Adaptive Bayesian Calibration of Computer Models Using Sequential Monte Carlo

• Activity Number: 179
• Type: Contributed
• Date/Time: Monday, August 2, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Gardar Johannesson*+
• Companies: Lawrence Livermore National Laboratory
• Keywords: Bayesian computer model calibration ; sequential design of computer experiments ; uncertainty quantification

Abstract

As computer model simulations become ever more complex and computationally demanding, so becomes the task of characterizing their accuracy; uncertainty quantification (UQ). One branch of UQ aims at characterizing input uncertainty by comparing model simulations to observations. Given a fixed and continuous block of computer time for input UQ (on a supercomputer), we propose an adaptive and sequential Bayesian approach to input parameter calibration (input UQ). Our approach is able to adaptively decide which computer simulations to execute and rapidly carry out posterior input UQ analysis given new simulations. This is accomplished by fusing together (1) sequential design of computer experiments, (2) Bayesian computer model calibration using statistical emulators, and (3) adaptive sequential Monte Carlo methods.