THOMAS SANTNER 
Department of Statistics
Ohio State University
A Bayesian Composite Gaussian Process Model and its Application to Variable Selection
ABSTRACT
This talk will describe a variable selection methodology to identify the “active” inputs of a deterministic simulator code. While the basic methodology can be used in many settings, the description here will use an enhancement of a Bayesian composite Gaussian Process (BCGP) model. The basic BCGP model assumes that the simulator output can be described as the sum of draw from a GP that is meant to describe the long-range mean of the output plus a draw from an independent GP that describes small-scale deviations from the long-range mean. A prior is placed on the model parameters that insures the process describing the mean is smoother than that describing local deviations from the mean. In this talk, the correlation function is assumed to be a Gaussian correlation function. Based on a Bayesian fit to an enhancement of the basic BCGP model the inputs having smaller estimated correlation parameters are judged to be more active. A reference inactive input is added to the data to judge the size of the correlation parameter for inactive inputs. The importance of design is indicated.
Joint work with Casey Davis and Christopher Hans.
Joint seminar with the Department of Epidemiology and Biostatistics, TAMU School of Public Health
Friday, 10/26/18, BLOC 113, 11:30 AM
