Bayesian Statistics Seminar
North Carolina State University
presents
Dr. Murali Haran
Penn State University
"Towards Automated MCMC Algorithms for a Class of Spatial Models"
ABSTRACT
Markov chain Monte Carlo (MCMC) algorithms provide a general recipe for estimating properties of complicated distributions. While their use has become commonplace, constructing and running MCMC algorithms is still far from automatic even for many commonly used models. Issues encountered by MCMC users include having to fine-tune the algorithm for a given model to each new data set, determining appropriate starting values, deciding whether the algorithm is producing accurate estimates, and determining an appropriate length for the Markov chain. I will describe some approaches for automating these decisions in the context of linear and generalized linear Gaussian random field models. These are flexible models with a wide range of applications. I will discuss analytical approximations and block sampling approaches for constructing provably fast mixing, automated MCMC algorithms that take advantage of recent developments in MCMC theory and, in some cases, the ready availability of multiple processors. While my focus will be on spatial models, much of my discussion of MCMC algorithm construction and output analysis applies more broadly to MCMC-based inference.
Thursday, November, 5, 2009
4:00 - 5:00 pm
450 Riddick Hall