Repository | Book | Chapter

Recent advances in Artificial Intelligence (AI) emphasize the crucial importance of nonmonotonic reasoning in any adequate scheme of knowledge representation.¹ As Nute (1984) observes, humans notoriously rely upon nonmonotonic reasoning, and any "automated reasoning system should also reason nonmonotonically in a way which people can easily understand". Conclusions that can be inferred on the basis of a given set of premises may often be withdrawn or even overruled when new evidence is provided in the form of additional premises. Told only that a match has been struck, for example, humans will typically infer that that match did light and burn. But if they are also told that the match in question was wet when struck, they will tend to revise their inference and conclude that the match did not light. If told further that the match was wet but coated in paraffin when struck, most (reasonable) humans would infer that the match did light and burn after all, thus revising their inference once more — unless, of course, told no oxygen was present when the match was struck, in which case still another revision is called for. And this is precisely the character of nonmonotonic reasoning, as Nute suggests, i.e., "people draw conclusions based on incomplete information, but these conclusions are open to revision as better information becomes available".²

DOI: 10.1007/978-94-009-2699-8_10

Full citation: