Performance profiles for benchmarking of global sensitivity analysis algorithms

Abstract

Nowadays, sensitivity analysis (SA) is a methodology commonly used to identify important parameters that determine the behavior of the model. The SA of a model allows to determine how uncertainties in the model responses (outputs) can be assigned to the values of the model parameters (input variables). The related literature indicates that there are several methods to perform SA. This work addresses the benchmarking of four widely used methods for Global SA (GSA): Sobol-Jansen, Sobol-Baudin, Sobol-Owen and Sobol 2007, based on the concept of performance profile introduced by Dolan and Moré (2002) and the extension by Mahajan et al. (2012). To evaluate these methods, a set of 21 models and their variations were considered, which correspond to various applications in chemical engineering (such as heap leaching, water distribution network, milling, flotation circuit, among others). These comparisons show that, although the four GSA methods based on the decomposition of the variance proved to be quite stable, the Sobol-Jansen method presented the best performance, since it is the first to perform GSA in 83% of the models considered and maintains a high performance up to 100%.