Bayesian Optimal Experimental Design for Inverse Scattering
Abstract
Bayesian inference provides a rational framework for learning models from data under uncertainty. The optimal experimental design (OED) problem then asks: how do we acquire the "best" data, i.e., the data from which we learn the most? OED is particularly crucial for Air Force systems, since physical or numerical experiments are often highly resource intensive. The goal of Bayesian OED is to optimally design the data acquisition (e.g., sensor locations, what quantities are measured, which experiments), so that uncertainty in the inferred parameters or some predicted quantity of interest is minimized with respect to a criterion. We focus on expected information gain, i.e., the expected Kullback{Leibler divergence between the prior and posterior with respect to the data. Since OED subsumes Bayesian inversion, and Bayesian inversion presents sign cant challenges in high dimensions and with expensive forward models, this leads to challenges of the highest order. Our work addressed these challenges by advancing the state of the art in methods for Bayesian inversion, optimal experimental design, and supporting optimization and machine learning tools.
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 26, 2021
- Accession Number
- AD1148955
Entities
People
- George Biros
- Omar Ghattas
- Youssef Marzouk
Organizations
- University of Texas at Austin