Date of Award

3-1997

Document Type

Thesis

Degree Name

Master of Science

Department

Department of Aeronautics and Astronautics

First Advisor

Edward A. Pohl, PhD

Abstract

Robust parameter estimation is successfully applied to the Mixed Weibull (seven parameter) using the Method of Minimum Distance and the Method of Maximum Likelihood. That is, parameters can now be estimated for a mixture of two Weibull distributions where the true populations are co-located, partially co-located or highly separated. Both techniques provided very robust estimates that were far superior to current parameter estimation techniques. Sample sizes as low as ten with mixing proportions down to 0.1 were investigated. For the MLEs, innovative bounding techniques are presented to allow consistent and correct convergence using any reasonable point estimate. The likelihood function is solved numerically as a non-linear constrained optimization using a quasi-Newton method. Minimum Distance Estimates (over three hundred scenarios investigated) are derived for some variation or combination of the mixing proportion and the location parameter(s), individually and simultaneously (the Anderson-Darling and Cramer-von Mises statistics were used). In tact, the MDE for the mixing proportion was so effective that future researchers should consider some permanent combination Primary measures of success were based on comparison of CDFs. Mean square error (MSE) and integrated absolute difference (LAF) between the estimated and true distributions were measured including confidence intervals.

AFIT Designator

AFIT-GOR-ENY-97M-01

DTIC Accession Number

ADA323181

Download

Share

COinS