Toward a more robust variance-based global sensitivity analysis of model outputs

Tong, C

Lawrence Livermore National Laboratory

关键词: Sensitivity Analysis;    Sensitivity;    Sampling;    Accuracy;    99 General And Miscellaneous//Mathematics, Computing, And Information Science;   

DOI  :  10.2172/923115
RP-ID  :  UCRL-TR-235561
RP-ID  :  W-7405-ENG-48
RP-ID  :  923115

美国|英语

来源: UNT Digital Library

【 摘 要 】

Global sensitivity analysis (GSA) measures the variation of a model output as a function of the variations of the model inputs given their ranges. In this paper we consider variance-based GSA methods that do not rely on certain assumptions about the model structure such as linearity or monotonicity. These variance-based methods decompose the output variance into terms of increasing dimensionality called 'sensitivity indices', first introduced by Sobol' [25]. Sobol' developed a method of estimating these sensitivity indices using Monte Carlo simulations. McKay [13] proposed an efficient method using replicated Latin hypercube sampling to compute the 'correlation ratios' or 'main effects', which have been shown to be equivalent to Sobol's first-order sensitivity indices. Practical issues with using these variance estimators are how to choose adequate sample sizes and how to assess the accuracy of the results. This paper proposes a modified McKay main effect method featuring an adaptive procedure for accuracy assessment and improvement. We also extend our adaptive technique to the computation of second-order sensitivity indices. Details of the proposed adaptive procedure as wells as numerical results are included in this paper.