1993 Abstract for Washington University Technical Report with K. Stiefvater COMPUTING SPECIFICITY This note reports on an effort to implement a version of Poole's rule for specificity. Relatively efficient implementation relies on correcting and improving a pruning lemma of Simari-Loui [92]. This in turn requires revision of Poole's specificity concept. The resulting system is a useable knowledge representation system with first-order-language and defeasible reasoning. Sample input and output are included in an appendix. It is a good candidate for multiple inheritance applications; it is useful for planning, but limited by the underlying search for plans. Introduction Much reasoning proceeds by argument. Automating this kind of reasoning requires theories about how to adjudicate disputes involving competing arguments. In recent work this has been described as determining a preference among arguments. A syntactic criterion for preference, especially one based on a concept of defeat among arguments, would be of practical use. After numerous attempts at identifying a suitable convention that determines when one non-demonstrative argument defeats another, two conclusions can be drawn: (1) there may be many suitable conventions (though this should not necessarily be described as a ``clash of intuitions''); and (2) in cases where the convention is not fixed (either by the field as a whole, or by a representer of knowledge, or by agreement among the conflicting parties to a dispute), superiority of one argument over another should be determined by meta-argument during the (object-level) dispute. The latter alternative has not yet been fully explored. Presumably it would require drawing analogies to benchmark disputes, or concocting a set of non-demonstrative reasons for claiming one argument is better than another. Actual arguments among researchers about which convention to adopt proceed by the former, but off-line (without reference to a particular dispute). The early investigation of Loui [87] does the latter. Though the field may lack the courage to forge a convention, and may be willing to retreat to on-line (mid-dispute) meta-argument, further study of promising conventions is appropriate. Even if parties to a dispute or a knowledge representer were willing simply to adopt an imperfect convention, a syntactic criterion for argument preference, there are still not many good candidates that could be adopted. There are a few suitable inheritance systems (e.g., Horty-Thomason-Touretzky [90a,90b], Nute [90], Stein [90], Guha [90], Padgham [89]), but the number of first-order systems is small (e.g., Simari-Loui [92], Geffner-Pearl [92], Pollock [92]). Note that we are restricting the discussion to systems that provide a syntactic criterion for preference; hence, they do not require the user to supply ordering information explicitly (e.g., Ginsberg [88], Baker-Ginsberg [89], Konolige [88], Lin-Shoham [89], Eklund [91], etc.). A major factor is computation. The systems proposed appear expensive to implement; the queries to theses systems appear expensive to compute. Most computations will call a theorem-prover as a subroutine, providing a bound on computation for negation as failure. Most call the theorem-prover a combinatorial number of times. We cannot speak of ultimate tractability for these systems, since the underlying logic is semi-decidable. However, since the theorem-proving is a constant computation in practice, it is reasonable to ask what can be done about the number of times this computation is invoked (see Kautz-Selman [88,91], Stillman [90] for complexity results regarding inheritance, the less expressive kind of system with syntactic preference criterion). Systems with implicit preference are not often implemented due to a lack of interest in this last question (as many authors have recently noted; see the AAAI Spring Symposium on Implemented Knowledge Representation Systems [91]). This is especially unfortunate since the choice of a convention is informed by experience with using the convention.