|
Activity Number:
|
158
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Monday, August 9, 2004 : 2:00 PM to 3:50 PM
|
|
Sponsor:
|
Section on Statistical Computing
|
| Abstract - #300981 |
|
Title:
|
Combining Data Augmentation and Over-relaxation to Speed Up the EM Algorithm
|
|
Author(s):
|
Robert Creecy*+
|
|
Companies:
|
U.S. Census Bureau
|
|
Address:
|
SRD 3000 FB-4, Stop 9100, Washington, DC, 20233,
|
|
Keywords:
|
EM ; algorithm ; optimization
|
|
Abstract:
|
The EM algorithm is popular primarily because it is easy to implement and has stable convergence properties. However, in some applications it can converge quite slowly. Various methods proposed to speed up the algorithm work by significantly complicating the writing of the computer code to achieve speed-up. Meng and van Dyke demonstrate that it is possible to improve the speed of the standard EM algorithm without sacrificing simplicity or stability by a class of data augmentation methods and introducing a working parameter to index the class. Another approach to speeding up EM while maintaining its simplicity has more recently been proposed by Salkhutidinov and Roweis that uses an adaptive over-relaxed bound optimization method. They derive a general adaptive over-relaxed EM algorithm that can be applied to any specific model for which a standard EM algorithm can be defined. This paper takes these two ideas and combines them in a single algorithm that speeds up EM while maintaining its simplicity. The method is illustrated on several examples.
|
- The address information is for the authors that have a + after their name.
- Authors who are presenting talks have a * after their name.
Back to the full JSM 2004 program |
