Probability is perfect, but we can't elicit it perfectly

The challenge problems set out in Oberkampf, Helton, Joslyn, Wojtkiewicz and Ferson (this issue), hereafter referred to as OHJWF, raise a variety of interesting questions. The background to this workshop is the continuing growth in number, diversity and complexity of computer simulation codes being built to model a huge variety of important real-world systems. Users of such models are increasingly demanding to know about the uncertainties in the model outputs. Nobody seriously believes that the models are perfect. The model output will not predict exactly the real-world system behaviour, but how close can we expect it to be? There are several possible contributory causes of the error in a model’s predictions, and in order to understand and quantify the uncertainties in predictions it is important to recognise and analyse each of these components of uncertainty. One taxonomy of uncertainties in computer code outputs is given by Kennedy and O’Hagan (2001), and it is worth reviewing the main components brie‡y here.