20608 Multi-agent reinforcement learning system using some agent to reduce uncertainty
Atushi Miyamae, Seiichi Kawata, Takeshi Tateyama · The Proceedings of Conference of Kanto Branch · 2006
It is known that the uncertainty of the state transition tends to rise in large-scale multi agent environments and to worsen converge of reinforcement learning. In this paper, it is tried to propose the new method that restrain exploration of some agents in each episode to improve converge of learning. The feature of this technique is that the agents whose explorations are restrained support learning of the normal agents by lowering the uncertainty of the state transition. It is shown that the proposal technique is effective by the simulation.