A1640
Title: Comparing emulators systematically
Authors: Devin Francom - Los Alamos National Laboratory (United States) [presenting]
Abstract: Many emulation methods exist, and no one emulator works best for all situations. A systematic comparison of a collection of emulators is provided. The accuracy of predictions and uncertainty estimates is compared using a broad set of test functions and computer model datasets (all with scalar response). The comparison includes emulation methods like Gaussian processes, Bayesian additive regression trees, Bayesian multivariate adaptive regression splines, Bayesian projection pursuit regression, Bayesian neural networks, and Bayesian polynomial chaos. Further, an R package, duqling, is introduced to make the comparison reproducible. For example, if someone wonders how their favorite emulation method would compare (or one of the above methods under different settings), they can train and test under the exact same conditions of this analysis by using the R package.
