Quantifying uncertainty in environmental models using particle Markov chain Monte Carlo methods

Lawrence Murray (CSIRO)

NICTA SML SEMINAR

DATE: 2011-05-05
TIME: 11:00:00 - 12:00:00
LOCATION: NICTA - 7 London Circuit
CONTACT: JavaScript must be enabled to display this email address.

ABSTRACT:
Particle Markov chain Monte Carlo (PMCMC) methods are a recent family of techniques for joint parameter and state estimation in nonlinear state-space models. They employ Markov chain Monte Carlo (MCMC) over parameters and sequential Monte Carlo (or particle filters) over state, and are highly amenable to parallelisation in a high-performance computing environment.

This talk considers the application of PMCMC to environmental models, using a case study in marine biogeochemistry. Models in this field are nonlinear, non-Gaussian, have no closed-form transition density and may be accompanied by poor prior knowledge of initial conditions. In the PMCMC methodology, this can lead to noisy marginal likelihood estimates and poor mixing of the Markov chain. We consider how some of these issues might be mitigated, and in doing so how the methodology can be successfully applied to such difficult problems.
BIO:
Lawrence Murray is a research scientist with CSIRO Mathematics, Informatics and Statistics in Perth, Western Australia. He received the bachelor's degree in software engineering from the Australian National University in 2004, before the Ph.D. degree in informatics at the University of Edinburgh in 2009. His research interests include Bayesian inference in state-space models, machine learning, high performance computing and general-purpose computation on graphical processing units (GPGPU).