Robust Sensitivity Analysis of Complex Simulation Models subject to Noise

In this paper, Shapley values are used as a method for sensitivity analysis for models subject to noise and where the input variables are correlated. They are estimated using the ApproShapley algorithm, for which modifications are presented in order to increase the robustness of the estimates. It is also shown that input variables without functional influence on the model can still obtain a Shapley value larger than 0, when they are correlated with other input variables. The estimation of the Shapley values requires many evaluations of the model. To avoid this time-consuming step, a surrogate model is used. It is shown, that the uncertainty in the estimation of the Shapley values increases, when the magnitude of the noise increases. In a final application, the Shapley values for a compressor blade subject to manufacturing variability as well as wear and tear are computed. Analyzing different parameter sets, the profile parameters with the most functional influence are identified.