| The main research contribution in this dissertation is the development of a new neural adaptive control system design methodology. In this methodology, a neural network architectures, the columnar fuzzy neural network architecture (CFNA) based on distributed representations, has been developed. The columnar fuzzy neural network architecture (CFNA) can learn approximations of nonlinear, multivariable mappings. The CFNA is used to implement plant identification and the controller subsystems of neural model reference adaptive control systems. The adaptive performance of the CFNA is due to its on-line learning capability. An extended version of the error back propagation learning algorithm is developed for this purpose. The extended error back propagation algorithm utilizes a quadratic measure of output error. The new design methodology provides stability in the Liapunov sense through a constraint mechanism imposed on the learning algorithm in the weight space of the CFNA.; Knowledge acquired from expert human operators of complex nonlinear plants in the form of fuzzy rules is mapped into an initial set of weights for the CFNA controller. The learning ability of the CFNA controller underlies its potential capabilities to eliminate erroneous rules, learn necessary new rules, and slightly modify or fine-tune existing rules. |
