A comparison of three metamodel-based methods for global sensitivity analysis: GP modelling, HDMR and LAR-gPC

AbstractThree metamodel-based method are compared for computing the Sobol’ indices of models featuring uncertain input parameters, namely Gaussian Process (GP) modelling, High-Dimensional Model Representation (HDMR) and Least Angle Regression-based generalized Polynomial Chaos expansions (LAR-gPC). The approaches are applied to the computation of the Sobol’ indices of several test functions featuring 3-50 input random variables. The computational costs and convergence rates associated with each scheme are compared. Eventually the strengths and weaknesses of each technique are highlighted and discussed