Learning in Multi-Robot Systems

摘要

This paper 1 discusses why traditional reinforcement learning methods, and algorithms applied to those models, result in poor performance in dynamic, situated multi--agent domains characterized by multiple goals, noisy perception and action, and inconsistent reinforcement. We propose a methodology for designing the representation and the forcement functions that take advantage of implicit domain knowledge in order to accelerate learning in such domains, and demonstrate it experimentally in two different mobile robot domains. 1 Learning in Situated Domains Successful applications of RL methodologies to well-- behaved domains (Sutton 1988, Watkins 1989, Kaelbling 1990) have encouraged researchers to hypothesize about its value for learning on situated agents such as mobile robots. However, while simulation results are encouraging, work on physical robots has not yet repeated that success. We discuss why traditional such methods perform poorly in situated domains with multiple goals, n...