Title
Non-Collapsing Space-Filling Designs for Computer Experiments in Bounded Polygonal Regions
Speaker
Danel Draguljic, The Ohio State University
Abstract
When physical phenomena are difficult or even impossible to investigate using traditional physical experiments, scientists can sometimes simulate the input/output relationship of interest by implementing a mathematical model of the relationship using a (complicated) computer code. As in a physical experiment, the inputs to the simulator code can be purposely varied in order to determine their effects on the output(s). An experiment conducted using data obtained from a computer code is called a computer experiment. Obtaining the results from a computer experiment can be computationally intensive and may require many hours or even days to solve the underlying mathematical model. To understand better the process under investigation, a cheaper statistical surrogate to the computer code, sometimes called a metamodel or an emulator, is often developed based on a limited number of (training) runs from a computer experiment. The accuracy of emulators is compromised unless the inputs for the training data are as 'space-filling' as possible. Non-collapsing designs for computer experiments are highly desirable due to their good projection properties. This talk describes an algorithm for constructing designs that maximize a convex combination of the minimum interpoint and the average reciprocal distance when the input region is a bounded polytope in an effort to obtain space-filling and non-collapsing designs for computer experiments.
