Imprecise probability generalizes probability theory to allow for partial probability specifications, and is applicable when information is scarce, vague, or conflicting, in which case a unique probability distribution may be hard to identify. Thereby, the theory aims to represent the available knowledge more accurately. Imprecision is useful for dealing with expert elicitation, because:
People have a limited ability to determine their own subjective probabilities and might find that they can only provide an interval. As an interval is compatible with a range of opinions, the analysis ought to be more convincing to a range of different people.
Introduction Uncertainty is traditionally modelled by a probability distribution, as developed by Kolmogorov, Laplace, de Finetti, Ramsey, Cox, Lindley, and many others. However, this has not been unanimously accepted by scientists, statisticians, and probabilists: it has been argued that some modification or broadening of probability theory is required, because one may not always be able to provide a probability for every event, particularly when only little information or data is available—an early example of such criticism is Boole's critique of Laplace's work—, or when we wish to model probabilities that a group agrees with, rather than those of a single individual. Perhaps the most common generalization is to replace a single probability specification with an interval specification. Lower and upper probabilities, denoted by P _ ( A ) {\displaystyle {\underline {P}}(A)} and P ¯ ( A ) {\displaystyle {\overline {P}}(A)} , or more generally, lower and upper expectations (previsions), aim to fill this gap. A lower probability function is superadditive but not necessarily additive, whereas an upper probability is subadditive. To get a general understanding of the theory, consider:
the special case with P _ ( A ) = P ¯ ( A ) {\displaystyle {\underline {P}}(A)={\overline {P}}(A)} for all events A {\displaystyle A} is equivalent to a precise probability
P _ ( A ) = 0 {\displaystyle {\underline {P}}(A)=0} and P ¯ ( A ) = 1 {\displaystyle {\overline {P}}(A)=1} for all non-trivial events represents no constraint at all on the specification of P ( A ) {\displaystyle P(A)}
We then have a flexible continuum of more or less precise models in between. Some approaches, summarized under the name nonadditive probabilities, directly use one of these set functions, assuming the other one to be naturally defined such that P _ ( A c ) = 1 − P ¯ ( A ) {\displaystyle {\underline {P}}(A^{c})=1-{\overline {P}}(A)} , with A c {\displaystyle A^{c}} the complement of A {\displaystyle A} . Other related concepts understand the corresponding intervals [ P _ ( A ) , P ¯ ( A ) ] {\displaystyle [{\underline {P}}(A),{\overline {P}}(A)]} for all events as the basic entity.
History The idea to use imprecise probability has a long history. The first formal treatment dates back at least to the middle of the nineteenth century, by George Boole, who aimed to reconcile the theories of logic and probability. In the 1920s, in A Treatise on Probability, Keynes formulated and applied an explicit interval estimate approach to probability. Work on imprecise probability models proceeded fitfully throughout the 20th century, with important contributions by Bernard Koopman, C.A.B. Smith, I.J. Good, Arthur Dempster, Glenn Shafer, Peter M. Williams, Henry Kyburg, Isaac Levi, and Teddy Seidenfeld. At the start of the 1990s, the field started to gather some momentum, with the publication of Peter Walley's book Statistical Reasoning with Imprecise Probabilities (which is also where the term "imprecise probability" originates). The 1990s also saw important works by Kuznetsov, and by Weichselberger, who both use the term interval probability. Walley's theory extends the traditional subjective probability theory via buying and selling prices for gambles, whereas Weichselberger's approach generalizes Kolmogorov's axioms without imposing an interpretation. Standard consistency conditions relate upper and lower probability assignments to non-empty closed convex sets of probability distributions. Therefore, as a welcome by-product, the theory also provides a formal framework for models used in robust statistics and non-parametric statistics. Included are also concepts based on Choquet integration, and so-called two-monotone and totally monotone capacities, which have become very popular in artificial intelligence under the name (Dempster–Shafer) belief functions. Moreover, there is a strong connection to Shafer and Vovk's notion of game-theoretic probability.
… excerpt ends here. Continue reading the full article.
