Designing Computer Experiments – What Matters?

Host

Department of Applied Mathematics

Speaker

Jerome Sacks, Director Emeritus, Professor Emeritus of Statistical Science
National Institute of Statistical Sciences;Institute of Statistics and Decision Sciences, Duke University
http://www.niss.org/people/jerome-sacks

Description

Using a complex computer model for optimization, calibration, etc, typically requires a surrogate (approximation) to enable many (fast) predictions. Building a surrogate is done via a set of runs at designated inputs, that is, a computer experiment. Choices must be made to design the experiment and build the surrogate. An initial design requires a specification of a sample size (number of runs), \(n\), and a set of inputs, \(D\). Faced with a myriad of competing answers for choosing \(D\), what's a modeler to do?
This talk will describe evidence for evaluating competing methods leading to recommendations and findings, some at variance with common beliefs.