecture
Rights accessOpen Access
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.
CitationCasadiego, L., Pelechano, N. From one to many: Simulating groups of agents with reinforcement learning controllers. A: International Conference on Intelligent Virtual Agents. "Intelligent Virtual Agents: 15th International Conference, IVA 2015: Delft, The Netherlands, August 26-28, 2015: proceedings". Delft: 2015, p. 119-123.