Convergence of sensitivity analysis methods for evaluating combined influences of model inputs. (September 2019)

• Record Type: Journal Article
• Authors: Awad, Majdi; Senga Kiesse, Tristan; Assaghir, Zainab; Ventura, Anne
• Is Part Of: Reliability engineering & system safety. Volume 189(2019)
• Page Start: 109
• Page End: 122
• Publication Date: 2019-09
• Subjects: Sensitivity analysis -- Combined action -- Morris' extension method -- Sobol method -- Carbonation model; Reliability (Engineering) -- Periodicals; System safety -- Periodicals; Industrial safety -- Periodicals; Fiabilité -- Périodiques; Sécurité des systèmes -- Périodiques; Sécurité du travail -- Périodiques; 620.00452
• DOI: 10.1016/j.ress.2019.03.050
• Languages: English
• ISSNs: 0951-8320

Abstract

Highlights: Morris' extension method was applied to evaluate the combined actions of model inputs. Study of three types of convergence of combined actions results using formal criteria. New Morris' extension indices was proposed to achieve convergent results at lower cost. Morris' extension method and total interaction Sobol indices provide the same results. Convergence of screening and ranking can be reached before sensitivity indices stabilize. Abstract: This work aims at studying Morris' extension method to evaluate the contribution of combined variations of inputs to variations of a model output. There is a lack of studies on the Morris' extension method concerning crucial choices of the adequate number of trajectories to distinguish influential and non-influential groups of pairs of inputs, rank pairs of inputs according to their relative importance and reach out the stability of sensitivity indices values. The Morris' extension method was studied regarding the three previous issues via applications on simple and complex models, in comparison with total interaction indices of Sobol. Formal criteria were implemented to assess the convergence of sensitivity analysis results. Sensitivity indices based on the median of mixed elementary effects (MEE) were investigated and found to be competing with classical ones based on the mean of MEE, to achieve convergent results.