Date of Award

Winter 2003

Project Type


Program or Major

Engineering: Electrical

Degree Name

Doctor of Philosophy

First Advisor

L Gordon Kraft


The CMAC (Cerebellar Model Articulation Controller) neural network has been successfully used in control systems and other applications for many years. The network structure is modular and associative, allowing for rapid learning convergence with an ease of implementation in either hardware or software. The rate of convergence of the network is determined largely by the choice of the receptive field shape and the generalization parameter. This research contains a rigorous analysis of the rate of convergence with the standard CMAC, as well as the rate of convergence of networks using other receptive field shape. The effects of decimation from state-space to weight space are examined in detail. This analysis shows CMAC to be an adaptive lowpass filter, where the filter dynamics are governed by the generalization parameter. A more general CMAC is derived using wavelet-based receptive fields and a controllable decimation scheme, that is capable of convergence at any frequency within the Nyquist limits. The flexible decimation structure facilitates the optimization of computation for complex multidimensional problems. The stability of the wavelet-based CMAC is also examined.