Title: Optimal-design search under the IMSPE objective
Authors: Selden Crary - Unaffiliated (United States) [presenting]
Abstract: Searches for optimal statistical designs of computer experiments, under the integrated mean-squared prediction error (IMSPE) objective, are often thought to encounter insurmountable problems because of ill-conditioning of the covariance matrix, whenever two or more trial design points are proximal in the design domain. The customary resolution is to disallow proximal design points, but doing so can disallow optimal designs that were the goal of the search. An alternative approach is to recognize the IMSPE is a member of a special class of pole-free, low-degree-truncated rational functions with essential discontinuities. Examples are provided of how optimal-design searches can be completed without excluding proximal points and how optimal designs with proximal points can be successfully used for metamodel generation.