Calibrating high dimensional computationally expensive computer simulators with functional response
LE3 .A278 2015
Master of Science
Mathematics and Statistics
Mathematics & Statistics
Scientists across various fields rely on mathematical models in the form of computer simulators to explain complex phenomena. Because of this, the study of computer experiments has increased steadily over the past few decades. One common problem in computer experiments is calibration, that is, estimating unmeasurable parameters for future runs of the computer simulator. In this thesis we evaluate several methodologies for simplifying calibration of high dimensional and/or computationally expensive computer simulators with func-tional output. First, by scalarizing the computer simulator’s functional response, the calibration inverse problem becomes an optimization problem. We compare two scalarizing techniques: a simple Euclidean distance approach and a more elaborate log-likelihood ratio approach. Second, emulating the computationally expensive com-puter simulator, using a statistical surrogate model, may save time and resources. We investigate the performance of three statistical surrogate models: Gaussian Process models, Treed Gaussian Process models, and Bayesian Additive Regression Trees.
The author grants permission to the University Librarian at Acadia University to reproduce, loan or distribute copies of my thesis in microform, paper or electronic formats on a non-profit basis. The author retains the copyright of the thesis.