Emergent Collective Behaviors in a Multi-agent Reinforcement Learning Pedestrian Simulation: A Case Study

摘要: In this work, a Multi-agent Reinforcement Learning framework is used to generate simulations of virtual pedestrians groups. The aim study the influence two different learning approaches in quality generated simulations. case consists on simulation crossing groups embodied agents inside narrow corridor. This scenario classic experiment pedestrian modeling area, because collective behavior, specifically lanes formation, emerges with real pedestrians. paper studies algorithms, function approximation approaches, and knowledge transfer mechanisms performance learned behaviors. Specifically, RL-based schemas are analyzed. first one, Iterative Vector Quantization Q-Learning (ITVQQL), improves iteratively state-space generalizer based vector quantization. second scheme, named TS, uses tile coding as generalization method Sarsa(\(\lambda \)) algorithm. Knowledge approach use Probabilistic Policy Reuse incorporate previously acquired current processes; additionally, value also ITVQQL schema between consecutive iterations. Results demonstrate empirically that our RL generates individual behaviors capable emerging expected behavior occurred appears independently algorithm used, but depends extremely whether was applied or not. addition, techniques has remarkable final (measured number times task solved)