Benchmarking Active Subspace methods of global sensitivity analysis against variance-based Sobol' and Morris methods with established test functions. (March 2022)

• Record Type: Journal Article
• Authors: Sun, Xifu; Croke, Barry; Jakeman, Anthony; Roberts, Stephen
• Is Part Of: Environmental modelling & software. Volume 149(2022)
• Publication Date: 2022-03
• Subjects: Global sensitivity analysis -- Test function -- Active subspace -- Activity score -- Variance-based Sobol' method -- Morris method; Environmental monitoring -- Computer programs -- Periodicals; Ecology -- Computer simulation -- Periodicals; Digital computer simulation -- Periodicals; Computer software -- Periodicals; Environmental Monitoring -- Periodicals; Computer Simulation -- Periodicals; Environnement -- Surveillance -- Logiciels -- Périodiques; Écologie -- Simulation, Méthodes de -- Périodiques; Simulation par ordinateur -- Périodiques; Logiciels -- Périodiques; Computer software; Digital computer simulation; Ecology -- Computer simulation; Environmental monitoring -- Computer programs; Periodicals; Electronic journals; 363.70015118
• DOI: 10.1016/j.envsoft.2022.105310
• Languages: English
• ISSNs: 1364-8152

Abstract

Abstract: Active Subspaces is a recently developed concept that identifies essential directions of the response surface of a model, providing sensitivity metrics known as activity scores. We compare activity scoring with the Sobol' and the Morris global methods using a series of well-known benchmark test functions with exactly computable sensitivities. In the ranking context, we analyse the accuracy, efficiency, impact of sampling method, convergence rate, and confidence interval estimation through both bootstrapping and replication. Heat maps that show both numerical rankings and underlying sensitivities with increasing sample size are introduced as a key visualization tool for such analysis. Convergence is also assessed through four previous measures. Activity scores match the total-effect sensitivity index of Sobol' and the absolute mean of elementary effect of Morris in most test cases. Activity scoring can be more computationally efficient. Its potential can be enhanced by expanding methods for approximating the gradient of the model function. Highlights: Comparison of activity scoring with two reference sensitivity analysis methods. Various test functions implemented for a comprehensive analysis. Heat maps display sensitivities and numerical rankings of factors. Four previous factor ranking measures employed and investigated. Comparison made of sampling methods and of bootstrap versus replication for confidence intervals.