Social crowd controllers using reinforcement learning methods

Crowd Simulation is an area of research that is present in several disciplines and industries. Even though the visualization of crowds is an important subject, the behavior behind it helps to make it believable. Behavioral modeling can be a te- dious task because the agents have to mimic the complex...

Descripción completa

Detalles Bibliográficos
Autor: Casadiego Bastidas, Luiselena
Tipo de recurso: tesis de maestría
Fecha de publicación:2014
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2099.1/22235
Acceso en línea:https://hdl.handle.net/2099.1/22235
Access Level:acceso abierto
Palabra clave:Collective behavior -- Computer simulation
Crowds -- Computer simulation
Reinforcement Learning
Q-Learning
Comportament col·lectiu -- Simulació per ordinador
Multituds -- Simulació per ordinador
Àrees temàtiques de la UPC::Matemàtiques i estadística::Investigació operativa::Simulació
Descripción
Sumario:Crowd Simulation is an area of research that is present in several disciplines and industries. Even though the visualization of crowds is an important subject, the behavior behind it helps to make it believable. Behavioral modeling can be a te- dious task because the agents have to mimic the complexity of human reactions to situations. In this work, we propose an alternative model for crowd simulation of pedestrian movements using Reinforcement Learning methods for low level deci- sions. Taking the approach of microscopic models, we trained an agent to reach a goal point while avoiding obstacles that might be on its way, and trying to follow a coherent path during the walk. Once one agent has learned, its knowledge is passed to the rest of the members of the crowd by sharing the resulting Q-Table, expecting the individual behavior and interactions to lead to a crowd behavior. We presented states sets, an action set and reward functions general enough to adapt to different environments, allowing us to use the same knowledge in different scenario settings.