Computing failure probability is a fundamental task in many important practical problems. The computation, its numerical challenges aside, naturally requires knowledge of the probability distribution of the underlying random inputs. On the other hand, for many complex systems it is often not possible to have complete information about the probability distributions. In such cases the uncertainty is often referred to as epistemic uncertainty, and straightforward computation of the failure probability is not available. In this paper we develop a method to estimate both the upper bound and the lower bound of the failure probability subject to epistemic uncertainty. The bounds are rigorously derived using the variational formulas for relative entropy. We examine in detail the properties of the bounds and present numerical algorithms to efficiently compute them.
Introduction.Probability of failure is an important quantity in many applications involving system safety, risk management, reliability analysis, etc. Accurate computation of failure probability is thus of fundamental significance. A large amount of literature has been devoted to this task, ranging from more mathematically rigorous studies for model problems to more heuristic ones for engineering systems.This paper seeks to study the problem in a different context. That is, we study how to estimate the probability of failure when complete knowledge of the input probability distribution is not available. When lack of knowledge is the primary source of the uncertainty, it is often referred to as epistemic uncertainty. Though different definitions and classifications exist in the literature, in this paper we will use epistemic uncertainty to refer to the random inputs whose complete information about the probability distribution is not available and use aleatory uncertainty to refer to the random inputs whose probability distribution is fully prescribed.The study of the impacts of epistemic uncertainty is more difficult because many of the existing probabilistic tools do not readily apply. Some of the existing approaches include evidence theory [7], possibility theory [3], and interval analysis [6,12]. These method have their own advantages, though most do not address efficient numerical implementations. More recent studies employ approximation theory [1,5,13,8].This paper is largely motivated by the work of [2], where a method utilizing the variational formulas of relative entropy is developed to derive upper bounds for the predictions of epistemic uncertainty computations. In this paper we develop a methodology, similar to that of [2], for failure probability computation. Most notably we derive both the upper bound and the lower bound for the failure probability subject