Reinforcement Learning Control Based on TWO-CMAC Structure

To improve the effect of the on-line control, a reinforcement learning algorithm presents which is based on CMAC (Cerebella Model Articulation Controller, CMAC) in this article. The algorithm includes two parts, the related search unit and the self-adaptive comment unit. The related search unit CMAC adopts the learning algorithm with instructors, that is, the P gain of fixed-line regulator provides online learning samples data, through which CMAC could take part in controlling step by step; while the self-adaptive comment unit CMAC judges the controlling error, it also produces enhanced or punishment signal. Such an algorithm is implied in controlling the speed of small wheeled robots, and analysis the control performance of the single neuron adaptive PID in the parameters system changes by contrast, Matlab simulation datum shows the rapidly and validity of the reinforcement learning control.