Robust credit assigned CMAC

In this paper, the online learning capability and the robust property for the learning algorithms of cerebellar model articulation controllers (CMAC) are discussed. Both the traditional CMAC and fuzzy CMAC are considered. A credit assignment idea is adopted to provide fast learning for CMAC. The idea is to distribute errors proportional to the inverse of learning times, which are viewed as the credibility of addressed cells. In the paper, we also embed the M-estimator into the CMAC learning algorithms to provide the robust property against noise or outliers existing in training data. An annealing schedule is also adopted to suitably define a scale estimate required in the M-estimator. From example simulations, it is clearly evident that the proposed algorithm indeed has faster and more robust learning than traditional CMAC does. Besides, we also employ the proposed CMAC for an online learning control scheme used in the literature. The simulation results indeed show the effectiveness of the proposed approaches.