Published online by Cambridge University Press: 12 March 2014
Research in AI has recently begun to address the problems of nondeductive reasoning, i.e., the problems that arise when, on the basis of approximate or incomplete evidence, we form well-reasoned but possibly false judgments. Attempts to stimulate such reasoning fall into two main categories: the numerical approach is based on probabilities and the nonnumerical one tries to reconstruct nondeductive reasoning as a special type of deductive process. In this paper, we are concerned with the latter usually known as nonmonotonic deduction, because the set of theorems does not increase monotonically with the set of axioms.
It is generally acknowledged that nonmonotonic (n.m.) formalisms (e.g., [C], [MC1], [MC2] [MD], [MD-D], [Rl], [R2], [S]) are plagued by a number of difficulties. A key issue concerns the fact that most systems do not produce an axiomatizable set of validities. Thus, the chief objective of this paper is to develop an alternative approach in which the set of n.m. inferences, which somehow qualify as being deductively sound, is r.e.
The basic idea here is to reproduce the situation in First Order Logic where the metalogical concept of deduction translates into the logical notion of material implication. Since n.m. deductions are no longer truth preserving, our way to deal with a change in the metaconcept is to extend the standard logic apparatus so that it can reflect the new metaconcept. In other words, the intent is to study a concept of nonmonotonic implication that goes hand in hand with a notion of n.m. deduction. And in our case, it is convenient that the former be characterized within the more tractable context of monotonic logic.