Entailment in Probability of Thresholded Generalizations
This work addresses foundational issues in AI and logic for ensuring reliable probabilistic reasoning, though it appears incremental by building on existing systems like System-Z^+.
The paper tackles the problem of defining a nonmonotonic logic for thresholded generalizations to ensure probabilistic trustworthiness in entailment, showing that conclusions are improbable to be false given true premises and relating this to System-Z^+.
A nonmonotonic logic of thresholded generalizations is presented. Given propositions A and B from a language L and a positive integer k, the thresholded generalization A=>B{k} means that the conditional probability P(B|A) falls short of one by no more than c*d^k. A two-level probability structure is defined. At the lower level, a model is defined to be a probability function on L. At the upper level, there is a probability distribution over models. A definition is given of what it means for a collection of thresholded generalizations to entail another thresholded generalization. This nonmonotonic entailment relation, called "entailment in probability", has the feature that its conclusions are "probabilistically trustworthy" meaning that, given true premises, it is improbable that an entailed conclusion would be false. A procedure is presented for ascertaining whether any given collection of premises entails any given conclusion. It is shown that entailment in probability is closely related to Goldszmidt and Pearl's System-Z^+, thereby demonstrating that the conclusions of System-Z^+ are probabilistically trustworthy.