Date of Award

1991

Document Type

Thesis

Publisher

Edith Cowan University

Degree Name

Bachelor of Applied Sciences Honours

Faculty

Faculty of Science and Technology

First Supervisor

Dr J.W.L. Millar

Abstract

It is currently believed that artificial neural network models may form the basis for inte1ligent computational devices. The Boltzmann Machine belongs to the class of recursive artificial neural networks and uses a supervised learning algorithm to learn the mapping between input vectors and desired outputs. This study examines the parameters that influence the performance of the Boltzmann Machine learning algorithm. Improving the performance of the algorithm through the use of a naïve mean field theory approximation is also examined. The study was initiated to examine the hypothesis that the Boltzmann Machine learning algorithm, when used with the mean field approximation, is an efficient, reliable, and flexible model of machine learning. An empirical analysis of the performance of the algorithm supports this hypothesis. The performance of the algorithm is investigated by applying it to training the Boltzmann Machine, and its mean field approximation, the exclusive-Or function. Simulation results suggest that the mean field theory approximation learns faster than the Boltzmann Machine, and shows better stability. The size of the network and the learning rate were found to have considerable impact upon the performance of the algorithm, especially in the case of the mean field theory approximation. A comparison is made with the feed forward back propagation paradigm and it is found that the back propagation network learns the exclusive-Or function eight times faster than the mean field approximation. However, the mean field approximation demonstrated better reliability and stability. Because the mean field approximation is local and asynchronous it has an advantage over back propagation with regard to a parallel implementation. The mean field approximation is domain independent and structurally flexible. These features make the network suitable for use with a structural adaption algorithm, allowing the network to modify its architecture in response to the external environment.

Share

 
COinS