Date of Award

6-1997

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Department of Electrical and Computer Engineering

First Advisor

Steven K. Rogers, PhD

Abstract

The construction of Multi Layer Perceptron (MLP) neural networks for classification is explored. A novel algorithm is developed, the MLP Iterative Construction Algorithm (MICA), that designs the network architecture as it trains the weights of the hidden layer nodes. The architecture can be optimized on training set classification accuracy, whereby it always achieves 100% classification accuracies, or it can be optimized for generalization. The test results for MICA compare favorably with those of backpropagation on some data sets and far surpasses backpropagation on others while requiring less FLOPS to train. Feature selection is enhanced by MICA because it affords the opportunity to select a different set of features to separate each pair of classes. The particular saliency metric explored is based on the effective decision boundary analysis, but it is implemented without having to search for the decision boundaries, making it efficient to implement. The same saliency metric is adapted for pruning hidden layer nodes to optimize performance. The feature selection and hidden node pruning techniques are shown to decrease the number of weights in the network architecture from one half to two thirds while maintaining classification accuracy.

AFIT Designator

AFIT-DS-ENG-97-01

DTIC Accession Number

ADA327557

Download

Share

COinS