An EM approach to blind Kriging with adaptive mean for surrogate modeling

The present work introduces a new blind Kriging model for approximating expensive computer experiments. The core contribution is the imposition of a nonstationary Gaussian prior on the regression parameter so as to achieve more adaptive mean modeling along with variable selection. With the technique of Cholesky decomposition, all the parameters involved in blind Kriging are estimated by the expectation-maximization scheme automatically. Empirical studies with a benchmark dataset demonstrate that our method achieves comparative performance in terms of the prediction accuracy.