M&S Uncertainty Quantification

An emerging consensus in engineering holds that aleatory uncertainty should be propagated by traditional methods of probability theory but that epistemic uncertainty may require methods that do not confuse incertitude with variability by requiring every possibility be associated with a probability that it occurs. Therefore, although Monte Carlo shells that re-run calculations many times while varying input values according to specified distributions are useful, they are insufficient to properly account for epistemic and aleatory uncertainty. Three things are needed to build on the new consensus. The first is a clear and comprehensive roadmap for how various important problems (e.g., arithmetic and logical evaluations, backcalculations, sensitivity analyses, etc.) can be solved under this view. The second is software that enables analysts to automatically propagate through simulations the uncertainty implied by the significant digits of the numbers specifying a model. Such software would require very little training in uncertainty analysis to be useful to analysts. The third need is a software library of recommended methods for common calculations that is usable by modelers and analysts who may not themselves be experts in uncertainty quantification but who recognize the need for and benefits from it.