Speeding up multiagent reinforcement learning by coarse‐graining of perception: The hunter game
Akira Ito, Mitsuru Kanabuchi · Electronics and Communications in Japan (Part II Electronics) · 2001
Abstract Reinforcement learning is a promising technique for making agents learn to cooperate in the real world. When we try to apply the technique to a practical problem, however, the slowness of learning due to the increased number of states becomes a bottleneck. We try to speed up learning by reducing the effective number of states by coarse‐graining of perceptions at the early stage of learning, and at the same time try to ensure the overall performance by switching back to complete perception later. By only switching the perception granularity, however, it is difficult to rectify bad habits acquired at the early stage of learning. Hence, we make a coarse‐grained perception learner and a complete perception learner work in parallel, and switch the learner used for action selection at an appropriate time. This technique makes possible both fast early‐stage learning and overall good performance. © 2001 Scripta Technica, Electron Comm Jpn Pt 2, 84(12): 37–45, 2001