Parametric estimation of covariance function in Gaussian-process based Kriging models. Application to uncertainty quantification for computer experiments

Abstract : The parametric estimation of the covariance function of a Gaussian process is studied, in the framework of the Kriging model. Maximum Likelihood and Cross Validation estimators are considered. The correctly specified case, in which the covariance function of the Gaussian process does belong to the parametric set used for estimation, is first studied in an increasing-domain asymptotic framework. The sampling considered is a randomly perturbed multidimensional regular grid. Consistency and asymptotic normality are proved for the two estimators. It is then put into evidence that strong perturbations of the regular grid are always beneficial to Maximum Likelihood estimation. The incorrectly specified case, in which the covariance function of the Gaussian process does not belong to the parametric set used for estimation, is then studied. It is shown that Cross Validation is more robust than Maximum Likelihood in this case. Finally, two applications of the Kriging model with Gaussian processes are carried out on industrial data. For a validation problem of the friction model of the thermal-hydraulic code FLICA 4, where experimental results are available, it is shown that Gaussian process modeling of the FLICA 4 code model error enables to considerably improve its predictions. Finally, for a metamodeling problem of the GERMINAL thermal-mechanical code, the interest of the Kriging model with Gaussian processes, compared to neural network methods, is shown
Complete list of metadatas

Cited literature [116 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-00881002
Contributor : François Bachoc <>
Submitted on : Thursday, November 7, 2013 - 11:43:11 AM
Last modification on : Tuesday, May 14, 2019 - 11:03:29 AM
Long-term archiving on : Saturday, February 8, 2014 - 5:55:14 AM

Identifiers

  • HAL Id : tel-00881002, version 1

Citation

François Bachoc. Parametric estimation of covariance function in Gaussian-process based Kriging models. Application to uncertainty quantification for computer experiments. Statistics [math.ST]. Université Paris-Diderot - Paris VII, 2013. English. ⟨tel-00881002⟩

Share

Metrics

Record views

1319

Files downloads

2171