It is important to consider the relationship between sensitivity and uncertainty of a given model parameter. In general, a model prediction can be very sensitive to a given parameter across the range of plausible parameter values. However, in a specific case study, if that parameter can be specified with very little uncertainty, then it could also contribute very little to the overall model prediction uncertainty. This type of observation is an important part of any UA that can be determined by analysing results of the uncertainty propagation. Simple local measures of importance involve standard and more advanced derivative-based sensitivity measures (Figure 3) and are typically available from the results of approximate uncertainty propagation methods (e.g., FOEA and first-order reliability method). Results of a traditional Monte Carlo simulation experiment or GLUE- or MCMC-based sampling can be quickly analyzed to derive a more globally representative measure of the relative contribution of each input/ parameter uncertainty source based on the correlation between the model output of interest and the corresponding sampled values of each uncertain input/ parameter.
Was this article helpful?