4/14/2017

Post date: May 1, 2017 7:29:10 PM

Speaker: Mengyang Gu, Department of Applied Mathematics & Statistics, JHU

Title: an improved approach to Bayesian computer model calibration and prediction

Abstract:

We consider the problem of calibrating inexact computer model that might be biased to the reality. A discrepancy function is usually modeled and calibrated along with the computer model, leading to better results in prediction. However, calibration parameters in the computer model are sometimes unidentifiable because of the inclusion of the discrepancy function in the model. In this work, we propose a new stochastic process, called the scaled Gaussian stochastic process (S-GaSP) that bridges the gap between two frequently used approaches, namely the $L_2$ calibration approach and the GaSP calibration approach. A computationally feasible approach is introduced for this new model under the Bayesian framework with the default prior. The posterior propriety of this approach is also shown. The S-GaSP model not only provides a more general framework for calibration, but also performs surprisingly well in prediction by delicately combining the calibrated model and the discrepancy function. Numerical examples are provided to demonstrate the accuracy of the S-GaSP model in terms of the parameter estimation and out-of-sample prediction.