|
Activity Number:
|
164
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Monday, August 3, 2009 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Section on Bayesian Statistical Science
|
| Abstract - #304128 |
|
Title:
|
A Multi-Scale Adaptive Metropolis Algorithm
|
|
Author(s):
|
Matthew J. Heaton*+ and Scott Schmidler
|
|
Companies:
|
Duke University and Duke University
|
|
Address:
|
Box 90251, Durham, NC, 27708-0251,
|
|
Keywords:
|
Adaptive MCMC ; Multimodal distribution ; Expected squared Jumping distance
|
|
Abstract:
|
Adaptive MCMC algorithms have seen renewed research interest in recent years. However, popular adaptive algorithms can perform very poorly on multimodal distributions by adapting to local (as opposed to global) covariance structure. Here, a multi-scale adaptive Metropolis (MAM) algorithm is proposed which adapts a range of distinct proposal distributions at varying resolutions. The weights given to each scale are adapted based on expected squared jumping distance giving the algorithm greater potential to explore the entire support of the target distribution with minimal overhead. Formal convergence is established via conditions given by Roberts and Rosenthal (2007). The performance of the MAM algorithm is evaluated via simulation studies on complex multimodal distributions.
|
