Resource Allocation for Uncertainty Quantification and Reduction

Abstract

This dissertation explores resource allocation methods for quantification and reduction of prediction uncertainty from computational models, in order to enable risk assessment and decision-making in engineering systems. The proposed methods focus primarily upon epistemic uncertainty due to approximate models and inadequate data, since it can be reduced by refining the models or collecting additional information. In order to reduce the uncertainty contribution from each of these sources, there is a tradeoff decision of cost vs. value. Methods for selecting among available modeling options of different fidelity are explored with the goal of simultaneously minimizing prediction error and computational expense. Then, validation methods for incorporating data uncertainty are considered, and a methodology for directly linking the validation result to the prediction of interest is proposed. A test selection approach is developed that explores the impact of data for both model calibration and model validation on the prediction of interest. Finally, the risk-based resource allocation problem is formulated. The proposed formulation motivates the total spending in the overall uncertainty quantification framework and provides insight about how much is economically efficient.