Date of Award
Spring 2004
Document Type
Thesis - Restricted
Degree Name
Master of Science (MS)
Department
Electrical and Computer Engineering
First Advisor
Feng, Xin
Second Advisor
Povinelli, Richard
Third Advisor
Struble, Craig
Abstract
This thesis introduces a new method that combines supervised artificial neural networks and fuzzy logic inference for the research of feature analysis and evaluation. A multilayer neural network is employed to !team models of input patterns and to classify them into correct classes. The connection weights and hidden neurons of the neural network incorporate information such as relative importance of features or contributions of features for the discrimination of data. Since input patterns usually contain certain degree of uncertainty, a fuzzy logic system therefore is employed to infer out the feature evaluation results from the trained neural network. The neural network used in this study is a multilayer neural network trained by the backpropagation algorithm in a supervised manner. The outputs of the multilayer neual network represent each of classes of input patterns. The connection weights and hidden neuron effect of the fully trained multilayer neural network later then are interpreted by fuzzy logic to evaluate the importance and contributions of features for the discrimination of input patterns. The proposed method was applied on phenotype analysis for the consomic rat strains. Consomic rat strains are developed by the PhysGen project at the Medical College of Wisconsin (MCW) and those consomic rat strains are characterized by 214 phenotypes across 7 protocols. Those phenotypes are specifically targeted to the heart, lung, kidney, vasculature and blood functions in response to environmental stressors. Since environmental stressors such as hypoxia and salt intake can unmask deficiencies in normal homeostatic mechanisms and idiopathic mechanisms that contribute to disease, impacts of allelic variances of rat strains upon diseases that influence the heart, lung and blood systems can be revealed through the measured phenotypes. Comparing with other statistical methods, the proposed method can conduct phenotype analysis across multiple rat strains. Meanwhile analysts can quantitatively evaluate the contributions of phenotypes with respect to the classification results of rat strains. Also, the proposed method can quantitatively measure the relative significance of phenotypes and ranks phenotypes according to their importance measure.
Recommended Citation
Jiang, Hao, "Feature Evaluation through Supervised Neural Networks and Fuzzy Logic Technology and Applications on Phenotype Analysis for Consomic Rat Strains" (2004). Master's Theses (1922-2009) Access restricted to Marquette Campus. 4858.
https://epublications.marquette.edu/theses/4858