Getting fancy with probability (Q1187224)

The author first outlines some non-Bayesian theories of uncertainty where the ordinary probability function is replaced, e.g., by a probability interval, by a second-order distribution over a set of probability functions, or by a ``belief-function'' à la Shafer. It is argued that all these non-standard approaches can be viewed as special cases of the approach taking probabilities to be determined by sets of probability functions. Second, the usual principle of conditionalization (for updating probabilities in the light of new evidence) is compared with some non- standard approaches, e.g., Dempster's rules of ``combination'' and of ``conditioning'' for belief functions. Third, the author presents an example to show that the principle of conditionalization can be incompatible with two commonly accepted rules: the ``direct inference'' rule which says us to equate the probability of a single instance with the statistic frequency in the given reference class, and the ``strength rule'' which ``says to apply your general knowledge of chances when nothing more specific conflicts with it''.