Title: Kriging Models:  Friend or Foe?

Speaker: Jay Martin, Applied Research Lab at Penn State University

Date/Time: Wednesday, June 13, 2007, 11:00am - 12:00

Location: CSRI Building/Room 90 (Sandia NM)

Brief Abstract: This seminar will identify both the good aspects and the bad aspects on using the kriging model. The kriging model is a statistical model that estimates a Gaussian spatial process. It is capable of estimating complex responses but at the cost of increased complexity of the model form and computational expense at estimating model parameters when compared to linear regression models. This seminar will briefly discuss the origins of the kriging model, derive it form as a best linear unbiased predictor, and present some of the more common methods used to estimate the kriging model parameters. These methods include graphical methods, cross-validation, Maximum Likelihood Estimation (MLE), and the Bayesian method of Markov Chain Monte Carlo (MCMC). Emphasis will be placed on MLE and details will be given on computationally efficient algorithms to optimize the log-likelihood function. The seminar will conclude with some diagnostic tests that can be performed to determine if there are outlier observations that are overly influencing the model parameter estimation and comparison test for competing model forms.

CSRI POC: Laura Swiler, (505) 844-8093