Abstract

Sensitivity analysis is an indispensable part of engineering model-based analyses. It provides a means for the engineer to peek into the model and understand those parts of it that are important to the task at hand. As a results there is a substantial amount of both research done and literature produced on the topic. However, sensitivity analysis is inherently an uncertainty-focused domain and as such it suffers from the problems discussed son far in the DAWS series of reports, namely the lack of practical distinction between aleatory and epistemic uncertainty, the confounding of uncertainty with probability theory and the resulting misinterpretation of results.

In this report we discuss the wide variety of methods for sensitivity analysis, as they apply to three main settings - engineering control, empirical planning and robustness analysis. We base the current exposition on the foundation laid so far in the previous reports in the DAWS series and more specifically on the DAWS Uncertainty Framework and the DAWS Report on Propagation of Uncertainty through Computer Models.

This is v. 1.1.

Changes from the previous version:

• Added a worked example for variance-decomposition (Sobol') sensitivity indices.

If you need access to previous versions, please email p.hristov@liv.ac.uk