ABSTRACT
Abstract Safety assessment for highly critical systems differs from other performance evaluation tasks in various respects. Statistical evidence is usually insufficient for assigning model parameters with any confidence before operation of a new system, and for a long time into the operation period itself. On the other hand, a high degree of confidence is sought that the system will perform as safely as required. The assessors use disparate forms of evidence to reach this confidence, usually via their own expert judgement, a process which is poorly understood and subject to well-documented problems. Explicit, probabilistic formal reasoning is a way for the assessors to control the risks of intuitive judgement. We report on an exercise in using the formalism of Bayesian Belief Networks to support such formal probabilistic reasoning, the various difficulties encountered and methods for resolving them.
