Prediction and Inference with Incomplete Probabilistic Knowledge

摘要

A principal goal of modern reliability analysis is to design methods for appraising the reliability of large, logically complex systems that must operate at exceptionally high levels of reliability, but are subject to failure at uncertain times in uncertain ways. Nuclear power plants, the space shuttle, DARPA net and fly-by-wire aircraft are examples of such systems. Current practice is to design a probabilistic model of the stochastic behavior of a complex system and to augment it with numerical appraisal of the most important properties of the joint probability law that the model entails. A very different approach is to model the logical structure of a system so as to identify all logically obtainable events and to not model a probability law that governs event occurrence exactly. In place of assigning a specific probability law to a sample space that contains all obtainable system events, ask system analysts to assign numerical probabilities to some events that lie within their domain of expertise. This recasts the problem and changes the focus of the computational task.