Inverting Regional Sensitivity Analysis to reveal sensitive model behaviors

TL;DR

This paper introduces iRSA, an inverse approach to Regional Sensitivity Analysis, which identifies specific input variations that explain sensitive behaviors in complex models, providing interpretable insights.

Contribution

The paper proposes iRSA, a novel method that inverts classical RSA to find regions in input space explaining sensitive model behaviors, formalized as an optimization problem.

Findings

iRSA effectively identifies input regions linked to sensitive outputs.

The method provides interpretable graphical characterizations of sensitivity.

Demonstrated on environmental models with time series outputs.

Abstract

We address the question of sensitivity analysis for model outputs of any dimension using Regional Sensitivity Analysis (RSA). Classical RSA computes sensitivity indices related to the impact of model inputs variations on the occurrence of a target region of the model output space. In this work, we invert this perspective by proposing to find, for a given target model input, the region whose occurrence is best explained by the variations of this input. When it exists, this region can be seen as a model behavior which is particularly sensitive to the variations of the model input under study. We name this method iRSA (for inverse RSA). iRSA is formalized as an optimization problem using region-based sensitivity indices and solved using dedicated numerical algorithms. Using analytical and numerical examples, including an environmental model producing time series, we show that iRSA can provide a new graphical and interpretable characterization of sensitivity for model outputs of various dimensions.