Quantifying uncertainty importance when inputs are correlated

摘要

Some remaining open issues in the area of global quantitative sensitivity analysis are explored. The use of variance based methods for the sensitivity analysis of non-additive computational models in the presence of correlated input is investigated. While local sensitivity analysis approaches are valuable and still widespread in many branches of physical sciences, the problem setting is quite different for practitioners involved in safety or risk analysis. For these the degree of variation of the input factors is material, because a quantitative assessment of the uncertainty around some best estimate value for Y is customarily being sought. In this context, sensitivity analysis is aimed at priority setting, to determine what factor mostly needs better determination. All this inevitably requires a global quantitative approach, although various approaches to global sensitivity analysis are available which are mostly characterised by different definitions of sensitivity analysis. The paper is devoted to this problem, where different approaches to the concept of importance are investigated and discussed.