An Integrated Architecture and Feature Selection Algorithm for Radial Basis Neural Networks
AIR FORCE INST OF TECH WRIGHT-PATTERSON AFB OH SCHOOL OF ENGINEERING AND MANAGEMENT
Pagination or Media Count:
The research contribution of this thesis is the first known integrated architecture and feature selection algorithm for Radial Basis Neural Networks RBNNs. The objective is to apply the network iteratively to determine the final architecture and feature set used to evaluate a problem. Additionally, this thesis compares three different classification techniques, Discriminant Analysis DA, Feed-Forward Neural Networks FFN and RBNNs against several hard to solve problems. These problems were used to evaluate general classifier performance as well as the performance of the feature selection techniques. This thesis describes the classification techniques as well as the measures used to evaluate them. It next develops a new clustering technique used to determine the network architecture and the saliency measure used to select features for RBNNs. Next, the thesis applies these techniques to three general problems, Block-C, the University of Wisconsin Breast Cancer Data UWBCD and a noise corrupted version of Fishers Iris problem. Finally, the conclusions and recommendations for future research are provided.
- Computer Systems