Computer Models and Uncertainty Quantification

• Activity Number: 669
• Type: Invited
• Date/Time: Thursday, August 13, 2015 : 10:30 AM to 12:20 PM
• Sponsor: Memorial
• Author(s): Jerome Sacks*
• Companies: NISS
• Keywords: computer model ; validation ; surrogate ; Gaussian process ; model discrepancy

Abstract

Using computer models demands effective quantification of uncertainty, the sources of which can be many. Validation of computer models (treating discrepancy between physical reality and its computer model representation) was bolstered by the work of Kennedy and O'Hagan. When some inputs are calibration parameters (therefore unknown), the challenge is to find effective ways to handle the inevitable confounding of calibration parameters with model discrepancy. In a series of papers, stimulated by applications in the automotive world and in geophysical problems, Susie and her co-workers devised methods to cope with challenges not only from the confounding but also, more generally, from the implementation of complex Bayesian computations, the need for effective surrogates (emulators, approximations) to costly-to-run computer models, and function output (such as time-space dependent output of the computer model ). Moreover, by capitalizing on a posterior distribution for all unknowns (especially the discrepancy term) it became possible to advance and explore ways to predict (with quantified uncertainty) behavior under new conditions, an essential goal of users of computer models.