关键词: 增强学习, 小脑模型关节控制器, 库恩三角化, Markov决策过程

Abstract: A reinforcement learning controller based on CMAC neural networks using triangulation was studied and applied to the learning control of intercepting a ball in the RoboCup. By utilizing Kuhn triangulation based on simplex interpolation in the continuous state space of Markov Decision Processes (MDPs), the value functions of MDPs were approximated with linear triangulation so that the generalization ability of the CMACbased reinforcement learning controller could be improved. Simulation results on the learning control of intercepting a ball show that the CMACbased learning controller using triangulation is much more efficient than the learning controller based on CMAC uniform coding.

Key words: reinforcement learning, Cerebellar Model Articulation Controller(CMAC), Kuhn triangulation, Markov decision process