EXTENDING AND IMPROVING DESIGNS FOR LARGE-SCALE COMPUTER EXPERIMENTS

Abstract

This research develops methods that increase the inventory of space-filling designs (SFDs) for large-scale computer-based experiments. We present a technique enabling researchers to add sequential blocks of design points effectively and efficiently to existing SFDs. We accomplish this through a quadratically constrained mixed-integer program that augments cataloged or computationally expensive designs by optimally permuting and stacking columns of an initial base design to minimize the maximum absolute pairwise correlation among columns in the new extended design. We extend many classes of SFDs to dimensions that are currently not easily obtainable. Adding new design points provides more degrees of freedom for building metamodels and assessing fit. The resulting extended designs have better correlation and space-filling properties than the original base designs and compare well with other types of SFDs created from scratch in the extended design space. In addition, through massive computer-based experimentation, we compare popular software packages for generating SFDs and provide insight into the methods and relationships among design measures of correlation and space-fillingness. These results provide experimenters with a broad understanding of SFD software packages, algorithms, and optimality criteria. Further, we provide a probability-distribution model for the maximum absolute pairwise correlation among columns in the widely used maximin Latin hypercube designs.