Multi-Agent Reinforcement Learning
Yuta KAJII, Kazuaki Yamada · The Proceedings of JSME annual Conference on Robotics and Mechatronics (Robomec) · 2017
This paper investigates learning performances of a multi-agent reinforcement learning that is employed reliability rate, eligibility trace, transfer learning and wide perception range. Reliability rate is used to learn stable behaviors in dynamic environments by adjusting automatically discount rate of Q learning. Eligibility trace is employed to converge learning earlier by going retroactive to selected behaviors and propagating discounted reward to them. Transfer learning is used to converge group learning earlier by transferring gained rules by an agent to the other agent. Wide perception range is able to expected to solve perceptual aliasing problem. The proposed method is tested under a narrow path problem which is one of dispersion games.