The IV AMMCS International Conference

Waterloo, Ontario, Canada | August 20-25, 2017

AMMSCS 2017 Plenary Talk

Gaussian Processes and the Statistical Analysis of Computer Experiments

William Welch (University of British Columbia)

The presentation will be in three parts: the history, the current use, and the future of Gaussian processes (GPs) in the statistical analysis of computer models.
  1. Why are GPs popular for computer experiments? Over the last quarter century GPs have become widespread in statistical, engineering, and other disciplines for the analysis of computer models. Their popularity probably stems from three major advantages: (1) they adapt to nonlinear input-output relationships in a data-adaptive way; (2) they may require relatively few runs of the computer model (important if the computer code is computationally demanding) and (3) they provide a measure of prediction uncertainty that is often realistic even when modelling a deterministic function. These basic properties will be reviewed.
  2. How are GPs used to achieve scientific and engineering objectives? Computer experiments are carried out with one or more of many purposes in mind: sensitivity analysis, calibration of unknown parameters, optimization, propagation of input uncertainty to statistical properties of the output, and so on. A broad overview of how GPs serve these goals will be given.
  3. What about really complex phenomena? The sample size, i.e., number of computer-model evaluations, for decent prediction accuracy can be impractically large for some high-dimensional, complex computer codes and even for toy problems. Some recent work to adapt GPs to expand the functions that can be usefully modelled will be briefly described.
Will Welch joined the Department of Statistics, UBC as a Professor in 2003 and was Head of Department from 2003 until 2008. Prior to that he was at the University of Waterloo for 16 years. He also holds the honorary title of Visiting Professor in the Business School, Loughborough University, UK.
Welch’s research spans computer-aided design of experiments, quality improvement, the design and analysis of computer experiments, statistical methods for drug discovery, and machine/statistical learning. The work is highly cited: please see http://scholar. In 2000 he won the American Sta- tistical Association’s Statistics in Chemistry Prize.
Welch has served on the editorial boards of the Annals of Applied Statistics, the Canadian Journal of Statistics, and the Journal of Uncertainty Quantification. He has also served as Associate Director of the Canadian Statistical Sciences Institute (CANSSI).