Design of Experiments for Screening

Abstract

The aim of this paper is to review methods of designing screening experiments, ranging from designs originally developed for physical experiments to those especially tailored to experiments on numerical models. The strengths and weaknesses of the various designs for screening variables in numerical models are discussed. First, classes of factorial designs for experiments to estimate main effects and interactions through a linear statistical model are described, specifically regular and nonregular fractional factorial designs, supersaturated designs and systematic fractional replicate designs. Generic issues of aliasing, bias and cancellation of factorial effects are discussed. Second, group screening experiments are considered including factorial group screening and sequential bifurcation. Third, random sampling plans are discussed including Latin hypercube sampling and sampling plans to estimate elementary effects. Fourth, a variety of modelling methods commonly employed with screening designs are briefly described. Finally, a novel study demonstrates six screening methods on two frequently-used exemplars, and their performances are compared.

A review on simulation-based metamodeling in emergency healthcare: methodology, applications, and future challenges

“…For multivariate problems, Barton recommends conducting a sensitivity analysis to determine influential design parameters. We refer the reader to Woods and Lewis 29 for more details on screening methods.…”

Cluster sampling for Morris method made easy

“…MM defines the elementary effects for factor j obtained at x as where e j denotes the unit vector in the direction of the j th coordinate and Δ is a predefined integer multiple of 1/( p − 1) such that x + e j Δ ∈ Ω. It is intuitive from (1) that d j ( x ) can be regarded as the partial derivative of Y( x ) at x with respect to x j when Δ is small (Woods & Lewis, 2017). More importantly, randomly sampling the factor combination x from Ω produces an induced distribution of elementary effects corresponding to factor j ; denote it by F j .…”

“…Once the active factors have been identified, one can restrict attention to varying their values while setting the inactive factors to nominal values in the computational model. Factor screening approaches typically fall into two categories: model-based and model-free (Woods & Lewis, 2017), depending on whether an approximation model or metamodel is assumed for describing the underlying input-output (I/O) relationship implied by the computational model. Many factor screening techniques rely on some lower-order polynomial regression models, such as supersaturated designs (Phoa et al, 2009;Xing et al, 2013), group designs (Morris, 2006), frequency domain designs (Sanchez et al, 2006), and sequential bifurcation (Ankenman et al, 2015;Shen et al, 2010;Shen & Wan, 2009;Shi et al, 2014;Wan et al, 2006Wan et al, , 2010.…”

New bounds and search for maximin distance U‐type designs

“…By applying this algorithm, we obtain many new maximin distance designs with 2, 3, 4, and 5 levels. These designs are particularly suitable as initial designs for factor screening in physical and computer experiments involving a large number of factors, as emphasized by Moon et al (2012) and Woods and Lewis (2017).…”