Augmentation of definitive screening designs (DSD+)

Abstract

Run efficient second order designs suitable for both factor screening and the fitting of a second-order response model are of interest in complex system characterisation efforts particularly when those efforts encounter budgetary constraints limiting their test programme. Leading designs are the smaller sized definitive screening design and the larger sized design, fractional Box-Behken design. When specific domain knowledge of system factor interactions is available, either a priori or a posteriori , that domain knowledge can be exploited to generate the definitive screening design plus. This paper describes how to generate the single DSD+, how to augment a DSD to arrive at the DSD+ and compares the performance of the two designs under a variety of effect sparsity, effect heredity and model noise conditions. The DSD+, with its extra k – 1 runs for k factors is shown to greatly improve the screen and model fitting capabilities of the DSD.