Thursday 13 Mar 2014: Bayesian approaches to inverse problems: model selection and combination
Prof Andrew Webb - DSTL Porton Down
Harrison 101 15:00-16:00
Making inference about physical parameters of interest using observations from field experiments
together with detailed computer models of a complex physical system has received considerable
attention in the Bayesian inference literature. The forward model is usually a computer simulation
of a large scale physical system, often described by the solution of a complex set of differential
equations and widely known application domains include meteorological modelling; geophysics
research and modelling the dispersion of contaminants after an accidental of deliberate release.
The applications often share several features in common: observations are sparse (spatially and
temporally); disparate sensor types; computer simulations are demanding and approximations must
be made; the physics is not represented faithfully in the simulator - a source of model inadequacy;
numerical integration; aim is to address the inverse problem in order to achieve predictions for the
physical system at other times and locations.
In this presentation we
-present the work of Dstl on Bayesian approaches to inverse problems, where the specific
-problem is that of inverse dispersion modelling;
-present a practical application of Sequential Monte Carlo methods;
-discuss some of the technical issues;
-introduce model selection, averaging and combination approaches in inverse problems.