IMPROVEMENTS TO GRADIENT-ENHANCED KRIGING USING A BAYESIAN INTERPRETATION
Open Access
- 1 January 2014
- journal article
- research article
- Published by Begell House Inc. Publishers in International Journal for Uncertainty Quantification
- Vol. 4 (3) , 205-223
Abstract
Cokriging is a flexible tool for constructing surrogate models on the outputs of computer models. It can readily incorporate gradient information, in which form it is named gradient-enhanced Kriging (GEK), and promises accurate surrogate models in >10 dimensions with a moderate number of sample locations for sufficiently smooth responses. However, GEK suffers from several problems: poor robustness and ill-conditionedness of the surface. Furthermore it is unclear how to account for errors in gradients, which are typically larger than errors in values. In this work we derive GEK using Bayes' Theorem, which gives an useful interpretation of the method, allowing construction of a gradient-error contribution. The Bayesian interpretation suggests the "observation error" as a proxy for errors in the output of the computer model. From this point we derive analytic estimates of robustness of the method, which can easily be used to compute upper bounds on the correlation range and lower bounds on the observation error. We thus see that by including the observation error, treatment of errors and robustness go hand in hand. The resulting GEK method is applied to uncertainty quantification for two test problems.Keywords
This publication has 37 references indexed in Scilit:
- Speeding up Kriging through fast estimation of the hyperparameters in the frequency-domainComputers & Geosciences, 2013
- Improving variable-fidelity surrogate modeling via gradient-enhanced kriging and a generalized hybrid bridge functionAerospace Science and Technology, 2013
- Analysis and implementation issues for the numerical approximation of parabolic equations with random coefficientsInternational Journal for Numerical Methods in Engineering, 2009
- Probabilistic collocation for period-1 limit cycle oscillationsJournal of Sound and Vibration, 2008
- Why do we need and how should we implement Bayesian kriging methodsStochastic Environmental Research and Risk Assessment, 2007
- A Bayesian tutorial for data assimilationPhysica D: Nonlinear Phenomena, 2007
- A Stochastic Collocation Method for Elliptic Partial Differential Equations with Random Input DataSIAM Journal on Numerical Analysis, 2007
- An implicit, exact dual adjoint solution method for turbulent flows on unstructured gridsComputers & Fluids, 2004
- Summary of the Delft University Wind Turbine Dedicated AirfoilsJournal of Solar Energy Engineering, 2003
- Modeling uncertainty in flow simulations via generalized polynomial chaosJournal of Computational Physics, 2003