TY - GEN
T1 - Reliability-oriented Sensitivity Analysis using Shapley Additive Explanations and Polynomial Chaos Expansion
AU - Palar, Pramudita Satria
AU - Renganathan, Ashwin
N1 - Publisher Copyright:
© 2024 by the American Institute of Aeronautics and Astronautics, Inc. All rights reserved.
PY - 2024
Y1 - 2024
N2 - This paper introduces a novel sensitivity metric specifically designed for reliability analysis, utilizing Shapley Additive Explanations (SHAP) from game theory. The primary objective of this study is to present a reliability-oriented sensitivity index (ROSI) called SHAP-ROSI, which enables analysts to assess the relative impact of input variables on the failure probability. Unlike conventional SHAP which deals with global domains, the proposed metric calculates averaged SHAP values only within the failure domain. Not only does SHAP-ROSI serve as a reliability-oriented sensitivity metric, but it also provides valuable insights into the underlying mechanisms of how each input contributes to the failure probability through visualization in the form of a dependence plot. We conducted experiments on a three-variable analytical problem, a six-variable truss problem, and a seven-variable transonic airfoil problem to evaluate the efficacy of the proposed metric. The results demonstrate that the ranking produced by SHAP-ROSI aligns with the widely used Sobol indices. Furthermore, the SHAP visualization offers valuable information on which variables should be controlled to minimize the failure probability.
AB - This paper introduces a novel sensitivity metric specifically designed for reliability analysis, utilizing Shapley Additive Explanations (SHAP) from game theory. The primary objective of this study is to present a reliability-oriented sensitivity index (ROSI) called SHAP-ROSI, which enables analysts to assess the relative impact of input variables on the failure probability. Unlike conventional SHAP which deals with global domains, the proposed metric calculates averaged SHAP values only within the failure domain. Not only does SHAP-ROSI serve as a reliability-oriented sensitivity metric, but it also provides valuable insights into the underlying mechanisms of how each input contributes to the failure probability through visualization in the form of a dependence plot. We conducted experiments on a three-variable analytical problem, a six-variable truss problem, and a seven-variable transonic airfoil problem to evaluate the efficacy of the proposed metric. The results demonstrate that the ranking produced by SHAP-ROSI aligns with the widely used Sobol indices. Furthermore, the SHAP visualization offers valuable information on which variables should be controlled to minimize the failure probability.
UR - http://www.scopus.com/inward/record.url?scp=85192229009&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85192229009&partnerID=8YFLogxK
U2 - 10.2514/6.2024-0789
DO - 10.2514/6.2024-0789
M3 - Conference contribution
AN - SCOPUS:85192229009
SN - 9781624107115
T3 - AIAA SciTech Forum and Exposition, 2024
BT - AIAA SciTech Forum and Exposition, 2024
PB - American Institute of Aeronautics and Astronautics Inc, AIAA
T2 - AIAA SciTech Forum and Exposition, 2024
Y2 - 8 January 2024 through 12 January 2024
ER -