EcoSta 2017: Start Registration
View Submission - EcoSta2017
A0416
Title: A sequential split-conquer-combine approach for Gaussian process modeling in computer experiments Authors:  Ying Hung - Rutgers University (United States) [presenting]
Min-ge Xie - Rutgers University (United States)
Abstract: Gaussian process (GP) models are widely used in the analysis of computer experiments. However, two critical issues remain unresolved. One is the computational issue in GP estimation and prediction where intensive manipulations of an $n$-by-$n$ correlation matrix are required and become infeasible for large sample size $n$. The other is how to improve the naive plug-in predictive distribution which is known to underestimate the uncertainty. We introduce an unified framework that can tackle both issues simultaneously. It consists of a sequential split-conquer procedure, an information combining technique using confidence distributions (CD), and a CD-based predictive distribution. This framework provides estimators and predictors that maintain the same asymptotic efficiency as the conventional method but reduce the computation dramatically. The CD-based predictive distribution contains comprehensive information for statistical inference and provides a better quantification of predictive uncertainty comparing with the plug-in approach. Simulations are conducted to evaluate the accuracy and computational gains. The proposed framework is demonstrated by a data center example based on tens of thousands of computer experiments generated from a computational fluid dynamic simulator.