dlm.allegra@btl.CSNET.UUCP (04/23/87)
May 7th 10:30 AM AT&T Bell Laboratories - Murray Hill 1E-449 REPORTING THE NON-MONOTONIC NEWS: Keeping the Beat Local Benjamin Grosof Stanford University "Non-monotonic" reasoning systems are ones in which some conclusions have a default or retractable status. A prime motivation for such systems is to build agents that revise their beliefs in response to news from their environment. Efficient updating is problematic, however, because adding new information in general may require the revision of many, or even all, previous retractable conclusions. An understanding is needed of the "partial monotonicities" of updating, i.e. of the irrelevance of updates to parts of the previous retractable conclusions. To define non-monotonic theories, we introduce a formalism based on McCarthy's circumscription that directly expresses, as axioms, both default beliefs and preferences among default beliefs. It has a strong semantics based on first- and second-order logic. We characterize non-monotonic theories as hierarchically decomposable in a manner more analogous to programming languages than to ordinary monotonic logics. We then give a set of results about partial monotonicities of updating. We discover some surprising differences between updates consisting of default axioms and those consisting of non-retractable axioms. These results bear on a wide variety of applications of non-monotonic reasoning. Sponsor: R.J.Brachman