Winds from Parallel Bayesian Hierarchical Models: Introducing the Gibbs Coupler

• Activity Number: 398
• Type: Topic Contributed
• Date/Time: Thursday, August 12, 2004 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics and the Environment
• Author(s): Timothy Hoar*+ and Doug Nychka and Ralph F. Milliff
• Companies: National Center for Atmospheric Research and National Center for Atmospheric Research and Colorado Research Associates
• Keywords: parallel computing ; spatio-temporal modelling ; Gibbs Sampler

Abstract

Advances in computing power are allowing researchers to use Bayesian Hierarchical Models (BHMs) on problems previously considered computationally infeasible. The BHM in question is estimated in a massive Gibbs Sampler and combines the information from a scatterometer on board a polar-orbiting satellite and the result of a numerical weather prediction model to produce an ensemble of high-resolution tropical surface wind fields with physically realistic variability at all spatial scales. Since the model is only valid for a time period (epoch) much shorter than the data record, a separate model for each epoch is estimated. However, each model has an autoregressive component which connects the current time with the previous time. If the calculations are kept separate, the full conditional distributions for the states at the endpoints are not conditioned on states outside the epoch. This presentation discusses the procedure of simultaneously estimating multiple BHMs on a massively parallel platform while still providing a mechanism for maintaining proper distributions for the states at the endpoints.