Learning convergence in cyclic learning

Abstract

The learning convergence of the cerebellar model articulation controller (CMAC) in cyclic learning is discussed. The authors demonstrate the following findings. To begin, assuming the training samples are noiseless, the learning algorithm converges if and only if the learning rate is chosen from a set of possible values (0, 2). Second, if the learning rate is dynamically reduced when the training samples include noise, the learning algorithm will converge with probability one. Third, given a modest but fixed learning rate ε in the noise situation, the mean square error of the weight sequences generated by the CMAC learning algorithm will be constrained by O. (ε). To put these findings to the test, certain simulation experiments are carried out.