|
Activity Number:
|
532
|
|
Type:
|
Topic Contributed
|
|
Date/Time:
|
Thursday, August 2, 2007 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Section on Bayesian Statistical Science
|
| Abstract - #308384 |
|
Title:
|
Resampling for Inverse Problems Involving High-Dimensional Computer Simulators
|
|
Author(s):
|
Herbert Lee*+ and Matthew Taddy and Bruno Sanso
|
|
Companies:
|
University of California, Santa Cruz and University of California, Santa Cruz and University of California, Santa Cruz
|
|
Address:
|
School of Engineering, Santa Cruz, CA, 95064,
|
|
Keywords:
|
Bayesian statistics ; Importance Sampling ; Gaussian process ; soil permeability
|
|
Abstract:
|
The two most common Bayesian approaches to computer model inverse problems are surrogate models or direct MCMC using the simulator. However, in some cases the simulator may have already been run on a large number of input cases, and may no longer be available for additional runs. When tens of thousands or more runs are available, standard surrogate models such as Gaussian processes are too computationally expensive to fit. When no additional runs are available, direct MCMC is not possible. Thus we have developed a sampling importance resampling algorithm that is applicable in such cases. We demonstrate our approach on a hydrology problem where we want to draw posterior inference for an unobserved high-dimensional spatial permeability field using a million runs from a simulator of water flow.
|
