Title: Fast algorithms for iterative Bayesian PDE inversion
Other Titles: Bayessche inverse PDE Probleme
Language: English
Authors: Stadlbauer, Benjamin 
Qualification level: Diploma
Keywords: Bayesian Estimation; PDE Inversion; Iterative Metropolis
Advisor: Heitzinger, Clemens 
Issue Date: 2018
Number of Pages: 27
Qualification level: Diploma
Abstract: 
In this article we present algorithms to perform Bayesian inversion based on physical models, in particular based on partial differential equations. We are interested in identifying parameters of the PDEs that affect functionals of the solutions for which experimental data are available. Markov-chain Monte-Carlo methods like the Metropolis algorithm provide the algorithmic foundation. We present an adaptation and extension of this procedure to be able to perform multi-dimensional Bayesian inversion where not all measurements have to be present prior to the estimation, but become available in batches as time passes. Namely, based on the Delayed-Rejection Adaptive-Metropolis (DRAM) algorithm, we introduce an iterative approach, where we use the posterior of the last Metropolis run as the prior for the new run, where we use new measurements in each iteration. This allows to examine some information about the parameters already during the estimation process. Therefore a density estimator needs to be introduced. We make use of the Improved Fast Gauss Transform (IFGT) which allows us to perform a faster evaulation of the kernel density estimator, reducing the runtime from quadratic to nearly linear. Applications using a nano-capacitor sensor array are presented as well, where we estimate the radii of over 4000 nano-electrodes.
URI: https://resolver.obvsg.at/urn:nbn:at:at-ubtuw:1-117239
http://hdl.handle.net/20.500.12708/6999
Library ID: AC15188969
Organisation: E101 - Institut für Analysis und Scientific Computing 
Publication Type: Thesis
Hochschulschrift
Appears in Collections:Thesis

Files in this item:

Show full item record

Page view(s)

8
checked on Apr 9, 2021

Download(s)

33
checked on Apr 9, 2021

Google ScholarTM

Check


Items in reposiTUm are protected by copyright, with all rights reserved, unless otherwise indicated.