Relating Default Logic and Circumscription David W. Etherington1 Artificial Intelligence Principles Research Department AT&T Bell Laboratories 600 Mountain Avenue Murray Hill, NJ 07974 ether^allegra(a btl.csnet Abstract approaches are different on both the syntactic and Default logic and the various forms of cir• semantic levels. cumscription were developed to deal with similar Default logic and circumscription were developed problems. In this paper, we consider what is known at approximately the same time as attempts to formal• about the relationships between the two approaches ize nonmonotonic inference, which was recognized as and present several new results extending this an important facet of intelligent behaviour. In the knowledge. We show that there are interesting cases intervening years, there has been considerable activity in which the two formalisms do not correspond, as in this area. Both formalisms have been extensively well as cases where default logic subsumes cir• studied, and several new paradigms have been pro• cumscription. We also consider positive and negative posed. These two, however, have had remarkable results on translating between defaults and cir• staying power, and applications and refinements con• cumscription, and develop a context in which they can tinue to surface. Since both attempt to capture similar be evaluated. phenomena, the natural question is whether either subsumes the other. Is there a direct mapping by which default 1. Introduction theories subsume circumscription, or vice versa"! Circumscription [McCarthy 1980, 1986; Lifschitz There is, as yet, no definitive answer to this question. 1984] and default logic [Reiter 1980] are both formal• In what follows, we draw together results that begin to isms for reasoning in the absence of complete informa• provide answers. Because there are many criteria that tion. Often, the available knowledge about the state can be used to determine "subsumption", a plurality of of the world (problem domain, task) is incomplete, in "answers" is perhaps the best that can be provided. the sense that required details are unknown. Certain As we proceed we outline the assumptions that under• plausible assumptions can sometimes be made - to fill lie the partial answers that can be provided to date. in some of these missing details - that may further the We generally refer to Lifschitz' [1984] generaliza• goals of the reasoning system. These assumptions may tion of circumscription, which we simply call "cir• or may not prove correct when further information cumscription". This version allows the denotations of becomes available, for example through further obser• terms, as well as of predicates, to be varied. Where vation. The availability of additional information may thus lead to the retraction of certain conclusions. This appropriate in what follows, we draw attention to the property is called nonmonotonicity. effect of the presence or absence of variable terms. In what follows, we present very brief sketches of Both default logic and circumscription can be semantic underpinnings for the two formalisms. These used to enforce "policies" (defaults, preferences, ...) sketches are then contrasted and compared to help that (ideally) lead to highly plausible conjectures in determine some of the relationships between the tech• the absence of definite information. The actual world niques. is assumed to be among some restricted subset of those worlds that meet the criteria of what is known. This selection of "preferred" worlds may be externally jus• 2. Circumscription tified by convention, probabilities, etc, but these exter• nal justifications do not play a direct role in the One form of preference is the "Closed-World theory. Beyond these intuitive similarities, the two Assumption", or CWA. This is the assumption that all the facts about the world have been stated [Reiter 1978]. Those facts that do not follow from the given 1 Parts of this work were done at the University of facts can thus be assumed to be false. The assumption British Columbia, and supported in part by an I. W. of complete knowledge about the world can be prob• Killam Predoctoral Scholarship and NSERC grant A7642. lematic if applied indiscriminately, but it is sometimes Etherington 489 reasonable to assume complete information, especially consistent to believe certain "gating facts" or "justifi- about particular predicates. If such an assumption is cations^ (£). A default theory consists of a set of warranted, the world can be characterized by its being defaults and a set of first-order axioms. Ideally, the among those models having the smallest possible defaults induce (possibly several) stable, maximal, spe• extension for the predicates in question, given the con• cializations - called extensions- of the theory given by straints of what is known. Syntactically, this amounts the axioms. to assuming the negation of every atomic fact (over In spite of the apparent simplicity and naturalness the predicates for which complete information is of defaults, the semantics of default logic is more assumed) not entailed by the original theory. complex than that of circumscription. There are two In general, this idea can be extended to predicates reasons for this. First, defaults are by nature global: about which only incomplete information is available, each refers to everything that is believed or not so long as it is reasonable to assume that they cover as believed. This means that the semantics must be based few individuals as necessary. In such cases, the on sets of models (which we call world-descriptions), models satisfying this "generalized" CWA [Minker rather than individual models, to allow the concepts of 1982] need not agree on the extensions of the predi• belief and unbelief to be represented. Second, since cates concerned, only on the minimality (with respect the justification (p) and consequent (y) of a default to the original theory) of those extensions. need not coincide, defaults can predicate conjectures on the continuing lack of belief in certain propositions The idea of minimization can be generalized from without asserting or otherwise enforcing this lack of subset minimality of predicates' extensions to minimal• belief. Thus, although local comparisons between ity according some arbitrary pre-order on models. world-descriptions can be used to partially represent Circumscription is a syntactic device that characterizes the preference relation, information not inherent in what is true in a theory's minimal models.2 The the world-descriptions must be considered in the final theory is augmented by a second-order axiom that can determination of preference. The semantics must be obtained deterministically, given a particular somehow make this information available. choice of an ordering on models. This axiom represents the assumption that the world is among the Given these caveats, we can define a semantics theory's minimal models. The conjectures of interest for default logic analogous to the minimal model then follow from this extended theory by normal semantics of circumscription. The defaults define a deduction. preference ordering that, in turn, determines a set of A typical application would be to axiomatize the minimal world-descriptions. These must be further "normal" state of the world and then circumscriptively restricted to exclude those that violate the lack-of- minimize the set of abnormalities. conjecturing that belief constraints that implicitly gave them rise. Each things are as normal as possible [McCarthy 1986]. This of the remaining world-descriptions consists of all the models of some extension of the default theory, and approach allows circumscription to be used for some 3 forms of default reasoning, although the full details each extension can be so characterized. remain to be worked out. 4. Default Logic as Subsuming 3. Default Logic Circumscription Default logic sanctions conjectures based on infer• Since circumscription's proof-theory avoids the ence rules, called defaults. These have the form consistency checking (in general not semidecidable) associated with default logic proofs, it would be nice if P , and can be read "If a is believed, and p is not y circumscription subsumed default logic. This would disbelieved, then y can be assumed". Because of P's allow us to deal only with circumscriptive theories. role, defaults allow conjectures based on both what is Unfortunately, this is not the case. known and what is not known. Intuitively, the defaults provide criteria for preferring world- Theorem 1 descriptions over one another. Roughly speaking, each default. " , sanctions a conjecture (y) that Default logic can make conjectures that cannot specializes a world-description provided certain be obtained by generalized circumscription "prerequisites" (a) are believed, so long as it remains without variable terms. | 2 For simplicity's sake, we ignore the issue of incom• pleteness. See [Perlis & Minker 1986] or [Etherington 3 This semantics is developed in detail elsewhere [Eth• 1987a] for a discussion. erington 1987 a,b]. 490 KNOWLEDGE REPRESENTATION Etherlngton 491 The reasons that only prerequisite-free defaults (provability) vs local (consistency) comparisons in the have modular translations become clear if we contrast model-theory (proof-theory), and statements about the semantics we have sketched for default logic and "unnamed" individuals. In all but the last of these the minimal-model semantics of circumscription. The categories, default logic is stronger.