Evaluating the probability that the outputs of a computer code with random inputs will meet a set of evaluation criteria

摘要

Several methods (Bayesian, testing, confidence) are presented for evaluating the probability 'p' that the outputs of a computer code, given random inputs, will meet a set of evaluation criteria. They all lead to the same expression for the confidence β that p exceeds some desired value q, based on the results of running the code independently N times. This expression is independent of the number of evaluation criteria. The 'coverage' approach to sampling, which yields a confidence β_(co) that the extreme values of the outputs from N independent runs of the code bound a region into which a single run of the code will fall with at least a probability γ, is shown to solve a different problem and to be unsuitable as a method for evaluating 'p', except in the special case when there is only one evaluation criterion. The method has broad application. It is developed in this paper in the context of evaluating the likelihood that a nuclear plant will meet the applicable safety criteria.