2017/09/27 by Chkonia, Nikoloz, Crary, Selden
#FOS: Computer and information sciences #Methodology (stat.ME)
paper · doi:10.48550/arxiv.1709.09599
This paper is an extension of Part I of a series about Nu-class multifunctions. A method is presented for computing the integrated mean-squared prediction error (IMSPE) in the design of computer experiments, when the prediction domain is a circular disk. The method is extensible to more than two factors, to prediction domains other than squares or disks, and to a variety of assumed covariance functions. Three example optimal designs, under Gaussian covariance with known hyperparameters, are found using the method: an n=1 design centered on the disk; an n=2 continuously rotatable design with assumed inversion symmetry about the center of the disk; and an n=4 twin-point design similar to the n=4 twin-point design observed previously for a square prediction domain [1]. The four-point design on the disk demonstrates that non-round boundaries are not a prerequisite for the occurrence of twin-point optimal designs.