Devin Francom, Bruno Sanso and Ana Kupresanin
11/01/2019 12:49 PM
Statistics
Many computer models output functional data, and in some cases, these functional data have similar, but misaligned, shape characteristics. We introduce a general ap- proach for building emulators for computer models that output misaligned functional data when key values in the functional response (landmarks) can be easily identified. This approach has two main parts: modeling the aligned (using the landmarks) func- tional data, and modeling the functions that map the misaligned data to the aligned space (warping functions). As the warping functions are required to be monotonic, we give special attention to modeling monotonic functional response data. We discuss how our approach can be used for a variety of typical emulators, such as Gaussian processes, Bayesian multivariate adaptive regression splines, and Bayesian additive re- gression trees, and how sensitivity analysis can be performed. We demonstrate these approaches to build emulators for a high-energy-density physics computer model used to simulate inertial confinement fusion ignition experiments.
