Including Epistemic Uncertainty in Engineering Decision Making

Abstract

Most engineering systems are designed on the basis that their probability of under- performance is appropriately low. This is done either explicitly, by calculating this probability of failure, or implicitly, by prescribing a demand on the system that is large enough. In either case, engineering design requires estimation of probabilistic quantities that are subjected to different sources of uncertainty. The most common approach in practice, is to only include what is typically considered aleatory uncertainty. That is, the uncertainty in the future occurrence of events, or future demand on the system (e.g. probability distribution of extreme flooding). The inclusion of uncertainty due to model specification and quantity and quality of data used to build the probabilistic models is less acknowledged when defining design quantities.

In the present work, we discuss the inclusion of epistemic uncertainty in the estimation of probability models, and its implication for design and decision making. The analysis is done via a synthetic example of a simple design problem specified by a prescribed return period for the demand, or the performance of the system. Particularly, we propose and discuss the use of Bayesian predictive posterior estimates of the quantities of analysis as a transparent and consistent way of including epistemic uncertainties in design.

Since engineering design deals almost exclusively with the tails of the probability distributions (low frequency events), we show that including epistemic uncertainty might result in a higher probability of failure than expected, or equivalently, in higher demands for a specified return period. This is shown to be particularly exacerbated in data-scarce, and/or very low probability design contexts.