Global sensitivity analysis based on entropy

ABSTRACT: This paper dealswith the sensitivity analysis ofmodel output, using entropy. By the past, variance has been used, and some measures directly related to data distributions have been performed. We expect entropy to give another point of view, refining the other sensitivity analysis, andmay be to solve some situationswhere the other methods were useless. Two different indices based on entropy are presented in this work: those defined by Krzykacz-Hausmann 2001,, based on conditional entropy which use directly the definition of Shannon’s entropy, and those later suggested byLiu, Chen, and Sudjianto 2006, based on theKullbak-Leibler entropy, whichmeasure the difference between two probability distributions. After a brief presentation of the two entropy-based indices, two application cases are studied, one analytical function and one industrial computer code. Our goal is to show that some indices are better suited for some specific cases, and that they all enable different but relevant sensitivity analyses to be performed.