|
Activity Number:
|
493
|
|
Type:
|
Invited
|
|
Date/Time:
|
Thursday, August 7, 2008 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Section on Statistical Computing
|
| Abstract - #300272 |
|
Title:
|
Global Optimization with Model Reference Adaptive Search and Expectation-Maximization
|
|
Author(s):
|
Jeffrey Heath*+ and Michael Fu and Wolfgang Jank
|
|
Companies:
|
Centre College and University of Maryland, College Park and University of Maryland, College Park
|
|
Address:
|
600 West Walnut St., Danville, KY, 40422,
|
|
Keywords:
|
EM Algorithm ; Model Reference Adaptive Search ; Global Optimization ; Mixture Models
|
|
Abstract:
|
It is well-known that the likelihood function of Gaussian mixture models can have many local, suboptimal maxima. While the Expectation-Maximization (EM) algorithm is the standard tool for estimating mixture-model parameters, it is easily trapped into such local maxima. We propose a systematic way of estimating mixture-model parameters based on the global optimization method Model Reference Adaptive Search (MRAS). One of the advantages of MRAS is that global convergence can be proved rigorously. We adapt MRAS to the Gaussian mixture model, and provide a theoretical proof of global convergence to the optimal solution of the likelihood function. We combine the updating procedure of MRAS with that of EM to construct the MRAS-EM algorithm for Gaussian mixtures. We provide numerical experiments which illustrate the performance of the MRAS algorithm relative to the EM algorithm.
|
