Real time learning control of high d.o.f. robots: automatic generation of discrete states and learning transition models
Hajime Kimura, Shigeru Kobayashi · Society of Instrument and Control Engineers of Japan · 2003
We present a model-based RL approach to cope with continuous space of high D.O.F. robots, combining model learning and an actor-critic method. The model learner generates a discrete state-transition model that helps improvement of both the policy and state-representation. In general, model-based methods tends to fail in non-Markovian problems, but the proposed method, using actor-critic, can find good policies in such environments.