Design for sensitivity analysis (Q2799901)

``Sensitivity analysis is the study of how uncertainty in the output of a model can be apportioned to different sources of uncertainty in the model input'' (from the Introduction). The authors consider models with \(k\) inputs and scalar output. Three measures of sensitivity are proposed: elementary effects of the \(i\)-th input (\(i = 1,\dots,k\)) and two variance-based sensitivity indices. In this paper the authors present a survey on methods for estimating these sensitivity measures. First they discuss the case where the analyst can choose (design) the inputs. For ``cheap'' models (for which a single model evaluation will take one second or less of computer time) they propose Monte Carlo estimation using good space-filling input designs. For ``expensive'' models designs based on Fourier analysis (FAST), random balance designs and screening designs are preferred. In the case with given input data the authors discuss the use of emulators (a simpler metamodel) and of scatter-plot smoothing.NEWLINENEWLINEFor the entire collection see [Zbl 1327.62001].