Knowledge representation and defeasible reasoning (Q1188702)

[The article of this volume will not be reviewed individually.] This volume provides a collection of 16 studies that focus on four principal areas of knowledge representation research in artificial intelligence and related fields in linguistics, philosophy, and decision theory: defeasible reasoning and the frame problem, representation problems and ordinary language, belief revision and its rules of inference, and logical problems in representing knowledge and coping with uncertainty. The first part of the book contains three papers that focus on various aspects of the frame problem. The first of these (by \textit{Nute}) presents a particular non-monotonic formalism, Defeasible Logic, that properly represents the Yale Shooting Problem and other familiar examples which often cause problems for non-monotonic systems. Yet another formal proposal for monotonically solving the frame problem is provided in terms of the Situation Calculus (by \textit{Schubert}) which entirely avoids (traditional) frame axioms and introduces axioms about what actions are required to produce given types of changes (so-called explanation-closure axioms) instead. Furthermore, a default strategy which is essentially the `sleeping dog' strategy of STRIPS is shown to be deductively sound when appropriately based on explanation closure; the potential of this approach is explored with respect to external events, continuous change, action composition using sequencing, conditionals and iteration, and, most of all, concurrency. In the third of these papers (by \textit{Dunn}), basically a brief and pointed rejoinder to Fodor who argues that the frame problem is unsolved because the relevance problem is unsolved, debates on a proper characterization of the frame problem are put on a firm formal basis introducing the notion of a computationally relevant fact (or relevant predication), the formalization of which is achieved in the context of the Anderson-Belnap relevance logics and currently actively explored in several database applications. This part of the book also contains a presentation of a system of defeasible inference (closely related to the logic of conditionals developed by Adams) that combines the probabilistic and the logical approach to non-monotonic reasoning (\textit{Geffner} and \textit{Pearl}); it operates very much like natural deduction systems in logic and, yet, can be justified on probabilistic grounds. The main contribution of the proposed framework for defeasible inference is the emergence of a more precise proof-theoretic and semantic account of defaults, and the association of defaults with a formal notion of irrelevance as a set of sufficient conditions under which belief in the consequent of a given default can be preserved upon acquiring new information (thus this notion plays a role similar to standard frame axioms). The second part on representation problems and ordinary language covers diverse topics, such as self-reference as an important attribute of intelligent systems and its relation to non-monotonicity, reification, and intentionality (\textit{Perlis}); a linguistic study (by \textit{Gillon}) employing standard truth-conditional semantic analysis considers ``bare plurals'' in English, i.e. noun phrases without determiners, as plural indefinite noun phrases, thus arguing against an alternative analysis of English bare plural by Carlson. Two of the papers are concerned with pragmatic issues of language use. Trying to advance the meta-level of linguistic descriptions, \textit{Belnap} and \textit{Perloff} propose a special analytic device, the ``stit'' sentence, whose canonical form helps clarify the role of English sentences as being agentives, the agentive content of imperatives, and the troubles usually encountered in the analysis of ``refraining'' sentences (acting vs. non-acting as proper reading). The final contribution of this section (by \textit{McCafferty}) argues that indirect speech acts derive from the speaker plans for the linguistic context as well as from the knowledge of `domains plans' which is real-world script-like knowledge that encodes the customary ways of achieving some end. This approach to discourse analysis adopts a contextual change theory of meaning, Allen's concept of a speaker's plan and applies this analytic framework in the context of longer conversations. The third section on belief revision and its inference rules presents three approaches dealing with probabilistic models and one working with a logical framework. This is a paper (by \textit{Cross}) on rational inquiry and the logic of rational belief change (or theory change). It argues against Gärdenfors's theorem that one principle about belief revision, the Ramsay rule, should be given up in favor of the Preservation principle. In particular, it provides formal evidence that, consistent with the Ramsay rule, rational agents who modify their beliefs in response to new information cannot avoid reasoning non-monotonically and so any adequate theory of belief revision must take non-monotonic reasoning into account. The series of papers dealing with probabilistic models of belief revision starts with \textit{Pearl}'s discussion of Jeffrey's Rule of belief updating, the information requirements it poses, and gives a brief survey of graphoid theory (as a formalism to express relevance) in order to illustrate the kind of structural knowledge that should supplement the algebraic description of belief states (and thus make Jeffrey's Rule actually workable in any specific case). This part goes on with a discussion (by \textit{Seidenfeld}) on questions of consensus for inferences and decisions within the context of Bayesian theory. It considers, in particular, the problem of extending the norms of rationality from individuals to groups, as illustrated by the case of combining uncertainty and utility judgments of different experts. This section ends with the derivation and proof of three triviality theorems considering conditionals and conditional probabilities (\textit{Leblanc} and \textit{Roeper}). The final part of the book treats logical problems in representing knowledge, with emphasis on statistical models of inferencing and decision making in the face for uncertainty. It begins with a presentation (by \textit{Carpenter} and \textit{Thomason}) which stresses the observation that semantic networks and their relation to formal logic is far more complex than generally supposed. Starting from the consideration of monotonic inheritance networks, possible definitions of non-monotonic inheritance procedures are discussed distinguishing between reasoning mechanisms which are credulous and those which are skeptical in the face of conflicting evidence. An integration of methods used in decision theory and artificial intelligence work on planning to arrive at the defeasible specification of utilities of states in non-trivial planning tasks is proposed in the second paper (by \textit{Loui}). Based on considerations that classical logic is inadequate for coping with typical natural language statements that are fuzzy sentences and sentences expressing degrees of belief, a formal logic of assertions is developed which is capable of handling these phenomena properly (\textit{Giles}). The final paper (by \textit{Aleliunas}) proposes practical and logically correct theories of rational belief (probabilistic logic) that are more general than the standard probability theory, yet remain capable of many of the inferences of this latter theory. Finally, a discussion of recent attempts to devise new theories of rational belief reveals that they are either incorrect or reincarnations of already existing ones.