An efficient approximation algorithm for variance global sensitivity by Bayesian updating

Abstract

Variance global sensitivity (VGS) is defined by the mean square difference between output expectation and conditional one on input realization, and it can calculate the mean contribution of the input within its distribution region and guide the effective modulation of output variance. The Monte Carlo simulation (MCS) and quasi MCS are commonly used to estimate VGS, but they are time-consuming respectively due to double-loop framework and computation related to input dimension. Thus, a novel method is proposed to estimate VGS by elaborately using Bayesian updating. In the proposed algorithm, the input realizations are firstly treated as observations to construct a likelihood function. Then by Bayesian updating, all conditional output expectations on different input realizations, which are required in estimating VGS and most time-consuming, can be obtained as the posterior and estimated by the sample of simulating the output expectation. The proposed algorithm shares the sample of solving output expectation to obtain all conditional ones required for solving VGS, which makes the computational effort of estimating VGS equivalent to that of estimating output expectation, thus improving the efficiency of estimating VGS. Numerical and engineering examples fully substantiate the novelty and effectiveness of this algorithm.