Bayesian Analysis of Learning in Risk Analyses

Abstract

The analysis of technological risks must often proceed with sparse data concerning the systems being evaluated. To expand the empirical basis from which conclusions are reached, data concerning similar systems, or similar components within other systems, are used to estimate the probabilities of various safeguard failures. New systems and components, however, often incorporate improvements that increase their reliability. When systems or components improve over time, learning has taken place. The concept of learning has been studied through the use of learning curves. Learning curves are usually applied, however, to the progress of a single component rather than to progress across systems. In this article, a Bayesian methodology for analyzing “learning” with respect to failure rates across similar systems or components is presented. The goal is to provide a quantitative expression of the knowledge about the failure probability of a new system or component.

KEY WORDS:

• Learning curves
• Reliability
• Extra-binomial variation
• Sub-binomial variation