Date of Award

Summer 1975

Document Type

Thesis - Restricted

Degree Name

Master of Science (MS)

Department

Civil, Construction, and Environmental Engineering

Abstract

In pattern recognition it is desirable that the classifier be easy to obtain and evaluate. To this end, Specht and Meisel each proposed classifiers based upon polynomial approximations of a potential function used to estimate the underlying distribution of the training samples. This estimate is then used in a Bayes' classifier. Meisel proposed a direct polynomial approximation of a normal distribution. This results in a classifier of the form: "formula" Specht proposed that the dot product be first expanded, taking the approximation only on the cross product terms. This results in a polynomial of order R rather than 2R. The resulting classifier is of the form: "formula" In view of the limitations imposed by the use of a finite approximation, it is reasonable to question the value of these classifiers in realizing general nonlinear decision surfaces. Specifically, do these classifiers result in more accurate classifications than a more simple linear classifier, and if so at what cost in increased complexity? To shed light on these more practical aspects of polynomial discriminant functions is the purpose of this Thesis.

Share

COinS

Restricted Access Item

Having trouble?