This talk is part of the "Cognitive Robotics Group/KR Group" seminar series. We'll be covering default reasoning later in the term. As such, you may not understand all of this talk nor necessarily appreciate the significance of it, but you may find it interesting all the same. If you would like to be on the mailing list for these seminars, please email jabaier -AT- cs - DOT- toronto -DOT edu. This is Jorge Baier, who not only TAs our course, but is also the organizer of these meetings. ------------------------------------------------------------------------- Date: Monday, Oct. 2nd Time: 3-4pm Room: PT266 ------------------------------------------------------------------------- The Truth about Defaults Hector Levesque Abstract Virtually all of the work on defaults in AI has concentrated on default reasoning: given a theory T containing facts and defaults of some sort, we study how an ideal agent should reason with T, typically in terms of constructs like fixpoints, or partial orders, or nonmonotonic entailment relationships. In this talk, we investigate a different question: under what conditions should we consider the theory T to be true, or believed to be true, or all that is believed to be true? By building on truth in this way, we end up with a logic of defaults that is classical, that is, a logic with an ordinary monotonic notion of entailment. And yet default reasoning emerges naturally from these ideas. We will show how to characterize the default logic of Reiter and the autoepistemic logic of Moore in purely truth-theoretic terms. We will see that the variant proposed by Konolige is in fact a link between the two, and that all three fit comfortably within a single logical language, that we call O3L. Finally, we will present first steps towards a proof theory (with axioms and rules of inference) for O3L. Among other things, this allows us to present ordinary sentence-by-sentence derivations that correspond to different sorts of default reasoning. This is joint work with Gerhard Lakemeyer.