A Bayesian view of assessing uncertainty and comparing expert opinion (Q1116567)

The author argues in favor of a Bayesian approach to assess uncertainty in the fields of artificial intelligence and expert systems (or else when you can consult an expert). The meaning of `Bayesian approach' is not the naive one (i.e. repeated use of Bayes' theorem), but is intended as the ubiquitous specification of probabilities to represent uncertainty, assessing a fully-detailed, high-dimensional joint probability distribution for all the quantities about which the decision maker is uncertain in order to be able to choose an effective decision. Once the class of well-calibrated experts has been defined, a partial ordering is introduced in this class by means of the concept of refinement. Moreover, a total ordering that preserves the aforementioned partial ordering can be induced by means of the so-called strictly proper scoring rules. Finally, there is no mention of de Finetti's theory which allows a gradual and coherent extension of subjective probability, starting from a few events of initial interest: this approach is very general and also useful for testing the consistency of probability assessments, whose overall design on the set of all possible envisaged situations cannot be easily achieved, and which is computational feasible.