Date of Award

Fall 1992

Document Type

Thesis - Restricted

Degree Name

Master of Science (MS)

Department

Computer Science

First Advisor

Brown, Ronald H.

Second Advisor

Belfore, Lee A. II

Third Advisor

Feng, Xin

Abstract

This thesis is concerned with the application of Kohonen topology-preserving neural network maps (KNNs) to nonlinear dynamic system identification problems. Kohonen maps were developed by Dr. Teuvo Kohonen in the early 1980's as a means of distributing neurons (points in n-dimensional space) throughout the n-dimensional space in an equiprobable manner. Here the concern for system approximation supersedes the desire for the equiprobable properties of the KNN and focuses on the achievement of a quality identification model. The quality of the identification model will be judged on the basis of speed of convergence to the KNN estimate, both in terms of computational complexity and in terms of iteration number, and on the basis of the KNN's ability to effectively represent the nonlinear system. Some of the ideas presented in this thesis include incorporating a priori knowledge into the KNN, preferential training the KNN, a modified conscience mechanism for training the KNN, and certain applications and examples of these KNNs as well as the use of uniform weight initialization and resetting [3] to nonlinear dynamic system identification using KNNs.

Share

COinS

Restricted Access Item

Having trouble?