Date of Award

12-2017

Degree Name

Doctor of Philosophy

Department

Electrical and Computer Engineering

First Advisor

Dr. Ikhlas Abdel-Qader

Second Advisor

Dr. Bradley Bazuin

Third Advisor

Dr. Azim Houshyar

Abstract

A Multiple Classifier System (MCS) is designed to combine classification results of an ensemble of different classifiers and consequently to produce the highest possible classification output. MCS has recently drawn growing attention and has become a necessity, especially when a problem involves a large class of noisy data or when using a single pattern classifier that has serious drawbacks in its results. A wide range of pattern recognition applications have benefited from the implementation of MCS, these include areas such as handwriting recognition, incremental learning, data fusion, feature selection, and a large variety of medical applications.

To achieve optimal ensemble performance, two design components must be optimized carefully which are diversity and the selection of combining rule. This dissertation is focused on designing an ensemble decision combining rule which leads the MCS to deliver the highest possible accuracy. Several models for decision combining rules, using an ensemble system of N classifiers and M classes, are developed. The proposed system can be considered as a unifying framework that works with any algebraic decision combining rule. While the results affirm that there is no single decision combining rule that can outperform in every classification problem, they clearly present the framework to design an optimum decision combining rule based on the statistics of the classifiers. Based on the predication extracted from the theoretical models, a novel algorithm that achieves optimal classification accuracy is presented in this study.

The proposed algorithm is tested on six datasets, the experimental results agree with the trend predicted by theoretical derivations. Results based on the proposed algorithm show that the performance of an ensemble always achieves at least the performance of the best performing individual classifier and evades selecting the least performing classifier. In addition, the results of the proposed algorithm show a comparable performance in classification accuracy compared to the random forest with less computational operations which makes it a good candidate for real time classification problems. Finally, the proposed model serves as an in-depth exploration into the performance of MCS and brings to the forefront of classification research significant insights.

Access Setting

Dissertation-Campus Only

Restricted to Campus until

12-2019

Share

COinS