Protected by Copyscape Duplicate Content Checker



Maximum limit of CO2 350ppm...!

CO2 World Ocean Acidification...!

World CO2
since 1750 (cubic feet) - CO2 Global Warming

UNIVERSAL STANDARD TIME - UTC - GMT [Greenwich-UK ! ] - Date Line Pacific FIJI

7 TIMES - do they do for you ? - 7 HORAS = LE SIRVE ALGUNO ?



PoodWaddle World Population and Productive Land Count !

YOUR CUTE unforgettable games online

YOUR CUTE unforgettable games online

Other web addresses

Other Web sites:

JOINTQUEST.COM ; ; : Pure Logic Book : George Frederick Thomson Broadhead

Blog of Venezuela - A country gone to another reality ! AND eve-nt !

DO NOT MISS the "Modern" Efforts of a ancient violent dead God and Jesus !

Friday, August 5, 2011

...THE progress of Artificial Intelligence and LOGIC...!

Progress in Logic-Based AI

YEAR 2000.

Progress in Logic-Based AI

Hubert Dreyfus claims that ``symbolic AI'' is a ``degenerating research program'', i.e. is not making progress. It's hard to see how he would know, since he makes no claim to have read much of the recent literature.

In defending ``symbolic AI'', I shall concentrate on just one part of symbolic AI--the logic-based approach. It was first proposed in [McCarthy, 1959], attracted only intermittent following at first, but has had an increasing number of workers since 1970. gif I think other approaches to AI will also eventually succeed, perhaps even connectionism. To contradict an earlier Dreyfus metaphor ``AI at the Crossroads Again'', it isn't a crossroads but a race including logic based AI, SOAR, connectionism and several other contenders.

How goes the logic-based runner? In fits and starts, as problems are identified and progress made.

Within logic-based AI, I shall emphasize one development--formalized nonmonotonic reasoning, because it illustrates intermittent but definite progress. It was first proposed in [McCarthy, 1977], gained momentum with the 1980 special issue of Artificial Intelligence, and summarized in the collection [Ginsberg, 1987]. It has continued to develop, see e.g. [Lifschitz, 1993].

Minsky [Minsky, 1975] mentioned the need for something like nonmonotonic reasoning, but used this fact as evidence for the inadequacy of logic-based approaches to AI and the need for approaches not based on logic. This isn't how things have gone. Nonmonotonic reasoning has developed as a branch of mathematical logic, using the concepts, such as that of interpretation and model, that logicians have been developing since the 1930s.

The circumscription method of nonmonotonic reasoning would have been entirely comprehensible to Hilbert and probably even to Frege. However, no formalized nonmonotonic logic was developed until the 1970s. This is typical of the slow progress in developing mathematical logical formalisms. Just consider the sequence, Aristotle, Leibniz, Boole, Frege, Hilbert, Gödel. Each step would have been comprehensible to the predecessor, yet it took a long time for the new ideas to appear. Formalized nonmonotonic reasoning is surely not the last step in the chain aimed at formalizing useful reasoning. Formalizing nonmonotonic reasoning required realizing that there is proper reasoning that is not conclusive and that is often not the same as probabilistic reasoning.

For this reason, neither Dreyfus or anyone else is entitled to conclude that if the logic oriented AI problem hasn't been solved in 40 years, it won't ever be solved. To do that it would be necessary to prove a theorem about limitations of logic. Not even showing that there has been no progress at all would be conclusive. However, Dreyfus makes no reference to nonmonotonic reasoning in his book. That's about 1,000 papers he doesn't know about. However, in answer to a question at a book-selling talk, he said that claiming progress in nonmonotonic reasoning is progress towards AI is like claiming that climbing a tree is progress towards reaching the moon--thus recycling a metaphor from the book.

Although formalized nonmonotonic reasoning was discovered in connection with AI, many logicians pursue it as a purely mathematical study, independent of applications to AI or logic programming (another non-entry in Dreyfus's index).

The Future of Logic Based AI

The review editors asked me to say what I think the obstacles are to human-level AI by the logic route and why I think they can be overcome. If anyone could make a complete list of the obstacles, this would be a major step towards overcoming them. What I can actually do is much more tentative.

Workers in logic-based AI hope to reach human-level in a logic based system. Such a system would, as proposed in [McCarthy, 1959], represent what it knew about the world in general, about the particular situation and about its goals by sentences in logic. Other data structures, e.g. for representing pictures, would be present together with programs for creating them, manipulating them and for getting sentences for describing them. The program would perform actions that it inferred were appropriate for achieving its goals.

Logic-based AI is the most ambitious approach to AI, because it proposes to understand the common sense world well enough to express what is required for successful action in formulas. Other approaches to AI do not require this. Anything based on neural nets, for example, hopes that a net can be made to learn human-level capability without the people who design the original net knowing much about the world in which their creation learns. Maybe this will work, but then they may have an intelligent machine and still not understand how it works. This prospect seems to appeal to some people.

Common sense knowledge and reasoning is at the core of AI, because a human or an intelligent machine always starts from a situation in which the information available to it has a common sense character. Mathematical models of the traditional kind are imbedded in common sense. This was not obvious, and many scientists supposed that the development of mathematical theories would obviate the need for common sense terminology in scientific work. Here are two quotations that express this attitude.

One service mathematics has rendered to the human race. It has put common sense back where it belongs, on the topmost shelf next to the dusty canister labelled `discarded nonsense'. --E. T. Bell

All philosophers, of every school, imagine that causation is one of the fundamental axioms or postulates of science, yet, oddly enough, in advanced sciences such as gravitational astronomy, the word `cause' never occurs ... The law of causality, I believe, like much that passes muster among philosophers, is a relic of a bygone age, surviving, like the monarchy, only because it is erroneously supposed to do no harm tex2html_wrap_inline136 .--B. Russell, ``On the Notion of Cause'', Proceedings of the Aristotelian Society, 13 (1913), pp. 1-26.

The ``Nemesis'' theory of the mass extinctions holds that our sun has a companion star that every 13 million years comes close enough to disrupt the Oort cloud of comets, some of which then come into the inner solar system and bombard the earth causing extinctions. The Nemesis theory involves gravitational astronomy, but it doesn't propose a precise orbit for the star Nemesis and still less proposes orbits for the comets in the Oort cloud. Therefore, the theory is formulated in terms of the common sense notion of causality.

It was natural for Russell and Bell to be pleased that mathematical laws were available for certain phenomena that had previously been treated only informally. However, they were interested in a hypothetical information situation in which the scientist has a full knowledge of an initial configuration, e.g. in celestial mechanics, and needs to predict the future. It is only when people began to work on AI that it became clear that general intelligence requires machines that can handle the common sense information situation in which concepts like ``causes'' are appropriate. Even after that it took 20 years before it was apparent that nonmonotonic reasoning could be and had to be formalized.

Making a logic-based human-level program requires enough progress on at least the following problems:

extensions of mathematical logic
Besides nonmonotonic reasoning other problems in the logic of AI are beginning to take a definite form including formalization of contexts as objects. This can provide a logical way of matching the human ability to use language in different ways depending on context. [McCarthy, 1987], [Guha, 1991], [McCarthy, 1993], [Buvac and Mason, 1993].

elaboration tolerance
Formalisms need to be elaboratable without a human having to start the formalism over from the beginning. There are ideas but no articles as yet.

concurrent events
[Gelfond, Lifschitz and Rabinov, 1991] treats this using the situation calculus, and I have some recent and still unpublished results aimed at a simpler treatment.

The treatment of mental objects such as beliefs (much discussed in the philsophical literature) and the corresponding term concept, e.g. ``what he thinks electrons are'' (hardly at all discussed in the formal literature).

We need a better understanding of what are to be considered objects in the logic. For example, a full treatment of the missionaries-and-cannibals problem together with reasonable elaborations must allow us to say, ``There are just two things wrong with the boat.''

introspection and transcendence
Human intelligence has the ability to survey, in some sense, the whole of its activity, and to consider any assumption, however built-in, as subject to question. Humans aren't really very good at this, and it is only needed for some very high level problems. Nevertheless, we want it, and there are some ideas about how to get it. What may work is to use the context mechanism as discussed in [McCarthy, 1993] to go beyond the outermost context considered so far.

Unfortunately, too many people concentrated on self-referential sentences. It's a cute subject, but not relevant to human introspection or to the kinds of introspection we will have to make computers do.

levels of description
If one is asked how an event occurred, one can often answer by giving a sequence of lower level events that answer the question for the particular occurrence. Once I bought some stamps by going to the stamp selling machine in the airport and putting in six dollars, etc. Each of these subevents has a how, but I didn't plan them, and cannot recall them. A stamp buying coach would have analyzed them to a lower level than I could and would be able to teach me how to buy stamps more effectively. For AI we therefore need a more flexible notion than the computer science theories of how programs are built up from elementary operations.

Dreyfus asks why anyone should believe all this can be done. It seems as good a bet as any other difficult scientific problem. Recently progress has become more rapid, and many people have entered the field of logical AI in the last 15 years. Besides those whose papers I referenced, these include Raymond Reiter, Leora Morgenstern, Donald Perlis, Ernest Davis, Murray Shanahan, David Etherington, Yoav Shoham, Fangzhen Lin, Sarit Kraus, Matthew Ginsberg, Douglas Lenat, R. V. Guha, Hector Levesque, Jack Minker, Tom Costello, Erik Sandewall, Kurt Konolige and many others. There aren't just a few ``die-hards''.

However, reaching human level AI is not a problem that is within engineering range of solution. Very likely, fundamental scientific discoveries are still to come.


Common Sense in Lenat's Work

Douglas Lenat is one of the few workers in AI at whose recent work Dreyfus has taken a peek.

Dreyfus, p. xvii and xviii, writes:

When, instead of developing philosophical theories of the transcendental conditions that must hold if the mind is to represent the world, or proposing psychological models of how the storage and retrieval of propositional representations works, researchers in AI actually tried to formulate and organize everyday consensus knowledge, they ran into what has come to be called the commonsense-knowledge problem. There are really at least three problems grouped under this rubric:

How everyday knowledge must be organized so that one can make inferences from it.
How skills or know-how can be represented as knowing-that.
How relevant knowledge can be brought to bear in particular situations.

While representationalists have written programs that attempt to deal with each of these problems, there is no generally accepted solution, nor is there a proof that these problems cannot be solved. What is clear is that all attempts to solve them have run into unexpected difficulties, and this in turn suggests that there may well be in-principle limitations on representationalism. At the very least these difficulties lead us to question why anyone would expect the representationalist project to succeed.

That's not too bad a summary except for the rhetorical question at the end. Why should one expect it to be easy, and why should one expect it not to succeed eventually in reaching human level intelligence? Most of the people who have pursued the approach have seen enough of what they regard as progress to expect eventual success. I have referred to some of this progress in my account of the invention and development of formalized nonmonotonic reasoning.

Mostly I agree with what Lenat said (as Dreyfus quotes him in the book), and I don't find much support for Dreyfus's assertions that empathy rather than just verbalizable understanding is required in order to understand human action. I think the example on p. xix of what ``it'' means in

Mary saw a dog in the window. She wanted it.

is within the capability of some current parsers that use semantic and pragmatic information.

However, I think the following assertion of Lenat's [Lenat and Guha, 1990] quoted by Dreyfus on p. xxv is an oversimplification.

These layers of analogy and metaphor eventually `bottom out' at physical-somatic-primitives: up, down, forward, back, pain, ssssscold, inside, seeing, sleeping, tasting, growing, containing, moving, making noise, hearing, birth, death, strain, exhaustion, tex2html_wrap_inline136

The contact of humans (and future robots) with the common sense world is on many levels, and our concepts are on many levels. Events that might bottom out physically--as informing someone of something may physically bottom out in making a noise--often don't bottom out epistemologically. We may assert that A informed B of something without our being able to describe the act in terms of making noise or typing on a keyboard.

While I don't agree with Lenat's formulation, the success of Cyc doesn't depend on its correctness. Cyc perfectly well can (and indeed does) store information obtained on several levels of organization and used by programs interacting with the world on several levels.

All this doesn't guarantee that Cyc will succeed as a database of common sense knowledge. There may be to big a conceptual gap in the AI community's ideas of what are the usefully stored elements of common sense knowledge.


The Degenerating Research Program

In the first edition of Dreyfus's book there were some challenges to AI. Dreyfus said computers couldn't exhibit ``ambiguity tolerance'', ``fringe consciousness'' and ``zeroing in''. These were left so imprecise that most readers couldn't see any definite problem at all. In the succeeding 30 years Dreyfus has neither made these challenges more precise nor proposed any new challenges, however imprecise. It's a pity, because AI could use a critic saying, ``Here's the easiest thing I don't see how you can do''. That part of Dreyfus's research program has certainly degenerated.

However, I can give a definite meaning to the phrase ``ambiguity tolerance'' that may not be too far from Dreyfus's vague idea, and with which formalized nonmonotonic reasoning can deal. The idea is that a concept that may be ambiguous in general is to be taken by default as unambiguous in a particular case unless there is reason to do otherwise.

Here's an example.

Suppose that some knowledge engineer has the job of making an adviser for assistant district attorneys. The prosecutor answers some questions about the facts of the case, and the program suggests asking for indictments for certain crimes. We suppose that attempting to bribe a public official is one of these crimes.

We ask whether the knowledge engineer must have anticipated the following three possible defenses against the charge, i.e. have decided whether the following circumstances still justify an indictment.

The defendant's lawyer claims that his client did not know the person he offered money to fix his drunk driving convictions was the commisioner of motor vehicles. His client thought he was just an influential lawyer.
The defendant's lawyer claims that while his client may have thought he was bribing the commissioner of motor vehicles, he really wasn't, because the Governor had never properly signed the commission.
The defendant put an advertisement in the Criminal Gazette offering $5,000 to any public official who would fix his conviction. Must the prosecution exhibit a specific public official the defendant was attempting to bribe in order to get a conviction for ``attempting to bribe a public official''.

There may be further potential ambiguities in the statute. If we demand that the knowledge engineer have resolved all of them before he can write his expert system, we are asking for the impossible. Legislators, lawyers and judges don't see all the ambiguities in advance.

Notice that in most cases of bribing a public official, there was a specific individual, and he really was a public official and this was really known to the defendant. Very likely, the legislators had not thought of any other possibilities. The nonmonotonic reasoning approach to ambiguity tolerance says that by default the statute is unambiguous in a particular case. Indeed this is how the law works. The courts will not invalidate a law because of a general amiguity; it has to be ambiguous in a significant way in the particular case.

Since the expert system writer cannot anticipate all the possible ambiguities, he must make his system ambiguity tolerant.

When an ambiguity is actually pointed out to the expert system, it would be best if it advised looking at cases to see which way the statute had been interpreted by judges. I don't know whether to be a useful adviser in statutory criminal law, the expert system would have to have a library of cases and the ability to reason from them.

I have not written logical formulas for ambiguity tolerance, i.e. expressing the default that a concept, possibly ambiguous in general, is to be considered unambiguous in particular cases unless there is evidence to the contrary. However, I would be strongly motivated to give it high priority if Dreyfus were to offer to bet money that I can't.

To conclude: Dreyfus has posed various challenges to AI from time to time, but he doesn't seem to make any of them precise. Here is my challenge to Dreyfus, whereby he might rescue his research program from degeneration.

What is the least complex intellectual behavior that you think humans can do and computers can't? It would be nice to have more details than were given in connection with ``ambiguity tolerance'' and ``zeroing in''.

Author & Citation Info | Friends PDF Preview | InPho Search | PhilPapers Bibliography
Logic and Artificial Intelligence
First published Wed Aug 27, 2003; substantive revision Fri May 9, 2008

Artificial Intelligence (which I'll refer to hereafter by its nickname, “AI”) is the subfield of Computer Science devoted to developing programs that enable computers to display behavior that can (broadly) be characterized as intelligent.[1] Most research in AI is devoted to fairly narrow applications, such as planning or speech-to-speech translation in limited, well defined task domains. But substantial interest remains in the long-range goal of building generally intelligent, autonomous agents.[2]

Throughout its relatively short history, AI has been heavily influenced by logical ideas. AI has drawn on many research methodologies; the value and relative importance of logical formalisms is questioned by some leading practitioners, and has been debated in the literature from time to time.[3] But most members of the AI community would agree that logic has an important role to play in at least some central areas of AI research, and an influential minority considers logic to be the most important factor in developing strategic, fundamental advances.

The relations between AI and philosophical logic are part of a larger story. It is hard to find a major philosophical theme that doesn't become entangled with issues having to do with reasoning. Implicatures, for instance, have to correspond to inferences that can be carried out by a rational interpreter of discourse. Whatever causality is, causal relations should be inferrable in everyday common sense settings. Whatever belief is, it should be possible for rational agents to make plausible inferences about the beliefs of other agents. The goals and standing constraints that inform a rational agent's behavior must permit the formation of reasonable plans.

In each of these cases, compatibility with an acceptable account of the relevant reasoning is essential for a successful philosophical theory. But the methods in the contemporary philosophical inventory are too crude to provide anything like an adequate account of reasoning that is this complex.

Bringing an eclectic set of conceptual tools to the problem of idealized reasoning in realistic settings, and using computers to model and test the theories, research in AI has transformed the study of reasoning—especially of practical, common sense reasoning. This process and its outcome is well documented in Russell & Norvig 2003.

The new insights and theories that have emerged from AI are, I believe, of great potential value in informing and constraining many areas of philosophical inquiry. The case of philosophical logic that forms the theme of this article may provide support for the more general point. Although logic in AI grew out of philosophical logic, in its new setting it has produced new theories and ambitious programs that would not have been possible outside of a community devoted to building full-scale computational models of rational agency.

I imagine that the audience for this entry will consist primarily of philosophers who have little or no familiarity with AI. In writing this survey, I have tried to concentrate on the issues that arise when logic is used in understanding problems in intelligent reasoning and guiding the design of mechanized reasoning systems. Logic in AI is a large and rapidly growing field—I haven't tried to achieve anything like complete coverage. In Section 3 and Section 4 I have tried to provide an overview with some historical and technical details concerning nonmonotonic logic and reasoning about action and change, a topic that is not only central in AI but that should be of considerable interest to philosophers. The remaining sections provide brief and inadequate sketches of selected topics, with references to the primary literature.

For a recent, comprehensive, up-to-date collection of survey papers and original contributions to the field of logic-based AI, with extensive references to the literature, see Minker 2000b. Minker 2000a is a useful introduction to the book and to the field. I highly recommend this volume as a beginning point for readers who wish to pursue this topic further.

1. Logic and Artificial Intelligence
2. John McCarthy and Common Sense Logicism
3. Nonmonotonic Reasoning and Nonmonotonic Logics
4. Reasoning about Action and Change
5. Causal Reasoning
6. Spatial Reasoning
7. Reasoning about Knowledge
8. Towards a Formalization of Common Sense
9. Logical Approaches to Natural Language and Communication
10. Taxonomic Representation and Reasoning
11. Contextual Reasoning
12. Prospects for a Logical Theory of Practical Reason
Other Internet Resources
Related Entries

1. Logic and Artificial Intelligence
1.1 The Role of Logic in Artificial Intelligence

Theoretical computer science developed out of logic, the theory of computation (if this is to be considered a different subject from logic), and some related areas of mathematics.[4] So theoretically minded computer scientists are well informed about logic even when they aren't logicians. Computer scientists in general are familiar with the idea that logic provides techniques for analyzing the inferential properties of languages, and with the distinction between a high-level logical analysis of a reasoning problem and its implementations. Logic, for instance, can provide a specification for a programming language by characterizing a mapping from programs to the computations that they license. A compiler that implements the language can be incomplete, or even unsound, as long as in some sense it approximates the logical specification. This makes it possible for the involvement of logic in AI applications to vary from relatively weak uses in which the logic informs the implementation process with analytic insights, to strong uses in which the implementation algorithm can be shown to be sound and complete. In some cases, a working system is inspired by ideas from logic, but acquires features that at first seem logically problematic but can later be explained by developing new ideas in logical theory. This sort of thing has happened, for instance, in logic programming.

In particular, logical theories in AI are independent from implementations. They can be used to provide insights into the reasoning problem without directly informing the implementation. Direct implementations of ideas from logic—theorem-proving and model-construction techniques—are used in AI, but the AI theorists who rely on logic to model their problem areas are free to use other implementation techniques as well. Thus, in Moore 1995b (Chapter 1), Robert C. Moore distinguishes three uses of logic in AI; as a tool of analysis, as a basis for knowledge representation, and as a programming language.

A large part of the effort of developing limited-objective reasoning systems goes into the management of large, complex bodies of declarative information. It is generally recognized in AI that it is important to treat the representation of this information, and the reasoning that goes along with it, as a separate task, with its own research problems.

The evolution of expert systems illustrates the point. The earliest expert systems, such as MYCIN (a program that reasons about bacterial infections, see Buchanan & Shortliffe 1984), were based entirely on large systems of procedural rules, with no separate representation of the background knowledge—for instance, the taxonomy of the infectious organisms about which the system reasoned was not represented.

Later generation expert systems show a greater modularity in their design. A separate knowledge representation component is useful for software engineering purposes—it is much better to have a single representation of a general fact that can have many different uses, since this makes the system easier to develop and to modify. And this design turns out to be essential in enabling these systems to deliver explanations as well as mere conclusions.[5]
1.2 Knowledge Representation

In response to the need to design this declarative component, a subfield of AI known as knowledge representation emerged during the 1980s. Knowledge representation deals primarily with the representational and reasoning challenges of this separate component. The best place to get a feel for this subject is the proceedings of the meetings that are now held every other year: see Brachman et al. 1989, Allen et al. 1991, Nebel et al. 1992, Doyle et al. 1994, Aiello et al. 1996, Cohn et al. 1998, Cohn et al. 2000, and Fensel et al. 2002.

Typical articles in the proceedings of the KR and Reasoning conferences deal with the following topics.

Topics in logical theory and the theory of computation, including
Nonmonotonic logic
Complexity theory
Studies in application areas, including
Temporal reasoning
Formalisms for reasoning about planning, action and change
Reasoning about context
Reasoning about values and desires
Reasoning about the mental states of other agents, and especially about knowledge and belief
Spatial reasoning
Reasoning about vagueness
Studies in application techniques, including
Logic programming
Description logics
Theorem proving
Model construction
Studies of large-scale applications, including
Cognitive robotics
Merging, updating, and correcting knowledge bases

These topics hardly overlap at all with the contents of the Journal of Symbolic Logic, the principal research archive for mathematical logic. But there is substantial overlap in theoretical emphasis with The Journal of Philosophical Logic, where topics such as tense logic, epistemic logic, logical approaches to practical reasoning, belief change, and vagueness account for a large percentage of the contributions. Very few JPL publications, however, deal with complexity theory or with potential applications to automated reasoning.
1.3 Philosophical Logic

I do not know of a good history of philosophical logic. In fact, the distinction between mathematical and philosophical logic may well be incidental in relation to the overall goals of the subject, since technical rigor and the use of mathematical methods seem to be essential in all areas of logical research. However, the distinction between the two subfields has been magnified by differences in the sorts of professional training that are available to logicians, and by the views of individuals on what is important for the field.

The statement of policy presented in Volume 1, no. 1 of the Journal of Symbolic Logic (1936) lists bringing together the mathematicians and philosophers working in logic among the goals of the new journal. Probably at this time both the mathematicians and the philosophers shared a sense that their subject was considered to be somewhat marginal by their colleagues, and may have felt a primary loyalty to logic as a subject rather than to any academic discipline. Articles in the first volume of the JSL were divided about equally between professional mathematicians and philosophers, and the early volumes of the JSL do not show any strong differences between the two groups as to topic.

This situation changed in the 1960s. The 1969 volume of the JSL contained 39 articles by mathematicians, and only nine by philosophers. By the early 1970s, many philosophers felt that philosophical papers on logic were unlikely to be accepted by the JSL, and that if they were accepted they were unlikely to be read by philosophers. At this point, the goals of the two groups had diverged considerably. Mathematicians were pursuing the development of an increasingly technical and complex body of methods and theorems. Many philosophers felt that this pursuit was increasingly irrelevant to the goal of illuminating philosophical issues. These divisions led to the founding of the Journal of Philosophical Logic in 1972. The list of sample topics in the first issue included:

Contributions to branches of logical theory directly related to philosophical concerns, such as inductive logic, modal logic, deontic logic, quantum logic, tense logic, free logic, logic of questions, logic of commands, logic of preference, logic of conditionals, many-valued logic, relevance logics;
Contributions to philosophical discussions that utilize the machinery of formal logic …;
Discussions of philosophical issues relating to logic and the logical structure of language, …;
Philosophical work relating to the special sciences, ….

Most of the articles over the subsequent 28 years of the JPL belong to the first of these four categories. But the description with which this list begins is not particularly illuminating: why should these particular topics be of interest to philosophers? I believe that the most important feature they share is a sense that despite successes in formalizing areas of mathematical logic, the scope of logic remained severely limited. There are unsolved problems in formalizing the nonmathematical sciences that seem to require thinking through new and different logical issues (quantum logic and the logic of induction, for instance). The remaining topics cover a part, at least, of the even more pressing problems involved in extending logical theory to nonscientific reasoning. The dominant goal, then, of philosophical logic is the extension of logical methods to nonmathematical reasoning domains. This goal has a theoretical dimension if (as many philosophical logicians seem to feel) it requires reworking and extending logical formalisms.

The development and testing of applications, such as the problem of formalizing the reasoning involved in getting to the airport, posed as a challenge in McCarthy 1959 (see Section 2.2, below), doesn't even appear as a category in the list of JPL topics, and in fact most of the philosophical logic literature is theoretical in nature, and is tested using philosophical techniques. Essentially, this means that the theories are motivated and tested with small-scale, artificial examples, selected by the theoreticians. These examples usually serve more as demonstrations or illustrations than as tests.
1.4 Logic in AI and Philosophical Logic

The rough comparison in Section 1.2 of the contents of the main publications for research in logical AI and philosophical logic suggests the following picture. Theoretical work in logical AI and in philosophical logic overlap to a large extent. Both are interested in developing nonmetamathematical applications of logic, and the core topics are very similar. This overlap is due not only to commonality of interest, but to direct influence of philosophical logic on logical AI; there is ample evidence, as we will see, that the first generation at least of AI logicists read and were influenced by the literature in philosophical logic.

Since that point, the specialties have diverged. New logical theories have emerged in logical AI (nonmonotonic logic is the most important example) which are not widely known in philosophical logic. Other differences are due to the AI community's interest in the theoretical analysis of algorithms and, of course, with their sense of the importance of implementations. Some have to do with the emerging development in computer science of ambitious applications using unprecedentedly large bodies of logical axioms. The sheer size of these applications produces new problems and new methodologies. And other differences originate in the interest of philosophical logicians in some topics (metaphysical topics, for instance) that are primarily inspired by purely philosophical considerations.

Concern for applications can be a great influence on how research is carried out and presented. The tradition in philosophical logic predates applications in automated reasoning, and to this day remains relatively uninterested in such applications. The methodology depends on intuitions, but without any generally accepted methodology for articulating and deploying these intuitions. And the ideas are illustrated and informed by artificial, small-scale examples.[6] In general, the philosophical literature does not deal with implementability or efficiency of the reasoning, or indeed with any features of the reasoning process. And it is hard to find cases in which the philosophical theories are illustrated or tested with realistic, large-scale reasoning problems.

These differences, however, are much more a matter of style than of substance or of strategic research goals. It is difficult to think through the details of the reasoning process without the computational tools to make the process concrete, and difficult to develop large-scale formalizations of reasoning problems without computational tools for entering, testing, and maintaining the formalizations. Because the core theoretical topics (modal, conditional and temporal logic, belief revision, and the logic of context) are so similar, and because the ultimate goal (the formalization of nonmathematical reasoning) is the same, I think of logic in AI as a continuous extension of the philosophical logic tradition.

The early influence of philosophical logic on logic in AI was profound. The bibliography of McCarthy & Hayes 1969, one of the most influential early papers in logical AI, illustrates the point well. There are 58 citations in the bibliography. Of these, 35 refer to the philosophical logic literature. (There are 17 computer science citations, one mathematical logic citation, one economics citation, and one psychology citation.) This paper was written at a time when there were hardly any references to logical AI in the computer science literature. Naturally, as logical AI has matured and developed as a branch of computer science, the proportion of cross-disciplinary citations has decreased. A sampling of articles from the first Knowledge Representation conference, Brachman et al. 1989, held in 1989, shows only 12 philosophical logic citations out of a total of 522 sampled citations; a sampling of articles from Cohn et al. 1998, held in 1998, shows 23 philosophical logic citations out of a total of 468 sampled.[7]

Despite the dramatic decrease in quantity of explicit citations, the contemporary literature in logical AI reflects an indirect acquaintance with the earlier literature in philosophical logic, since many of the computational papers that are explicitly cited in the modern works were influenced by this literature. Of course, the influence becomes increasingly distant as time passes, and this trend is accelerated by the fact that new theoretical topics have been invented in logical AI that were at best only dimly prefigured in the philosophical literature.

Although philosophical logic is now a relatively small field in comparison to logical AI it remains a viable area of research, with new work appearing regularly. But references to contemporary research in philosophical logic are rare in the AI literature. Similarly, the papers currently published in The Journal of Philosophical Logic, at least, do not show much influence from AI.[8] In Europe, the lines are harder to draw between professional divisions among logicians: some European journals, especially the Journal of Logic, Language, and Information, are successful in maintaining a focus in logic while attracting authors from all the disciplines in which logic is represented.
1.5 The Role of Artificial Intelligence in Logic

The importance of applications in logical AI, and the scale of these applications, represents a new methodology for logic—one that would have been impossible without mechanized reasoning. This methodology forces theoreticians to think through problems on a new scale and at a new level of detail, and this in turn has a profound effect on the resulting theories. The effects of this methodology will be illustrated in the sections below, dealing with various topics in logical AI. But the point is illustrated well by reasoning about action and change. This topic was investigated in the philosophical literature. Reasoning about change, at least, is part of tense logic, and the consequences of action are investigated in the literature on “seeing to it that”; see, for instance, Belnap 1996. The latter theory has no very robust account of action. The central construct is a variation on a branching-time modality of the sort that has been familiar since Prior 1967. Although it represents an interesting development in philosophical logic, the scale of the accomplishment is very different from the research tradition in logical AI reported in Section 4, below. The formalisms in this tradition not only support the formalization of complex, realistic planning problems, but provide entirely new insights into reasoning about the causal effects of actions, the persistence of states, and the interactions between actions and continuous physical processes. Developments such as this would have been impossible without the interactions between the logical theories and large-scale, practical applications in automated planning.

In Carnap 1955, Rudolf Carnap attempted to clarify intensional analyses of linguistic meaning, and to justify from a methodological point of view, by imagining how the analysis could be applied to the linguistic usage of a hypothetical robot. Carnap hoped that the fact that we could imagine ourselves to know the internal structure of the robot would help to make the case for an empirical science of semantics more plausible. This hope proved to be unjustified; the philosophical issue that concerned Carnap remains controversial to this day, and thought experiments with robots have not proved to be particularly rewarding in addressing it. Real robots, though, with real applications,[9] are a very different matter. Though it is hard to tell whether they will prove to be helpful in clarifying fundamental philosophical problems, they provide a laboratory for logic that is revolutionary in its potential impact on the subject. They motivate the development of entirely new logical theories that I believe will prove to be as important for philosophy as the fundamental developments of the late nineteenth century proved to be.

The emergence of separate mathematical and philosophical subspecialties within logic was not an entirely healthy thing for the field. The process of making mathematical logic rigorous and of demonstrating the usefulness of the techniques in pursuing mathematical ends that was pursued so successfully in the first half of the twentieth century represents a coherent refinement of logical methodology. All logicians should be pleased and proud that logic is now an area with a body of results and problems that is as substantial and challenging as those associated with most areas of mathematics.

But these methodological advances were gained at the expense of coverage. In the final analysis, logic deals with reasoning—and relatively little of the reasoning we do is mathematical, while almost all of the mathematical reasoning that nonmathematicians do is mere calculation. To have both rigor and scope, logic needs to keep its mathematical and its philosophical side united in a single discipline. In recent years, neither the mathematical nor the philosophical professions—and this is especially true in the United States—have done a great deal to promote this unity. But the needs of Computer Science, provide strong unifying motives. The professional standards for logical research in Computer Science certainly require rigor, but the field also puts its practitioners into contact with reasoning domains that are not strictly mathematical, and creates needs for innovative logical theorizing.

The most innovative and ambitious area of Computer Science, in terms of its coverage of reasoning, and the one that is closest in spirit to philosophical logic, is AI. This article will attempt to provide an introduction, for outsiders who are familiar with logic, to the aspects of AI that are closest to the philosophical logic tradition. This area of logic deserves, and urgently needs, to be studied by historians. But I am not a historian, and this entry does not pretend to be a history.
2. John McCarthy and Common Sense Logicism[10]
2.1 Logical AI

The most influential figure in logical AI is John McCarthy. McCarthy is one of the founders of AI, and has consistently advocated a research methodology that uses logical techniques to formalize the reasoning problems that AI needs to solve. All but the most recent work in McCarthy's research program can be found in Lifschitz 1990a, which also contains an introduction to McCarthy's work Lifschitz 1990b; for additional historical background, see Israel 1991.

McCarthy's methodological position has not changed substantially since it was first articulated in McCarthy 1959 and elaborated and amended in McCarthy & Hayes 1969. The motivation for using logic is that—even if the eventual implementations do not directly and simply use logical reasoning techniques like theorem proving—a logical formalization helps us to understand the reasoning problem itself. The claim is that without an understanding of what the reasoning problems are, it will not be possible to implement their solutions. Plausible as this Platonic argument may seem, it is in fact controversial in the context of AI; an alternative methodology would seek to learn or evolve the desired behaviors. The representations and reasoning that this methodology would produce might well be too complex to characterize or to understand at a conceptual level.

From McCarthy & Hayes 1969, it is clear that McCarthy thinks of his methodology for AI as overlapping to a large extent with traditional philosophy, but adding to it the need to inform the design of programs capable of manifesting general intelligence. This idea is not uncongenial to some philosophers (see, for instance, Carnap 1956 (pp. 244-247) and Pollock 1995). In practice, the actual theories that have emerged from McCarthy's methodology are influenced most strongly by work in philosophical logic, and the research tradition in logical AI represents a more or less direct development of this work, with some changes in emphasis. This review will concentrate on logical AI in relation to philosophical logic, without further comment on relations to philosophy in general or to the feasibility of developing human-level intelligent systems.
2.2 The Formalization of Common Sense

McCarthy's long-term objective is to formalize common sense reasoning, the prescientific reasoning that is used in dealing with everyday problems. An early example of such a problem, mentioned in McCarthy 1959, is getting from home to the airport. Other examples include:

Narrative understanding. The reasoning involved in reconstructing implicit information from narratives, such as sequencing of eventualities, and inferred causal connections.
Diagnosis. For instance, detecting faults in physical devices.
Spatial Reasoning. For instance, reasoning about the parts of rigid bodies and their shapes, and their relation to the shape of the whole.
Reasoning about the attitudes of other agents. For instance, making informed guesses about the beliefs and desires of other agents, not from “keyhole observation” but from conversational clues of the sort that could be obtained in a brief, interactive interview.

Stated baldly, the goal of formalizing common sense would probably seem outrageous to most philosophers, who are trained to think of common sense as rather elusive. But whether or not the ultimate goal is appropriate and achievable, the specific formalization projects that have emerged from this program have been successful in several ways. They have succeeded in breaking new territory for logic by extending the scope of the reasoning problems to which logical techniques can be successfully applied. They have demonstrated that logical techniques can be an important part of the solutions to specific AI problems—planning is the most successful of these, but some success has been achieved in other areas as well.[11] They form the basis of one approach to developing complete, autonomous agents.[12] And they have illuminated many specific forms of nonscientific reasoning—for instance, qualitative reasoning about the behavior of physical devices.[13]

Although McCarthy has advocated this program of formalization since 1959—an almost prehistorical date for AI—the program has not been taken on board and pursued by a dedicated community until recently. Research in this area has gained much momentum over the last 10 years. For further discussion and details, see Section 8.
3. Nonmonotonic Reasoning and Nonmonotonic Logics
3.1 Nonmonotonicity

Aristotle believed that most reasoning, including reasoning about what to do and about sublunary natural phenomena, dealt with things that hold “always or for the most part.” But Aristotelian logic deals only with patterns of inference that hold without exception. We find at the very beginning of logic a discrepancy between the scope of logical theory and common sense reasoning. Nonmonotonic logic is the first sustained attempt within logical theory to remedy this discrepancy. As such, it represents a potential for a sweeping expansion of the scope of logic, as well as a significant body of technical results.

The consequence relations of classical logics are monotonic. That is, if a set G of formulas implies a consequence C then a larger set G ? A will also imply C. A logic is nonmonotonic if its consequence relation lacks this property. Preferred models provide a general way to induce a nonmonotonic consequence relation. Invoke a function that for each G produces a subset MG of the models of G; in general, we will expect MG to be a proper subset of these models. We then say that G implies C if C is satisfied by every model in MG. As long as we do not suppose that MG?{A} ? MG, we can easily have an implication relation between G and C without imposing this relation on supersets of G.[14]

This model theoretic behavior corresponds to expectation-guided reasoning, where the expectations allow certain cases to be neglected. Here is an important difference between common sense and mathematics. Mathematicians are trained to reject a proof by cases unless the cases exhaust all the possibilities; but typical instances of common sense reasoning neglect some alternatives. In fact, it is reasonable to routinely ignore outlandish possibilities. Standing in my kitchen in California, wondering if I have time to wash my dishes before leaving for work, I do not take the possibility of an earthquake into account.

There seem to be many legitimate reasons for neglecting certain cases in common sense reasoning. A qualitative judgment that the probability of a case is negligible is one reason. But, for instance, in a planning context it may be reasonable to ignore even nonnegligible probabilities, as long as there is no practical point in planning on these cases.

The motivations for nonmonotonicity seem to involve a number of complex factors; probability (perhaps in some qualitative sense), normality, expectations that are reasonable in the sense that one can't be reasonably blamed for having them, mutual acceptance, and factors having to do with limited rationality. As far as I know, no one has succeeded in disentangling and clarifying these motivating considerations. In the early stages of its emergence in logical AI, many researchers seem to have thought of nonmonotonic reasoning as a general method for reasoning about uncertainty; but by the end of the 1980s, implementations of fully quantitative probabilistic reasoning were not only possible in principle, but were clearly preferable in many sorts of applications to methods involving nonmonotonic logic. A plausible and realistic rationale for nonmonotonic logic has to fit it into a broader picture of reasoning about uncertainty that also includes probabilistic reasoning.[15]
3.2 Historical Motivations

Three influential papers on nonmonotonic logic appeared in 1980: McDermott & Doyle 1980, Reiter 1980, and McCarthy 1980. In each case, the formalisms presented in these papers were the result of a gestation period of several years or more. To set out the historical influences accurately, it would be necessary to interview the authors, and this I have not done. However, there seem to have been two motivating factors: strategic considerations having to do with the long-range goals of AI, and much more specific, tactical considerations arising from the analysis of the reasoning systems that were being deployed in the 1970s.

Section 2.2 drew attention to McCarthy's proposed goal of formalizing common sense reasoning. The brief discussion above in Section 3.1 suggests that monotonicity may be an obstacle in pursuing this goal. An additional motive was found in Minsky 1974, which was widely read at the time. This paper presents an assortment of challenges for AI, focusing at the outset on the problem of natural language understanding.[16] Minsky advocates frame-based knowledge representation techniques[17] and (conceiving of the use of these representations as an alternative to logic), he throws out a number of loosely connected challenges for the logical approach, including the problem of building large-scale representations, of reasoning efficiently, of representing control knowledge, and of providing for the flexible revision of defeasible beliefs. In retrospect, I think most AI researchers would agree that these problems are general challenges to any research program in AI (including the one Minsky himself advocated at the time) and that logical techniques are an important element in addressing some, perhaps all, of the issues. (For instance, a well structured logical design can be a great help in scaling up knowledge representation.)

Minsky apparently intended to provide general arguments against logical methods in AI, but McDermott & Doyle 1980 and McCarthy 1980 interpret Minsky 1974 as a challenge that can be met by developing logics that lack the monotonicity property. Perhaps unintentionally, Minsky's paper seems to have provided some incentive to the nonmonotonic logicians by stressing monotonicity as a source of the alleged shortcomings of logic. In fact, the term ‘monotonicity’ apparently makes its first appearance in print in this paper.

The development of nonmonotonic logic also owes a great deal to the applied side of AI. In fact, the need for a nonmonotonic analysis of a number of AI applications was as persuasive as the strategic considerations urged by McCarthy, and in many ways more influential on the shape of the formalisms that emerged. Here, I will mention three such applications that appear to have been important for some of the early nonmonotonic logicians: belief revision, closed-world reasoning, and planning.
3.2.1 Belief Revision
Doyle 1979 provides an analysis and algorithm for a “truth maintenance system”. The TMS answers a general need, providing a mechanism for updating the “beliefs” of knowledge bases. The idea of the TMS is to keep track of the support of beliefs, and to use the record of these support dependencies when it is necessary to revise beliefs.

In a TMS, part of the support for a belief can consist in the absence of some other belief. This introduces nonmonotonicity. For instance, it provides for defaults; that Wednesday is the default day for scheduling a meeting means the belief that the meeting will be on Wednesday depends on the absence of special-case beliefs entailing that it will not be on Wednesday.

The TMS algorithm and its refinements had a significant impact on AI applications, and this created the need for a logical analysis. (In even fairly simple cases, it can be hard in the absence of analytic tools to see what consequences a TMS should deliver.) This provided a natural and highly specific challenge for those seeking to develop a nonmonotonic logic. The TMS also provided specific intuitions: the idea that the key to nonmonotonicity has to do with inferences based on unprovability was important for the modal approaches to nonmonotonic logic and for default logic. And the TMS's emphasis on interactions between arguments began a theme in nonmonotonic logic that remains important to this day. (See the discussion of argument-based approaches, in Section 3.4, below.)
3.2.2 Closed-world reasoning

The study of databases belongs to computer science, not specifically to AI. But one of the research paradigms in the scientific analysis of databases uses logical models of the representations and reasoning (see Minker 1997 for a recent survey of the field), and this area has interacted with logical AI. The deductive database paradigm was taking shape at about the same time that many AI researchers were thinking through the problems of nonmonotonic logic, and provided several specific examples of nonmonotonic reasoning that called for analyses. Of these, perhaps the most important is the closed-world assumption, according to which—at least as far as simple facts are concerned, represented in the database as positive or negative literals—the system assumes that it knows all that there is to be known. It is the closed world assumption that justifies a negative answer to a query “Is there a direct flight from Detroit to Bologna?” when the system finds no such flight in its data. This is another case of inference from the absence of a proof; a negative is proved, in effect, by the failure of a systematic attempt to prove the positive. This idea, which was investigated in papers such as Reiter 1978 and Clark 1978 also provided a challenge for nonmonotonic logics, as well as specific intuitions—note that again, the idea of inference rules depending on the absence of a proof is present here.
3.2.3 Planning

Rational planning is impossible without the ability to reason about the outcomes of a series of contemplated actions. Predictive reasoning of this sort is local; in a complex world with many features, we assume that most things will be unchanged by the performance of an action. But this locality has proved to be difficult to formalize. The problem of how to formalize this “inertia”[18] is known as the Frame Problem.

It is very natural to suppose that inertia holds by default; variables are unchanged by the performance of an action unless there is a special reason to think that they will change. This suggests that nonmonotonic temporal formalisms should provide an appropriate foundation for reasoning about action and change. So attempts to formalize the reasoning needed in planning also created a need for nonmonotonic logics. One of the earliest attempts to formalize nonmonotonic reasoning, Sandewall 1972, addresses the frame problem. Inertial defaults are an especially important and instructive case study; I will say no more about them here, since they are discussed in detail in Section 4.4, below.
3.3 The Earliest Formalisms

The three 1980 papers mentioned at the beginning of Section 3.2 represent three approaches to nonmonotonic logic that remain important subfields to this day: circumscription (McCarthy), modal approaches (Doyle & McDermott) and default logic (Reiter).

In McCarthy 1993a, McCarthy urges us, when considering the early history of circumscription, to take into account a group of three papers: McCarthy 1986, 1980, and 1987. The first paper connects the strategic ideas of McCarthy & Hayes 1969 with the need for a nonmonotonic logic, and sketches the logical ideas of domain circumscription, which is now classified as the simplest case of circumscription. The second paper provides more thorough logical foundations, and introduces the more general and powerful predicate circumscription approach. The third paper concentrates on developing techniques for formalizing challenging common sense examples.

All forms of circumscription involve restricting attention to models in which certain sets are minimized; for this reason, circumscription can be grouped with the preferred models approaches to nonmonotonicity: see Section 3.4, below. McCarthy's formalism is fairly conservative; though it raises interesting logical issues in higher-order logic and complexity, it uses familiar logical frameworks. And much of the focus is on the development of formalization techniques. The other varieties of nonmonotonic logic, including default logic and the modal nonmonotonic logics, raise issues of the sort that are familiar to philosophical logicians, having to do with the design of new logics, the systematic investigation of questions concerning validity, and managing the proliferation of alternative logics.

As the discussion above of truth maintenance indicated, it is very natural to think of nonmonotonic inferences as being hedged. That is, a nonmonotonic inference may require not merely the presence of a set of proved conclusions, but the absence of certain other conclusions. The general form of such a rule is:

In the presence of {A1,…,An} and in the absence of {B1,…,Bn}, conclude C.

An important special case of DR is a normal default, a simple rule to the effect that C holds by default, conditionally on assumptions A1,…,An. This can be formalized by taking the condition that must be absent to simply be the negation of the conclusion.

In the presence of {A1,…,An} and in the absence of ¬C, conclude C.

At first sight, it is somewhat perplexing how to formalize this notion of nonmonotonic inference, since it seems to require a circular definition of provability that can't be replaced with an inductive definition, as in the nonmonotonic case. The difficulty with the early theory of Sandewall 1972 is that it does not address this difficulty successfully. McDermott & Doyle 1980 and Reiter 1980 use fixpoint definitions to solve the problem. In both cases, the logical task is (1) to develop a formalism in which rules like DR can be expressed, and (2) to define the relation between a theory DT (which may incorporate such rules) and the theories E which could count as reasonable consequences of DT. In the terminology that later became standard, we need to define the relation between a theory DT and its extensions.

In retrospect, we can identify two sorts of approaches to nonmonotonic logic: those based on preference and those based on conflict. Theories of the first sort (like circumscription) involve a relatively straightforward modification of the ordinary model-theoretic definition of logical consequence that takes into account a preference relation over models. Theories of the second sort (like default logic) involve a more radical rethinking of logical ideas. The possibility of multiple extensions—different possible coherent, inferentially complete conclusion sets that can be drawn from a single set of premises—means that we have to think of logical consequence not as a function taking a set of axioms into its logical closure, but as a relation between a set of axioms and alternative logical closures. Since logical consequence is so fundamental, this represents a major theoretical departure. With multiple extensions, we can still retrieve a consequence relation between a theory and a formula in various ways, the simplest being to say that DT nonmonotonically implies C if C is a member of every extension of DT. Still, the conflict-based account of consequence provides a much richer underlying structure than the preferential one.

Reiter approaches the formalization problem conservatively. Nonmonotonicity is not expressed in the language of default logic, which is the same as the language of first-order logic. But a theory may involve a set of default rules—rules of the form DR. Reiter 1980 provides a fixpoint definition of the extensions of such a theory, and develops the theoretical groundwork for the approach, proving a number of the basic theorems.

Of these theorems, I mention one in particular, which will be used in Section 4.5, in connection with the Yale Shooting Anomaly. The idea is to take a conjectured extension (which will be a set T*) and to use this set for consistency checks in a proof-like process that applies default rules in successively to stages that begin with W.

We define a default proof process T0,T1,… for W, D, relative to T*, as follows.

Let T0 = W.
If no default rule in D is nonvacuously applicable to Ti relative to T*, then Ti+1 = ThFOL(Ti), where ThFOL(T) is the logical closure in FOL of T.
Otherwise, choose some default rule

A : B1,…,Bn

that is nonvacuously applicable to Ti relative to T*, and let

Ti+1 = ThFOL(Ti?{ C}).

In other words, as long as we can nonvacuously close the stage we are working on under an applicable default, we do so; otherwise, we do nothing. A theorem of Reiter's says that, under these circumstances:

T is an extension of if and only if there is a proof process T0, T1,... for W, D, relative to T, such that

T = 8
i=0 Ti.

Thus, we can show that T is an extension by (1) using T for consistency checks in a default reasoning process from , (2) taking the limit T' of this process, and (3) verifying that in fact T' = T.

The modal approach represents a “higher level of nonmonotonic involvement” than default logic. The unprovability construct is represented explicitly in the language, by means of a modal operator L informally interpreted as ‘provable’ (or, as in McDermott & Doyle 1980, by the dual of this operator).[19] Although McDermott and Doyle's terminology is different from Reiter's, the logical ideas are very similar—the essence of their approach, like Reiter's, is a fixpoint definition of the extensions of a nonmonotonic logic. Incorporating nonmonoticity in the object language creates some additional complexities, which in the early modal approach show up mainly in proliferation of the logics and difficulties in evaluating the merits of the alternatives. As better foundations for the modal approach emerged, it became possible to prove the expected theorems concerning equivalence of modal formalisms with default logic.[20]

Reiter's paper Reiter 1980 appears to have developed primarily out of tactical considerations. The earlier paper Reiter 1978 is largely concerned with providing an account of database queries. Unlike the other seminal papers in nonmonotonic logic, Reiter's shows specific influence from the earlier and independent work on nonmonotonicity in logic programming—the work seems to have been largely inspired by the need to provide logical foundations for the nonmonotonic reasoning found in deductive databases. Doyle and McDermott's paper shows both strategic and tactical motivation—citing the earlier literature in logicist AI, it motivates nonmonotonic logic as part of a program of modeling common sense rationality. But the theory is also clearly influenced by the need to provide a formal account of truth maintenance.
3.4 Approaches to Nonmonotonic Logic

Nonmonotonic logic is a complex, robust research field. Providing a survey of the subject is made difficult by the fact that there are many different foundational paradigms for formalizing nonmonotonic reasoning, and the relations between these paradigms is not simple. An adequate account of even a significant part of the field requires a something like a book-length treatment. A number of books are available, including Brewka et al. 1997, Brewka 1991, Schlechta 1997, Lukaszewicz 1990, Antoniou 1997, Besnard 1992, and Marek & Truszczynski 1994. Two collections are especially useful: Ginsberg 1987 and Gabbay et al. 1994. The former is a useful source for readers interested in the early history of the subject, and has an excellent introduction. The handbook chapters in Gabbay et al. 1994 provide overviews of important topics and approaches. My current recommendation for readers interested in a quick, readable introduction to the topic would be Brewka et al. 1997 and self-selected chapters of Gabbay et al. 1994. I will rely on these references for technical background, and will concentrate on intellectual motivation, basic ideas, and potential long-term significance for logic.
3.4.1 Preference Semantics

At the outset in Section 3.1, I mentioned how preferred models could be used to characterize a nonmonotonic consequence relation. This general model theory of nonmonotonicity emerged in Shoham 1988, five years after the work discussed in Section 3.2, and represents a much more general and abstract approach.

Preferential semantics relies on a function S taking a set K of models into a subset S(K) of K. The crucial definition of preferential entailment stipulates that A is a (nonmonotonic) consequence of G if every model M of S(Models(G)) implies A. Shoham's theory is based on a partial order ? over models: S(K) can then be characterized as the set of models in K that are ?-minimal in K. To ensure that no set can preferentially entail a contradiction unless it classically entails a contradiction, infinite descending ? chains need to be disallowed.

This treatment of nonmonotonicity is similar to the earlier modal semantic theories of conditionals—the similarities are particularly evident using the more general theories of conditional semantics, such as the one presented in Chellas 1975. Of course, the consequence relation of the classical conditional logics is monotonic, and conditional semantics uses possible worlds, not models. But the left-nonmonotonicity of conditionals (the fact that A box-arrow C does not imply [A ? B] Box-arrow C) creates issues that parallel those in nonmonotonic logics. Early work in nonmonotonic logic does not seem to be aware of the analogy with conditional logic. But the interrelations between the two have become an important theme more recently; see, for instance, Gärdenfors & Makinson 1994, Boutilier 1992, Pearl 1994, Gabbay 1995, Benferat et al. 1997, Delgrande 1998, Arlo-Costa & Shapiro 1992, Alcourrón 1995, and Asher 1995.

Preference semantics raises an opportunity for formulating and proving representation theorems relating conditions over preference relations to properties of the abstract consequence relation. This line of investigation began with Lehmann & Magidor 1992.
3.4.2 Modal and epistemic theories

Neither Doyle or McDermott pursued the modal approach much beyond the initial stages of McDermott 1982, and McDermott & Doyle 1980. With a helpful suggestion from Robert Stalnaker (see Stalnaker 1993), however, Robert C. Moore produced a modal theory that improves in many ways on the earlier ideas. Moore gives the modal operator of his system an epistemic interpretation, based on the conception of a default rule as one that licenses a conclusion for a reasoning agent unless something that the agent knows blocks the conclusion. In Moore's autoepistemic logic, an extension E of a theory T is a superset of T that is stable, i.e., that is deductively closed, and that satisfies the following two rules:

If A?E then ?A?E.
If A?E then ¬?A?E.

It is also usual to impose a groundedness condition on autoepistemic extensions of T, ensuring that every member of an extension has some reason tracing back to T. Various such conditions have been considered; the simplest one restricts extensions to those satisfying

E is the set of nonmodal consequences of T?{A : ?A?E} ? {¬?A : A?E}.

Autoepistemic logic remains a popular approach to nonmonotonic logic, in part because of its usefulness in providing theoretical foundations for logic programming. For more recent references, see Marek & Truszczynski 1991, Moore 1995b, Marek & Truszczynski 1989 Konolige 1994, Antoniou 1997, and Moore 1993.

Epistemic logic has inspired other approaches to nonmonotonic logic. Like other modal theories of nonmonotonicity, these use modality to reflect consistency in the object language, and so allow default rules along the lines of DR to be expressed. But instead of consistency, these use ignorance. See Halpern & Moses 1985 and Levesque 1987 for variations on this idea. These theories are explained, and compared to other nonmonotonic logics, in Meyer & van der Hoek 1995. In more recent work, Levesque's ideas are systematically presented and applied to the theory of knowledge bases in Levesque & Lakemeyer 2000.
3.5 Further Topics

This brief historical introduction to nonmonotonic logic leaves untouched even a number of general topics that might well be of interest to a nonspecialist. These include graph-based and proof-theoretic approaches to nonmonotonic logic, results that interrelate the various formalisms, complexity results, tractable special cases of nonmonotonic reasoning, relations between nonmonotonic and abductive reasoning, relations to probability logics, the logical intuitions and apparent patterns of validity underlying nonmonotonic logics, and the techniques used to formalize domains using nonmonotonic logics. For these and other topics I have to refer the reader to the literature. As a start, the chapters in Gabbay et al. 1994 are highly recommended.
4. Reasoning about Action and Change[21]
4.1 Priorian Tense Logic

Time and temporal reasoning have been associated with logic since the origins of scientific logic with Aristotle. The idea of a logic of tense in the modern sense has been familiar since at least the work of Jan Lukasiewicz (see, for instance, Lukasiewicz 1970), but the shape of what is commonly known as tense logic was standardized by Arthur Prior's work in the 1950s and 1960s: see Prior 1956, 1967, 1968.[22] As the topic was developed in philosophical logic, tense logic proved to be a species of modal logic; Prior's work was heavily influenced by both Hintikka and Kripke, and by the idea that the truth of tense-logical formulas is relative to world-states or temporal stages of the world; these are the tense-theoretic analogues of the timeless possible worlds of ordinary modal logic. Thus, the central logical problems and techniques of tense logic were borrowed from modal logic. For instance, it became a research theme to work out the relations between axiomatic systems and the corresponding model theoretic constraints on temporal orderings. See, for instance, Burgess 1984 and van Benthem 1983.

Priorian tense logic shares with modal logic a technical concentration on issues that arise from using the first-order theory of relations to explain the logical phenomena, an expectation that the important temporal operators will be quantifiers over world-states, and a rather distant and foundational approach to actual specimens of temporal reasoning. Of course, these temporal logics do yield validities, such as


(if A, then it was the case that A was going to be the case), which certainly are intuitively valid. But at most, these can only play a broadly foundational role in accounting for realistic reasoning about time. It is hard to think of realistic examples in which they play a leading part.

This characteristic, of course, is one that modal logic shares with most traditional and modern logical theories; the connection with everyday reasoning is rather weak. Although modern logical techniques do account with some success for the reasoning involved in verifying mathematical proofs and logic puzzles, they do not explain other cases of technical or common sense reasoning with much detail or plausibility. Even in cases like legal reasoning, where logicians and logically-minded legal theorists have put much effort into formalizing the reasoning, the utility of the results is controversial.
4.2 Planning Problems and the Situation Calculus

Planning problems provide one of the most fruitful showcases for combining logical analysis with AI applications. On the one hand there are many practically important applications of automated planning, and on the other logical formalizations of planning are genuinely helpful in understanding the problems and in designing algorithms.

The classical representation of an AI planning problem, as described in Amarel 1968, evidently originates in early work of Herbert Simon's, published in a 1966 CMU technical report, Simon 1966. In such a problem, an agent in an initial world-state is equipped with a set of actions, which are thought of as partial functions transforming world-states into world-states. Actions are feasible only in world-states that meet certain constraints (these constraints are now called the “preconditions” of the action). A planning problem then becomes a search for a series of feasible actions that successively transform the initial world-state into a desired world-state.

The Situation Calculus, developed by John McCarthy, is the origin of most of the later work in formalizing reasoning about action and change. It was first described in 1969, in McCarthy 1983; the earliest generally accessible publication on the topic is McCarthy & Hayes 1969.

Apparently, Priorian tense logic had no influence on Amarel 1968. But there is no important difference between Amarel's world-states and those of Priorian tense logic. The “situations” of the Situation Calculus are these same world-states, under a new name.[23] They resemble possible worlds in modal logic in providing abstract locations that support a consistent and complete collection of truths. As in tense logic, these locations are ordered, and change is represented by the variation in truths from one location to another. The crucial difference between the Situation Calculus and tense logic is that change in the situation is dynamic—changes do not merely occur, but occur for a reason.

This difference, of course, is inspired by the intended use of the Situation Calculus: it is meant to formalize Simon's representation of the planning problem, in which a single agent reasons about the scenarios in which a series of actions is performed.[24] In this model, what drives change is the performance of actions, so the fundamental model theoretic relation is the relation


between an action a, an initial situation s in which a is performed, and a resulting situation s' immediately subsequent to the performance of the action. Usually (though this is not absolutely necessary) the deterministic assumption is made that s' is unique. In general, actions can be successfully performed only under certain limited circumstances. This could be modeled by allowing for cases in which there is no s' such that RESULT(a,s,s'). But usually, it is assumed that RESULT is in fact a total function, but that in cases in which s does not meet the “preconditions” of a, there are no restrictions on the s' satisfying RESULT(a,s,s'), so that the causal effects of a will be entirely unconstrained in such cases.

A planning problem starts with a limited repertoire of actions (where sets of preconditions and effects are associated with each action), an initial situation, and a goal (which can be treated as a formula). A planning problem is a matter of finding a sequence of actions that will achieve the goal, given the initial situation. That is, given a goal G and initial situation s, the problem will consist of finding a sequence s1,...,sn of actions which will transform s into a final situation that satisfies G. This means (assuming that RESULT is a function) that G will be satisfied by the situation sn, where s0 = s and si+1 is the s' such that RESULT(ai+1,si,s'). The planning problem is in effect a search for a sequence of actions meeting these conditions. The success conditions for the search can be characterized in a formalism like the Situation Calculus, which allows information about the results of actions to be expressed.

Nothing has been said up till now about the actual language of the Situation Calculus. The crucial thing is how change is to be expressed. With tense logic in mind, it would be natural to invoke a modality like [a]A, with the truth condition

?s[a]A iff ?s'A, where RESULT(a,s,s').

This formalization, in the style of dynamic logic, is in fact a leading candidate; see Section 4.7, below.

But McCarthy & Hayes 1969 deploys a language that is much closer to first-order logic. (This formalization style is characteristic of McCarthy's work; see McCarthy 1979.) Actions are treated as individuals. And certain propositions whose truth values can change over time (propositional fluents) are also treated as individuals. Where s is a situation and f is a fluent, Holds(f,s) says that f is true in s.
4.3 Formalizing Microworlds

Since the pioneering work of the nineteenth and early twentieth century logicians, the process of formalizing mathematical domains has largely become a matter of routine. Although (as with set theory) there may be controversies about what axioms and logical infrastructure best serve to formalize an area of mathematics, the methods of formalization and the criteria for evaluating them are relatively unproblematic. This methodological clarity has not been successfully extended to other domains; even the formalization of the empirical sciences presents difficult problems that have not yet been resolved.[25]

The formalization of temporal reasoning, and in particular of reasoning about actions and plans, is the best-developed successful extension of modern formalization techniques to domains other than mathematical theories. This departure has required the creation of new methodologies. One methodological innovation will emerge in Section 4.5: the development of a library of scenarios for testing the adequacy of various formalisms, and the creation of specialized domains likes the blocks-world domain (mentioned above, in Section 4.2) that serve a laboratories for testing ideas. For more on the blocks world, see Genesereth & Nilsson 1987; Davis 1991. McCarthy's ideas about elaboration tolerance McCarthy 1999 provide one interesting attempt to provide a criterion for the adequacy of formalizations. Another idea that has emerged in the course of formalizing common sense domains is the importance of an explicit ontology; see, for instance, Fikes 1996; Lenat & Guha 1989. Another is the potential usefulness of explicit representations of context; see Guha, 1991. Another is the use of simulation techniques: see, for instance, Johnstone & Williamson 2007

Temporal reasoning, and in particular reasoning about actions and plans, is the best-developed domain. At least one important methodology will emerge in Section 4.5, below: the development of a library of scenarios for testing the adequacy of various formalisms, and the creation of specialized domains likes the blocks-world domain (mentioned above, in Section 4.2) that serve a laboratories for testing ideas. For more on the blocks world, see Genesereth & Nilsson 1987; Davis 1991. McCarthy's ideas about elaboration tolerance McCarthy 1999 provide one interesting attempt to provide a criterion for the adequacy of formalizations. Still other important ideas have emerged in the course of formalizing common sense domains. One is the importance of an explicit ontology; see, for instance, Fikes 1996; Lenat & Guha 1989. Another is the potential usefulness of explicit representations of context; see Guha, 1991.
4.4 Prediction and the Frame Problem

To tell whether a plan achieves its goal, you need to see whether the goal holds in the plan's final state. Doing this requires predictive reasoning, a type of reasoning that was, as far as I know, entirely neglected in the tense-logical literature. As in mechanics, prediction involves the inference of later states from earlier ones. But (in the case of simple planning problems at least) the dynamics are determined by actions rather than by differential equations. The investigation of this qualitative form of temporal reasoning, and of related sorts of reasoning (e.g., plan recognition, which seeks to infer goals from observed actions, and narrative explanation, which seeks to fill in implicit information in a temporal narrative) is one of the most impressive chapters in the brief history of common sense logicism.

The essence of prediction is the problem of inferring what holds in the situation that ensues from performing an action, given information about the initial situation. I will assume that the agent has complete knowledge about the initial situation—this assumption is usual in classical formalizations of planning.[26]

A large part of the qualitative dynamics that is needed for planning consists in inferring what does not change. Take a simple plan to type the word ‘cat’ using word processing software: my plan is to first enter ‘c’, then enter ‘a’, then enter ‘t’. Part of my confidence in this plan is that the actions are independent: for instance, entering ‘a’ does not also erase the ‘c’. The required inference can be thought of as a form of inertia. The Frame Problem is the problem of how to formalize the required inertial reasoning.

The Frame Problem was named and introduced in McCarthy & Hayes 1969. Unlike most of the philosophically interesting technical problems to emerge in AI, it has attracted the interest of philosophers; most of the relevant papers, and background information, can be found in Ford & Pylyshyn 1996; Pylyshyn 1987. Both of these volumes document interactions between AI and philosophy.

The quality of these interactions is discouraging; as a philosopher, I even find it somewhat embarrassing. Like any realistic common sense reasoning problem, the Frame Problem is open-ended, and can depend on a wide variety of circumstances. If I put $20 in my wallet and go to the store with the wallet in my pocket, I can safely assume that the $20 is still in my wallet. If I leave the $20 on the counter at the store while shopping, I can't safely assume it will be there when I get back. This may account for the temptation that makes some philosophers[27] want to construe the Frame Problem very broadly, so that very soon it becomes indiscernible from the problem of formalizing general common sense in arbitrary domains. Such a broad construal may serve to introduce speculative discussions concerning the nature of AI, but it loses all contact with the genuine, new logical problems in temporal reasoning that have been discovered by the AI community. It provides a forum for repeating some familiar philosophical themes, but it brings nothing new to philosophy. I find this approach disappointing because I believe that philosophy can use all the help it can get, and that the AI community has succeeded in extending and enriching the application of logic to common sense reasoning in dramatic ways that are highly relevant to philosophy. The clearest account of these developments to be found in the volumes edited by Pylyshyn is Morgenstern 1996. A recent extended treatment can be found in Shanahan 1997; also see Sandewall 1994. The purely logical Frame Problem can be solved using monotonic logic, by simply writing explicit axioms stating what does not change when an action is performed. This technique can be successfully applied to quite complex formalization problems.[28] But nonmonotonic solutions to the framework have been extensively investigated and deployed; these lead to new and interesting lines of logical development.

Some philosophers Fodor 1987; Lormand 1996 have felt that contrived propositions will pose special difficulties in connection with the Frame Problem. As Shanahan points out Shanahan 1997 [p. 24]) Fodor's “fridgeon” example is readily formalized in the Situation Calculus and poses no special problems. However, as Lormand suggests, Goodman's examples Goodman, 1946 do create problems if they are admitted as fluents; there will be anomalous extensions in which objects change from green to blue in order to preserve their grueness.

This is one of the few points that I can find in the philosophical literature on the Frame Problem that raises a genuine difficulty for the formal solutions. But the difficulty is peripheral, since the example is not realistic. Recall that fluents are represented as first-order individuals. Although fluents are situation-dependent functions, an axiom of comprehension is certainly not assumed for fluents. In fact, it is generally supposed that the domain of fluents will be a very limited set of the totality of situation-dependent functions; typically, it will be a relatively small finite set of important variables, and will be chosen in particular cases much as a set of variables is chosen in statistical modeling.

I know of no systematic account in the AI literature of how to choose an appropriate set of fluents, but it would certainly be part of such an account that all fluents should correspond to projectable predicates, in Goodman's sense.
4.5 Nonmonotonic Treatments of Inertia and a Package of Problems

The idea behind nonmonotonic solutions to the Frame Problem is to treat inertia as a default; changes are assumed to occur only if there is some special reason for them to occur. In an action-centered account of change, this means that absence of change is assumed when an action is performed unless a reason for the change can be found in axioms for the action.

For explicitness, I will use Reiter's default logic to illustrate the formalization. Recall that in Reiter's theory, defaults are represented as rules, not formulas, so that they are not subject to quantification. To formalize inertia, then, we need to use default rule schemata. For each fluent f, action a, and situation s, the set of these schemata will include an instance of the following schema:

IR(f,a,s) T : Holds(f,s) ? Holds(f,RESULT(a,s)) Holds(f,s) ? Holds(f,RESULT(a,s))

This way of doing things makes any case in which a fluent changes truth value a prima facie anomaly. But it follows from Reiter's account of extensions that such defaults are overridden when they conflict with the monotonic theory of situation dynamics. So if, for instance, there is a monotonic causal axiom for the action blacken ensuring that blackening a block will make it black in the resulting situation, then the appropriate instance of IR will be inefficacious, and there will be no extension in which a white block remains white when it is blackened.

The Frame Problem somehow managed to capture the attention of a wide community—but if one is interested in understanding the complex problems that arise in generalizing formalisms like the Situation Calculus, while at the same time ensuring that they deliver plausible solutions to a wide variety of scenarios, it is more useful to consider a larger range of problems. For the AI community, the larger problems include the Frame Problem itself, the Qualification Problem, the Ramification Problem, generalizability along a number of important dimensions including incomplete information, concurrency (multiple agents), and continuous change, and finally a large assortment of specific challenges such as the scenarios mentioned later in this section.

The Qualification Problem arises generally in connection with the formalization of common sense generalizations. Typically, these involve exceptions, and these exceptions—especially if one is willing to entertain far-fetched circumstances—can iterate endlessly. The same phenomenon, under a label like ‘the problem of ceteris paribus generalizations', is familiar from analytic philosophy. It also comes up in the semantics of generic constructions found in natural languages.[29] In a sense, this problem is addressed at a general level by nonmonotonic logics, which—though they do not provide a way to enumerate exceptions—do allow common sense generalizations to be formulated as defaults, as well as enabling further qualifications to be added nondestructively. Ideally, then, the initial generalization can be stated as an axiom and qualifications can be added incrementally in the form of further axioms.

The Qualification Problem was raised in McCarthy 1986, where it was motivated chiefly by generalizations concerning the consequences of actions; McCarthy considers in some detail the generalization that turning the ignition key in an automobile will start the car. Much the same point, in fact, can be made about virtually any action, including stacking one block on another—the standard action that is used to illustrate the Situation Calculus. A circumscriptive approach to the Qualification Problem is presented in Lifschitz 1987; this explicitly introduces the precondition relation between an action and its preconditions into the formalism, and circumscriptively minimizes preconditions, eliminating from preferred models any “unknown preconditions” that might render an action inefficacious.

Several dimensions of the Qualification Problem remain as broad, challenging research problems. For one thing, not every nonmonotonic logic provides graceful mechanisms for qualification. Default logic, for instance, does not deliver the intuitively desired conclusions. Suppose one formalizes the common sense generalization that if I press the ‘a’ key on my computer it will type ‘a’ as a normal default:
T?:?VALUE(text, RESULT(Press-a,s)) = VALUE(text,s) + ‘a’
VALUE(text, RESULT(Press-a,s)) = VALUE(text,s) + ‘a’

If I then formalize the exception to this generalization that if I press the ‘a’ key while the Alt key is depressed the cursor moves to the beginning of the current sentence as a normal default along the same lines, I get two extensions: one in which pressing ‘a’ while the Alt key is depressed adds ‘a’ to the text and another in which it moves the cursor. The problem is that default logic does not provide for more specific defaults to override ones that are more general. This principle of specificity has been discussed at length in the literature. Incorporating it in a nonmonotonic logic can complicate the theory considerably; see, for instance, Asher & Morreau 1991 and Horty 1994. And, as Elkan 1995 points out, the Qualification Problem raises computational issues.

Relatively little attention has been given to the Qualification Problem for characterizing actions, in comparison with other problems in temporal reasoning. In particular, the standard accounts of unsuccessful actions are somewhat unintuitive. In the formalization of Lifschitz 1987, for instance, actions with some unsatisfied preconditions are only distinguished from actions whose preconditions all succeed in that the conventional effects of the action will only be ensured when the preconditions are met. It is as if an action of spending $1,000,000 can be performed at any moment—although if you don't have the money, no effects in particular will be guaranteed. [30] And there is no distinction between actions that cannot even be attempted (like boarding a plane in London when you are in Sydney), actions that can be attempted, but in which the attempt can be expected to go wrong (like making a withdrawal when you have insufficient funds), actions that can be attempted with reasonable hope of success, and actions that can be attempted with guaranteed success. As J.L. Austin made clear in Austin 1961, the ways in which actions can be attempted, and in which attempted actions can fail, are a well developed part of common sense reasoning. Obviously, in contemplating a plan containing actions that may fail, one may need to reason about the consequences of failure. Formalizing the pathology of actions, providing a systematic theory of ways in which actions and the plans that contain them can go wrong, would be a useful addition to planning formalisms, and one that would illuminate important themes in philosophy.

The challenge posed by the Ramification Problem (characterized first in Finger 1987) is to formalize the indirect consequences of actions, where “indirect” effects are not delayed,[31] but are temporally immediate but causally derivative. If I walk into a room, the direct effect is that I am in now the room. There are also many indirect effects: for instance, my shirt also is now in the room. You can see from this that the formulation of the problem presupposes a distinction between direct consequences of actions (ones that attach directly to an action, and that are ensured by the successful performance of the action) and other consequences. This assumption is generally accepted without question in the AI literature on action formalisms. You can make a good case for its common sense plausibility—for instance, many of our words for actions (‘to warm’, to ‘lengthen’, ‘to ensure’) are derived from the effects that are conventionally associated with them. And in these cases, success is entailed: if someone has warmed something, this entails that it became warm. [32] A typical example is discussed in Lin 1995: a certain suitcase has two locks, and is open if and only if both locks are open. Then (assuming that actions are not performed concurrently) opening one lock will open the suitcase if and only if the other lock is open. Here, opening a lock is an action, with direct consequences; opening a suitcase is not an action, it is an indirect effect.

Obviously, the Ramification Problem is intimately connected with the Frame Problem. In approaches that adopt nonmonotonic solutions to the Frame Problem, inertial defaults will need to be overridden by conclusions about ramifications in order to obtain correct results. In case the left lock of the suitcase is open, for instance, and an action of opening the right lock is performed, then the default conclusion that the suitcase remains closed needs somehow to be suppressed. Some approaches to the Ramification Problem depend on the development of theories of common sense causation, and therefore are closely related to the causal approaches to reasoning about time and action discussed below in Section 4.6. See, for instance, Giunchiglia et al. 1997; Thielscher 1989; Lin 1995.

Philosophical logicians have been content to illustrate their ideas with relatively small-scale examples. The formalization of even large-scale mathematical theories is relatively unproblematic. Logicist AI is the first branch of logic to undertake the task of formalizing large examples involving nontrivial common sense reasoning. In doing so, the field has had to invent new methods. An important part of the methodology that has emerged in formalizing action and change is the prominence that is given to challenges, posed in the form of scenarios. These scenarios represent formalization problems which usually involve relatively simple, realistic examples designed to challenge the logical theories in specific ways. Typically, there will be clear common sense intuitions about the inferences that should be drawn in these cases. The challenge is to design a logical formalism that will provide general, well-motivated solutions to these benchmark problems.

Among the many scenarios that have been discussed in the literature are the Baby Scenario, the Bus Ride Scenario, the Chess Board Scenario, the Ferryboat Connection Scenario, the Furniture Assembly Scenario, the Hiding Turkey Scenario, the Kitchen Sink Scenario, the Russian Turkey Scenario, the Stanford Murder Mystery, the Stockholm Delivery Scenario, the Stolen Car Scenario, the Stuffy Room Scenario, the Ticketed Car Scenario, the Walking Turkey Scenario, and the Yale Shooting Anomaly. Accounts of these can be found in Shanahan 1997 and Sandewall 1994; see especially Sandewall 1994[Chapters 2 and 7].

Many of these scenarios are designed to test advanced problems that I will not discuss here—for instance, challenges dealing with multiple agents, or with continuous changes. Here, I will concentrate on one of the earliest, and probably the most subtle of these scenarios: the Yale Shooting Anomaly, first reported in Hanks & McDermott 1985 and published in Hanks & McDermott 1986; Hanks & McDermott 1987.

The Yale Shooting Anomaly involves three actions: load, shoot, and wait. A propositional fluent Loaded tracks whether a certain pistol is loaded; another fluent, Alive, tracks whether a certain person, Fred, is alive. load has no preconditions; its only effect is Loaded. The fluent shoot has Loaded as its only precondition and Alive as a negative effect; wait has no preconditions and no effects.

Causal information regarding the axioms is formalized as follows.
Load ?sHolds(Loaded, RESULT(load,s))
Shoot 1 ?s[Holds(Loaded,s) ? Holds(¬Alive, RESULT(shoot,s))]
Shoot 2 ?s[Holds(Loaded,s) ? Holds(¬Loaded, RESULT(shoot,s))]

There is no Wait Axiom—that is, wait has no preconditions and no effects.

We will formalize the inertial reasoning in this scenario using a nonmonotonic logic—to be specific, we use Reiter's default logic. The set D of defaults for this theory consists of all instances of the inertial schema IR. In the initial situation, Fred is alive and the pistol is unloaded.

IC1 Holds(Alive,s0)
IC2 ¬Holds(Loaded,s0)

The monotonic theory W of the scenario consists of: (1) the action axioms Load, Shoot 1 and Shoot 2 and (2) the initial conditions IC1 and IC2. Let s1 = RESULT(load,s0), s2 = RESULT(wait,s1), and s3 = RESULT(shoot,s2).

The Yale Shooting Anomaly consists of the fact that the theory allows an extension in which the actions are load; shoot; wait, and in the final situation s3, the pistol is unloaded and Fred is alive. The initial situation in the Anomaly and the three actions, with their resulting situations, can be pictured as follows.

s0 load
arrow s1 wait
arrow s2 shoot
arrow s3

The natural, expected outcome of these axioms is that the pistol is loaded and Fred is alive after waiting, so that shooting yields a final outcome in which Fred is not alive and the pistol is unloaded. There is no problem in showing that this corresponds to an extension; the problem is the presence of the other, anomalous extension, which looks like this.
s0 load
arrow Alive
s1 wait
arrow Alive
s2 shoot
arrow Alive

Here is a narrative version of this extension. At first, Fred is alive and the pistol is unloaded. After loading, the pistol is loaded and Fred remains alive. After waiting, the pistol becomes unloaded and Fred remains alive. Shooting is then vacuous since the pistol is unloaded, so finally, after shooting, Fred remains alive and the pistol remains unloaded.

The best way to see clearly that this is an extension is to work through the proof. Less formally, though, you can see that the expected extension violates just one default: the frame default for Alive is violated when Fred changes state in the last step. But the anomalous extension also violates only one default: the frame default for Loaded is violated when the pistol spontaneously becomes unloaded while waiting. So, if you just go by the number of defaults that are violated, both extensions are equally good.

The Yale Shooting Anomaly represents a major obstacle in developing a theory of predictive reasoning. A plausible, well-motivated logical solution to the Frame Problem runs afoul of a simple, crisp example in which it clearly delivers the wrong results. Naturally, the literature concerning the Yale Shooting Problem is extensive. Surveys of some of this work, with bibliographical references, can be found in Shanahan 1997; Morgenstern 1996.
4.6 Some Emergent Frameworks

Many formalisms have been proposed to deal with the problems surveyed in the previous section. Some are more or less neglected today. Several are still advocated and defended by leading experts; some of these are associated with research groups who are not only interested in developments of logical theory, but in applications in planning and cognitive robotics.

The leading approaches provide solutions to the main problems mentioned in Section 4.5, and to many of the scenarios designed to test and illustrate theories of reasoning about action and change. It is commonly agreed that good solutions need to be generalizable to more complex cases than the early planning formalisms, and that in particular the solutions they offer should be deployable even when continuous time, concurrent actions, and various kinds of ignorance are allowed. Also, it is generally agreed that the formalisms should support several kinds of reasoning, and, in particular, not only prediction and plan verification but retrodiction, i.e., construction of a sequence of states and actions from partial information, presented in narrative form.

I will describe four approaches here: (1) Features and fluents (Sandewall), (2) Motivated Action Theory (Morgenstern and Stein), (3) State Minimization in the Event Calculus (Shanahan) and Causal Theories (Lifschitz and others). My accounts of the first three will be fairly brief; fortunately, each approach is well documented in a single reference. I believe that the fourth approach is likely to be most interesting to philosophers, and that it contains elements that will be of lasting importance whatever changes future developments in this area may bring.
4.6.1 Features and fluents

This approach, described in Sandewall 1994, uses preference semantics as a way to organize nonmonotonic solutions to the problems of reasoning about action and change. Rather than introducing a single logical framework, Sandewall considers a number of temporal logics, including ones that use discrete, continuous, and branching time. The properties of the logics are systematically tested against a large suite of test scenarios.
4.6.2 Motivated Action Theory

This theory grew out of direct consideration of the problems in temporal reasoning described above in Section 4.5, and especially the Yale Shooting Scenario. In Morgenstern & Stein 1994, Morgenstern and Stein seek to find a general, intuitively motivated logical framework that solves the difficulties. They settle on the idea that unmotivated actions are to be minimized, where an action (“actions” construed generally enough to include any change) can be motivated directly, e.g. by an axiom, or indirectly, through chains of motivations. The key technical idea of the paper is a (rather complicated) definition of motivation in an interval-based temporal logic. In Morgenstern 1996, Morgenstern presents a summary of the theory, along with reasons for rejecting its causal rivals. The most important of these reasons is that these theories, based on the Situation Calculus, do not appear to generalize to cases allowing for concurrency and ignorance. She also cites the failure of early causal theories to deal with retrodiction.
4.6.3 State-Based Minimization in the Event Calculus

In Baker 1989, Andrew Baker presented a solution to the version of the Yale Shooting problem in the Situation Calculus, using a circumscriptive inertial axiom. The very brief account of circumscription above in Section 3 indicated that circumscription uses preferred models in which the extensions of certain predicates are minimized. In the course of this minimization, a set of parameters (including, of course, the predicates to be minimized) is allowed to vary; the rest are held constant. Which parameters vary and which are held constant is determined by the application.

In the earliest circumscriptive solutions to the Frame Problem, the inertial rule CIR is stated using an abnormality predicate.

?f?s?a[¬Ab(f,a,s) ? [Holds(f,s) ? Holds(f,RESULT(a,s))]]

This axiom uses a biconditional, so that it can be used for retrodiction; this is typical of the more recent formulations of common sense inertia. In circumscribing, the abnormality predicate is minimized while the Holds predicate is allowed to vary and all other parameters are fixed. This formalization succumbs to the Yale Shooting Anomaly in much the same way that default logic does. (Circumscription does not involve multiple extensions, so the problem emerges as the nonderivability of the conclusion that Fred is alive after the occurrence of the shooting.)

In Baker's reformulation of the problem, separate axioms ensure the existence of a situation corresponding to each Boolean combination of fluents, and the RESULT function is allowed to vary, while the Holds predicate is held constant. In this setting, the RESULT function needs to be specified for “counterfactual”actions—in particular, for shooting as well as for waiting in the Yale Shooting Anomaly. It is this feature that eliminates the incorrect model for that scenario; for details, see Baker 1989 and Shanahan 1997, Chapter 6.

This idea, which Shanahan calls “State-Based Minimization,” is developed and extended in Shanahan 1997, in the context of a temporal logic deriving from the Event Calculus of Kowalski and Sergot; see Kowalski & Sergot 1986. Shanahan's formalism has the advantage of being closely connected to implementations using logic programming.
4.6.4 Causal Theories

Recall that in the anomalous model of the Yale Shooting Scenario the gun becomes unloaded after the performance of the wait action, an action which has no conventional effects—the unloading, then, is uncaused. In the context of a nonmonotonic logic—and without such a logic, the Yale Shooting Anomaly would not arise—it is very natural to formalize this by treating uncaused eventualities as abnormalities to be minimized.

This strategy was pursued by Hector Geffner in Geffner 1992 1990, where he formalizes this simple causal solution to the Yale Shooting Anomaly. But the solution is presented in the context of an ambitious general project in nonmonotonic logic that not only develops properties of the preferred model approach and shows how to apply it to a number of reasoning problems, but that relates nonmonotonic logic to probabilities, using ideas deriving from Adams 1975. In Geffner 1992, the causal theory is sketched; it is not developed to show its adequacy in dealing with the battery of problems presented above, and in particular the Ramification Problem is left untouched.

The work beginning with Lifschitz 1987 has contributed to a sustained line of research in the causal approach—not only by Lifschitz and students of his such as Enrico Giunchiglia and Hudson Turner, but by researchers at other sites. For work in this area, and further references, see Thielscher 1989, Gustaffson & Doherty 1996, Baral 1995, Nakashima et al. 1997, Lifschitz 1997, Giunchiglia & Lifschitz 1998, Lin 1995, Haugh 1987, Lifschitz 1998, Turner 1999, McCain & Turner 1995, Elkan 1991, McCain & Turner 1997, Thielscher 1996, and Gelfond & Lifschitz 1998.

Here, I will briefly describe some of theories developed by the Texas Action Group, leading up to the causal solution presented in Turner 1999. Turner returns to the ideas of Geffner 1992, but places them in a simpler logical setting and applies them to the formalization of more complex scenarios that illustrate the interactions of causal inertia with other considerations, especially the Ramification Problem.

Ramification is induced by the presence of static laws which relate the direct consequences of actions to other changes. I will use a car-starting scenario to illustrate the difficulties. There is one action, turn-on, which turns on the ignition; let's suppose that this action has no preconditions. There is a fluent Ig tracking whether the ignition is on, a fluent Dead tracking whether the battery is dead, and a fluent Run tracking whether the engine is running. A static law says that if the ignition is on and the battery isn't dead, the engine is running. (Let's suppose that every other source of failure has already been eliminated in this scenario; the only possible reason for not starting is the battery.) We want to consider a transition in which turn-on is performed in a situation in which the ignition is not on, the battery is not dead, and the car isn't running.

Of course, we want a planning agent to be able to infer in such a case that a performance of turn-on will result in a situation in which the ignition is on, the battery isn't dead, and the engine is running. But contraposition of laws makes it difficult to devise a principled solution. Informally, this difficulty is this: we can conclude by contraposing our only static law that if the ignition is on and the engine isn't running, then the battery is dead. This law not only is true in our scenario, but would be used to explain a failed attempt to start the car. But if we allow it to be used for prediction, then it is hard to see how to rule out an outcome of turn-on in which the ignition is on, the battery is dead, and the engine isn't running. The battery is dead in this outcome because of causal inertia. The engine isn't running because of the contraposed causal law.

I recommend Gelfond & Lifschitz 1998 to readers who want to explore in some detail the problems of embedding a nonmonotonic solution to the Frame Problem in relatively expressive action languages. This paper presents an increasingly powerful and sophisticated series of action languages. Their language incorporates an ad hoc or at least purely syntactic solution to the Ramification Problem.

A theory in language B consists of two sets of axioms:

Static laws of the form F?L, where L is a literal and F is a conjunction of literals or is T.
Dynamic laws of the form A Causes L if F, where A is an action term and F is a boolean formula.

Gelfond and Lifschitz impose a weak closure condition on static laws: where s is a set of literals, s is restricted-closed with respect to a B theory T, RBClT(s) if and only if every literal that would be added by starting with s and forward-chaining through the static laws of B is already in s. In other words:

RBClT(s) iff for all static laws F?L of T, if s entails F then L?s.

The closure operation induced by this condition is used in defining RESULT(a,s) for the language B. Of course, this solution treats logically equivalent static laws differently; for instance, a theory whose only static law is [P ? ¬Q] ?R will in general determine a very different RESULT function than one whose only static law is [P ? ¬R] ?Q.

This has some somewhat counterintuitive effects. In the car-starting scenario, Gelfond and Lifschitz’ language B indeed yields the desired conclusion that the car will start when there is one dynamic law,

turn-on Causes Ig if T,

when the only static law is

[Ig ? ¬Dead] ? Run,

when the initial state is s = {¬Ig, ¬Dead, ¬Run}, and when the action turn-on is performed.

However, if we add a true static law saying that if the ignition is on and the engine isn't running the battery is dead, we get an anomaly. With the addition of this law, there is a model in which preserving the fact that the car is not running makes the battery become dead when the ignition is turned on.

If the two static laws, [Ig ? ¬Dead] ? Run and [Ig ? ¬Run] ? Dead, are reformulated using causal language, the former sounds correct, while the second is distinctly problematic: compare

If the battery isn't dead, the ignition's being on will cause the engine to run.


If the engine isn't running, the ignition's being on will cause the battery to be dead.

This makes it very plausible to suppose that the source of the problem is a representation of underlying causal information in action language B that is somehow inadequate.

Gelfond and Lifschitz go on to describe another action language, C, which invokes an explicit notion of causality—motivated, I believe, in part by the need to provide a more principled solution to the problem. Instead of describing that language, I will discuss the similar theory of Turner 1999.

Turner's idea is to treat Caused as a modal operator [ c ], making this the basis of a modal nonmonotonic logic. In the preferred models of this logic, the caused propositions coincide with the propositions that are true, and this must be the only possibility consistent with the extensional part of the model. To make this more explicit, recall that in the possible worlds interpretation of S5, it is possible to identify possible worlds with state descriptions, which we can represent as sets I of literals (atomic formulas and their negations). Making this identification, then, we can think of a model as a pair , where S is a set of interpretations (complete, consistent sets of literals) including I. The modal operator [ c ] is given the standard semantics: where S is a set of interpretations and where I ? S, S ?I [ c ] A if and only if S ?I' A for all I' ?S. satisfies a set of formulas T if and only if S ?I A for all A ? T.

Turner's preferred models of T are the pairs such that: (1) satisfies T, (2) S = {I}, and (3) is the unique interpretation meeting conditions (1) and (2) with I' = I. Condition (2) guarantees the “universality of causation”; it validates A ? [ c ]A. Condition (3) “grounds” causality in noncausal information (in the models in which we are interested, this will be information about the occurrence of events), in the strongest sense: it is uniquely determined by this information.

Although it is not evident from the formulation, Turner's account of preferred models is related to the constructions of more general nonmonotonic logics, such as default logic. Consult Turner 1999 for details.

The axioms that specify the effects of actions treat these effects as caused; for instance, the axiom schema for loading would read as follows:

Causal-Load [ c ]Holds(load, RESULT(load, s))[33]

Ramifications of the immediate effects of actions are also treated as caused. And the nonmonotonic inertial axiom schemata take the form

[[[?c?]Holds(f,s)] ? Holds(f,RESULT(a,s))] ? [?c?]Holds(f,R ESULT(a,s))


[[[?c?]¬Holds(f,s)] ? ¬Holds(f,RESULT(a,s))] ? [?c?]¬Holds(f,RESULT(a,s)).

Thus, a true proposition can be caused either because it is the direct or indirect effect of an action, or because it involves the persistence of a caused proposition. Initial conditions are also considered to be caused, by stipulation.

To illustrate the workings of this approach, let's consider the simplest case of inertia: we have a language with just one constant denoting a fluent, f, and one action-denoting constant, wait. As in the Yale Shooting Problem, there are no axioms for wait; this action can always be performed and has no associated effects. Let s1 be RESULT(wait,s 0). The theory T contains an initial condition for f, Holds(f,s0) and a statement that the initial condition is caused, [?c?]?Holds(f,s 0), as well as the inertial schemata.

Two models of T satisfy conditions (1) and (2): M1 = and M2 = , where I1 = {Holds(f,s0), Holds(f,s1)} and I2 = {Holds(f,s0), ¬Holds(f,s1)}.

M1 is the intended model, in which nothing changes. It satisfies Condition (3), since if satisfies T it satisfies [?c?]Holds(f,s 1) by the inertial axiom

[[[?c?]Holds(f,s)] ? Holds(f,s1)] ? [?c?]Holds(f,s 1).

Therefore, S = {I1}.

M2 is an anomalous model, in which the fluent ceases spontaneously. This model does not satisfy Condition (3), since M3 = also satisfies T; in particular, it satisfies the inertial axiom for f because it fails to satisfy Holds(f,s1). So, while M1 is a preferred model, M2 is not.

Turner's approach avoids the problem of contraposition by giving causal relations the form


When contraposed, this becomes


which does not have the form of a causal law. This solution is less than fully satisfactory at solving the intuitive difficulties, because Turner's semantics seems intuitively to validate formulas such as

[?c?][[¬Dead ? Ig] ? Run],

and in this form, contraposition would be restored. The task of clarifying the foundations of causal theories of action and change may not yet be complete.

But the apparent usefulness of a “principle of universal causality” in accounting for a range of problems in qualitative common sense reasoning will be tantalizing to philosophers. And the causal theory, as initiated by Geffner and developed by Turner, has many interesting detailed features. For instance, while philosophical work on causality has concentrated on the causal relation, this work in logical AI shows that a great deal can be done by using only a nonrelational causal predicate.

Morgenstern's two chief criticisms of the causal approach to reasoning about actions are that it does not give an adequate account of explanation[34] and that the logical context in which it works (the Situation Calculus) is limited. As work on the approach continues, progress is being made in these areas. But the constraints that a successful logic of action and change must meet are so complex that it seems to be a reasonable research methodology to concentrate initially on a restricted logical setting.
4.7 Action Formalisms and Natural Language

Although for many AI logicists, the goal of action formalisms is to illuminate an important aspect of common sense reasoning, most of their research is uninformed by an important source of insights into the common sense view of time—namely, natural language. Linguists concerned with the semantics of temporal constructions in natural language, like the AI community, have begun with ideas from philosophical logic but have discovered that these ideas need to be modified in order to deal with the phenomena. A chief discovery of the AI logicists has been the importance of actions and their relation to change. Similarly, an important discovery of the “natural language logicists” has been the importance of different kinds of events (including structured composite events) in interpreting natural language. From work such as this the idea of “natural language metaphysics” (see, for instance, Bach 1989) has emerged.

The goal of articulating a logical framework tailored to a representational system that is motivated by systematic evidence about meanings in natural languages is not acknowledged by all linguistic semanticists. Nevertheless, it is a significant theme in the linguistic literature. This goal is remarkably similar to those of the common sense logicists, but the research methodology is entirely different.

Can the insights of these separate traditions be reconciled and unified? Is it possible to constrain theories of temporal representations and reasoning with the insights and research methodologies of both traditions? In Steedman 1995 (and 2000, listed in the Other Internet Resources Section), these important questions are addressed, and a theory is developed that extends action formalisms like the Situation Calculus, and that incorporates many of the insights from linguistic semantics. The project reported in Steedman 2000 is still incomplete, but the results reported there make a convincing case that the event-based ideas from linguistics can be fruitfully combined with the action-centered formalisms in the AI literature. The possibility of this unification is one of the most exciting logical developments in this area, bringing together as it does two independent descendants of the earlier work in the logic of time.
5. Causal Reasoning

In Section 4.6, we traced the reasons for the development of theories incorporating causality in work on reasoning about action and change. This is not the only area of AI in which causality has emerged. Causality figures in qualitative reasoning about devices; for Herbert Simon's important work in this area, which goes back to the 1950s, see Simon 1952; 1977; Iwasaki & Simon 1986. Both these traditions are important. But the most robust and highly developed program in AI relating to causality is that of Judea Pearl and his students and associates, which derives from the use of causal diagrams in the formalism for reasoning about probabilities known as Bayesian Belief Networks.

Pearl's program has developed into a far-reaching campaign to rehabilitate causality in statistical thinking. I will not discuss this topic here. For one thing, this survey omits probabilistic reasoning in AI. For another, Pearl's views on causality are systematically and comprehensively presented in a recent book-length study, Pearl 2000.

But I do wish to point out that the work on causality discussed in Section 4.6 and Pearl's ideas do share some common themes. On both approaches: action is central for causality. Also there is a focus on causality as a tool in reasoning that is necessitated in part by limited resources. Another important theme is the deployment and systematic study of formalisms in which causality is related to other constructs (in particular, to probability and to qualitative change) and a variety of realistic reasoning problems are addressed.

These commonalities provide reason to hope that we will see a science of causality emerging from the AI research, unifying the contributions of the probabilistic, the qualitative physics, and the nonmonotonic traditions, and illuminating the various phases of causal reasoning.

Whether you take causality to be a fundamental construct in natural science, or a fundamental common sense phenomenon, depends on whether you have in mind an idealized nature described by differential equations or you have in mind the view of nature we have to take in order to act, either in everyday situations, or for that matter in designing experiments in the laboratory. The fact that, as Bertrand Russell noted (see Russell 1957), causality is not to be found as a theoretical primitive in contemporary physical theories is at odds with its seeming importance in so many familiar areas of reasoning. The rigorous theories emerging in AI that are beginning to illuminate the workings of causality are important not only in themselves, but in their potentiality to illuminate wider philosophical issues.
6. Spatial Reasoning

The precomputational literature in philosophical logic relating to spatial reasoning is very sparse in relation, for instance, to the temporal literature. The need to support computational reasoning about space, however, in application areas such as motion planning and manipulation in physical space, the indexing and retrieval of images, geographic information systems, diagrammatic reasoning, and the design of high-level graphics programs has led to new interest in spatial representations and spatial reasoning. Of course, the geometrical tradition provides an exceptionally strong mathematical resource for this enterprise. But as in many other AI-related areas, it is not clear that the available mathematical theories are appropriate for informing these applications, and many computer scientists have felt it worthwhile to develop new foundations. Some of this work is closely related to the research in qualitative reasoning mentioned above in Section 2.2, and in some cases has been carried out by the same individuals.

The literature in spatial reasoning is extensive; for references to some areas not discussed here, see Stock 1997, Kapur & Mundy 1988, Hammer 1995, Wilson 1998, Osherson & Lasnik 1990, Renz & Nebel 1999, Yeap & Jeffries 1999, Forbus et al. 1991, Chen 1990, Burger & Bhanu 1992, Allwein & Barwise 1996, Glasgow et al. 1995, and Kosslyn 1990. Here, I will discuss only one trend, which is closely connected with parallel work in philosophical logic.

Qualitative approaches to space were introduced into the logical literature early in the twentieth century by Lesniewski; see Lesniewski 1916, which presents the idea of a mereology, or qualitative theory of the part-whole relation between physical individuals. This idea of a logical theory of relations among regions or the objects that occupy them, which does not depend on construing regions as sets of points, remained an active area of philosophical logic, even though it attracted relatively few researchers. More recent work in the philosophical literature, especially Casati & Varzi 1999, Simons 1987, Casati & Varzi 1996, Clarke 1981, Clarke 1985, as directly influential on current computational work.

The Regional Connection Calculus (RCC), developed by computer scientists at the University of Leeds, is based on a primitive C relating regions of space: the intended interpretation of C(x, y) is that the intersection of the closures of the values of x and y is nonempty. (See Cohn et al. 1997, Cohn 1996) for details and references.) One area of research concerns the definability of shapes in RCC. The extent of what can be defined with this simple primitive is surprising, but the technicalities quickly become complex; see, for instance, Gotts 1994, Gotts 1996). The work cited in Cohn et al. 1997 describes constraint propagation techniques and encodings in intuitionistic propositional logic as ways of supporting implemented reasoning based on RCC and some of its extensions. More recent work based on RCC addresses representation and reasoning about motion, which of course combines spatial and temporal issues; see Wolter & Zakharyaschev 2000). For more information about qualitative theories of movement, with references to other approaches, see Galton 1997.
7. Reasoning about Knowledge

Epistemic logic is another area in which strong influences from philosophical logic can be traced on logic in computer science. The classical source for epistemic logic is Hintikka 1962, in which Jaakko Hintikka showed that a modal approach to single-agent epistemic attitudes could be informative and rewarding. This work discusses at length the question of exactly which constraints are appropriate for knowledge and belief, when these attitudes are viewed as explicated by a model theoretic relation over possible worlds; in both cases, Hintikka argues for S4 type operators.

In several papers (including McCarthy 1979), John McCarthy has recommended an approach to formalizing knowledge that uses first-order logic, but that quantifies explicitly over such things as individual concepts. In this section I'll discuss the approach taken by most computer scientists, who, unlike McCarthy, use a modal language to formalize propositional attitudes.

The logical aspects of modal epistemic logic were not significantly developed after Hintikka's 1962 presentation; instead, the philosophical literature (which is not extensive, compared with many other topics in the area) concentrates on the issue of hyperintensionality, or closure of epistemic attitudes under logical consequence. This topic is especially challenging, turning out to be closely related to the semantic paradoxes, and the philosophical literature is inconclusive. Intuitions seem to conflict, and it is difficult to find ways to model the important phenomena using logical techniques.[35]

Fagin et al. 1984 begins a tradition in computational logic that revives the modal approach to epistemic logic, developing generalized logical foundations and applications that had not occurred to the philosophers. The technical idea is to simplify the modality, using S5 (or deontic S5 for belief), but to introduce multiple agents, and to concentrate on reasoning having to do with agents' attitudes about one another's attitudes. Such logics have direct applications in the analysis of distributed systems, dynamic systems in which change is effected by message actions, which change the knowledge of agents according to rules determined by a communications protocol.

As such, this work belongs to a separate area of computer science, but one that overlaps to some extent with AI. Later, this work has interacted with a research tradition in economics that is concerned with the role of knowledge in games and bargaining; see, for instance, Geanakopolos 1994; Osborne & Rubenstein 1994 [Chapter 5].

For some reason, the multi-agent case did not occur to philosophical logicians.[36] This is another example of the way in which need for an application (in this case, the need for a theory of distributed systems) provided the inspiration for an important logical development. I will not present details concerning the logic here, since they are extensively and systematically recorded in Fagin et al. 1995; this is essential reading for anyone seriously interested in this topic.

Much of the interdisciplinary work in applications of the logic of knowledge is reported in the proceedings of a series of conferences initiated in 1986 with Halpern 1986. These conferences record one of the most successful collaborations of philosophers with logicians in Computer Science, although the group of involved philosophers has been relatively small. The focus of the conferences has gradually shifted from Computer Science to Economics.

AI applications deal with with knowledge in the form of stored representations, and the tradition in AI with which we are concerned here thinks of reasoning as the manipulation of symbolic representations. Also, it is mainly due to AI that the problem of limited rationality has become a topic of serious interest, providing a counterbalance to the idealizations of philosophy and economics.[37] So you would think that a logical model of propositional attitudes that is committed to closure under logical consequence would be highly unpopular in AI. But this is not so; the possible worlds approach to attitudes is not only the leading theory in the areas discussed in Fagin et al. 1995, but has even been advocated in robotics applications; see Rosenschein & Kaelbling 1995; Rosenschein 1989. Nevertheless, the issue of hyperintensionality has been investigated in the AI literature; see Perlis 1985; Konolige 1986; Lakemeyer 1997; Levesque 1984). Though there are some new positive results here, the AI work in this area, in my opinion, has been as inconclusive as that in philosophy.

The philosophical literature on a related topic, the logic of perception, has not been extensive; the main reference is Hintikka 1970.[38] But sensation is addressed in recent work in the AI literature which is concerned with developing logical frameworks for general-purpose applications in robotics. The main idea in this area is to add sensing actions to the repertoire of a planning formalism of the sort discussed in Section 4. The earliest work in this area was carried out in the 1980s by Robert Moore; see Moore 1995b; Moore 1985. For some of the contemporary work in cognitive robotics, see Baral et al. 2000, Bacchus et al. 1999, Golden & Weld, 1996, Pirri & Finzi 1999, and Thielscher 2000.
8. Towards a Formalization of Common Sense

John McCarthy's explicit long-term goal—the formalization of common sense knowledge—has been adopted and pursued by a relatively small, but active, subcommunity of AI researchers. A larger group (those involved in knowledge representation, cognitive robotics, and qualitative physics) can be considered to work on specialized projects that support the larger goal. Anything like a formalization of common sense is so far from being accomplished that—if it is achievable at all—it is not even possible to estimate when we could expect the task to be completed. However, since 2001 (the date of a symposium on common sense reasoning held at The Courant Institute—see The Common Sense 2001 Homepage) something like a cooperative, sustained effort in this direction has begun to emerge. This effort is yielding a better sense of how to develop a workable methodology for formalizing common sense, and of how to divide the larger problem up into more manageable parts. Many of the papers presented at this conference were presented in expanded form in 2004 in an issue of Artificial Intelligence. The introduction to this issue, Davis & Morgenstern 2004, provides a useful survey and appreciation of the general research area.

This cooperative formalization effort (1) seeks to account for many areas of knowledge, and at the same time (2) attempts to see how this formalized knowledge can be brought to bear on moderately complex common-sense reasoning problems. The first book-length treatment of this topic, Davis 1991, divides the general problem into the following subtopics.

Quantities and Measurements
Plans and Goals

Several of these topics overlap with concerns of the qualitative physics and qualitative reasoning community. Although it can be hard to tell where common sense ends and physics begins, the formalization of common sense reasoning can be seen as a more general formalization project that can draw on a tradition in qualitative physics that has gone through many years of development and by now is fairly mature.[39]

And a few of them overlap with the work on the formalization of planning that was described above in Section 4. Minds and society, however, are new and different topics; the former has to do with common sense psychology and its application in introspective and interpersonal reasoning, and the latter, of course, should have to do with social and political knowledge and reasoning, but this is the least-developed area of formalized common sense knowledge: the chapter on this topic in Davis 1991 is very brief, and discusses mutual attitudes and communication. More recently, Andrew S. Gordon and Jerry Hobbs have undertaken a large-scale, ambitious formalization of common-sense psychology. See, for instance, Hobbs & Gordon 2005.

A more recent book-length treatment of the formalization of common sense, Mueller, 2006, follows a similar pattern. More than half of the book is devoted to reasoning about actions and change. There are short chapters on space and mental states, and a longer treatment of default reasoning. Although logical techniques and formalization methods take center stage in this book, it also contains material on nonlogical methods and on implementations related to the formalizations.

Even when attempted on a moderate scale, the formalization of common sense knowledge puts considerable pressure on the resources of even the most powerful logical systems that were devised for the formalization of mathematics. As we tried to show in discussing the special case of action and planning in Section 4, this pressure may lead us to seek logics that can facilitate the formalization projects: for instance, nonmonotomic logics and logics that explicitly represent context.

When larger-scale formalizations are attempted, other challenges arise that are similar to those that software engineering tries to address. Even fairly small programs and systems of axioms are difficult to comprehend and can be highly unpredicable, yielding unexpected consequences and unanticipated interactions. The creation and use of larger programs and formalizations raises questions of how to enable teams of developers to produce coherent results when modules are integrated, how to maintain and test large systems, and how to use knowledge sources such as dictionaries and knowledge bases to automatically generate axioms.

You can think of the philosophical methodology of providing analyses as a collection of attempts to formalize or partially formalize various common sense notions. These attempts are far smaller in scale, less systematic, and more heterogeneous than the parallel effort that is emerging in AI. Philosophers have never chosen a specific domain comparable to the planning domain and mounted a sustained attempt to formalize it, along with a companion effort to develop appropriate logics. And no matter how complex the notions with which they which they are concerned, philosophers have never allowed their analyses to grow to the complexity where methodological issues arise similar to those that apply to the development and maintenance of large software systems.

The techniques emerging in AI are of great potential significance for philosophy because it is easy to suspect that many philosophically important phenomena have the sort of complexity that can only be dealt with by accepting the problems that go along with developing complex formalizations. Limitations of the philosophical methods that were used throughout the twentieth century and are still in use may make it impossible to produce theories that do justice to the subject matter.

It would therefore be a great mistake for philosophers to disparage and ignore the large-scale formalizations that are beginning to emerge in AI because these efforts begin to raise engineering issues. It may well be that, although philosophy requires us to address complex phenomena in a rigorous way, the traditional philosophical methods are capable of doing justice to the complexity. Methods that promise to do this are worth taking seriously.

Among other methods borrowed from computer science, the common sense reasoning community has sought to develop suites of “benchmark problems”. The idea is to publicize problems that are difficult but not impossibly difficult, to encourage the community to create solutions, and compare the solutions.

Probably the best-studied problem to date is Ernest Davis' “egg-cracking problem.” This is formulated as follows on the Common Sense Reasoning Problem webpage.

A cook is cracking a raw egg against a glass bowl. Properly performed, the impact of the egg against the edge of the bowl will crack the eggshell in half. Holding the egg over the bowl, the cook will separate the two halves of the shell with his fingers, enlarging the crack, and the contents of the egg will fall gently into the bowl. The end result is that the entire contents of the egg will be in the bowl, with the egg unbroken, and that the two halves of the shell are in the cook's fingers.

Variants: What happens if: The cook brings the egg to impact very quickly? Very slowly? The cook lays the egg in the bowl and exerts steady pressure with his hand? The cook, having cracked the egg, attempts to peel it off its contents like a hard-boiled egg? The bowl is made of looseleaf paper? Of soft clay? The bowl is smaller than the egg? The bowl is upside down? The cook tries this procedure with a hard-boiled egg? With a coconut? With an M&M? [40]

Along with the problem itself three solutions are posted: Shanahan 2004, Lifschitz 1998b, and a version of Morgenstern 2001. Comparing the solutions is instructive: similarities outweigh differences. All the authors think of this as a planning problem, and use a versions of the Situation Calculus or the Event Calculus in the formalization. Each axiomatization is modular, with, for instance, separate modules devoted to the relevant geometrical and material properties. Each author provides a “proof of concept” for the formalization by showing that the axioms support a proof of the correctness of a plan to crack the egg in the simple case. None of the authors considers all of Davis' elaborations of the problem, but the axioms are framed with elaboration in mind and some elaborations are considered. It isn't clear whether any of the authors actually implemented the formalization (for instance, using a theorem prover).

The egg-cracking case raises the problem of how to evaluate moderately large formalizations of common sense problems. Morgenstern and Shanahan express this issue explicitly. Morgenstern suggests that the important criteria are (1) Epistemological adequacy (correspondence to intuitive reasoning, as experienced by people who engage in it), (2) Faithfulness to the real world, (3) Reusability, and (4) Elaboration tolerance. It isn't clear whether the first two of these criteria are too subjective to be useful. To these, Shanahan adds (5) Usability, which probably is presupposed by Morgenstern's third criterion.

There is anecdotal evidence that the larger AI community is somewhat skeptical about such research projects—or, if not skeptical, at least puzzled about how to evaluate them. In considering these doubts, it is necessary to appreciate the complexity of these formalization problems, and the preliminary and tentative status of the research program. Nevertheless, this criticism has some legitimacy, the common sense reasoning community is sensitive to these criticisms, and is working to develop and refine the methods and criteria for evaluating this work.

As long as formalization problems remain relatively simple, we can treat formalization as an art rather than as a discipline with a well-articulated methodology. But the trends I have been discussing show that formalization of even moderate-sized realistic common sense reasoning problems is not merely an art. Just as programming systems, expert systems and knowledge bases have created corresponding software engineering disciplines, large-scale formalization projects require a carefully thought through and tested methodology.
9. Logical Approaches to Natural Language and Communication

Over the last twenty-five years or so, many profound relations have emerged between logic and grammar. Computational linguistics (or natural language processing) is a branch of AI, and it is fairly natural to classify some of these developments under logic and AI. But many of them also belong to an independent tradition in logical foundations of linguistics; and in many cases it is hard (and pointless) to attempt a classification. This sketch will concentrate on developments that have to do with reasoning about linguistics; other applications of logic to linguistics are described in van Benthem & ter Meulen 1996.
9.1 Parsing and Deduction

Grammar formalisms—special-purpose systems for the description of linguistic systems and subsystems—can be thought of as logics designed to axiomatize the association of linguistic structures with strings of symbols. You might be able to infer from such a system, for instance, that ‘assignments’ is the plural form of the nominalization of the verb ‘assign’. So you can look at the process of parsing a string of words—of finding the linguistic structures, if any, that are associated with it—as a search for a proof in a certain logical system.

This approach has been highly successful as an analytic tool. It makes model-theoretic techniques applicable to linguistic reasoning, This makes the underlying reasoning problems much more transparent, and makes it possible to apply many well-developed areas of logic to grammar formalisms. For more information on these topics, see Buszkowski 1996; Shieber 1992.
9.2 Feature Structure Logic

The usefulness and scope of logical methods in relation to linguistics is greatly increased by the development of techniques for analyzing the way information attaches to linguistic units. It is very natural to represent the information attaching, say, to a lexical item in the form of a set of functions (or attributes) that produce values in some linguistic domain. A pronoun x may have a number, a person, and a case: if x = ‘we’ then

number(x) = plural,
person(x) = first,
case(x) = nominative.

In more general cases, the values of these functions may themselves be linguistic units that take on values for certain attributes.

Allowing these functions to be partial provides a useful informational representation of the stages of a linguistic parse; much of the work of parsing involves completing this partial information, subject to constraints imposed by linguistic agreement conditions. Feature structures—sets of identities that serve to evaluate linguistic features—have a natural algebraic interpretation, and there is an elegant treatment of their logic. For more information and references, see Rounds 1996.
9.3 Logic and Discourse

The reasoning associated with discourse is the probably the least well understood area of computational linguistics. Although logical techniques do not yet play a major role in discourse, they seem to offer one of the most promising ways of providing a uniform account of the many forms of reasoning that are involved in generating and interpreting language in interactive conversation.

I will briefly mention three contributions to this area. Building on the fact that the rules governing conversation are exception-ridden, Alex Lascarides and Nicholas Asher have developed techniques for formalizing discourse phenomena based on nonmonotonic logic; see Asher & Lascarides 1994, Asher & Lascarides 1997. Jerry Hobbs and various co-workers look at the inference processes used in discourse as abductive, and propose to formalize abduction as a search for a proof in which certain “low-cost” assumptions may be made which serve as data or additional axioms for the proof. Hobbs et al. 1993 shows how an impressive range of discourse phenomena can be formalized using this idea. In practice, this abductive account looks rather similar to that of Lascarides and Asher, because it involves deploying axioms about discourse (in the form of Horn clause rules supplemented with weights giving the assumption costs of premises) that in effect are nonmonotonic.

In more recent work, Matthew Stone shows in Stone 1998 how modal logic can inform the complex reasoning involved in natural language generation. Generating a coherent, appropriately phrased text that usefully performs a task-oriented communication task is difficult to formalize because it requires the integration of complex and sophisticated domain information with discourse planning, user modeling, and linguistic constraints. Stone shows that modal logic can be used to modularize the formalization of the information required in this task; he also shows how modal theorem proving can be used to implement the reasoning.
10. Taxonomic Representation and Reasoning
10.1 Concept-Based Classification

Traditionally, the task of representing large amounts of domain information for general-purpose reasoning has been one of the most important areas of knowledge representation. Systems that exploit the intuitive taxonomic organization of domains are useful for this purpose; taxonomic hierarchies not only help to organize the process of knowledge acquisition, but provide a useful connection to rule-based reasoning.[40]

For domains in which complex definitions are a natural way to organize information, knowledge engineering services based on definitions of concepts have been extremely successful. Like variable-free versions of first-order logic (see, for instance, Quine 1960, these systems are centered on concepts or first-order predicates, and provide a number of mechanisms for their definition. The fundamental algorithm associated with these taxonomic logics is a classifier which inputs a system of definitions and outputs the entailment relations between defined and primitive concepts. For background on these systems, see Woods & Schmolze 1992 and Brachman et al. 1991.

The simplest taxonomic logics can be regarded as subsystems of first-order logic with complex predicates; but they have been extended in many ways, and the issues raised by many of these extensions overlap in many cases with topics in philosophical logic.
10.2 Nonmonotonic Inheritance

Much more complex logical issues arise when the organization of a domain into hierarchies is allowed to have exceptions. One way to approach this topic is to explore how to make a taxonomic logic nonmonotonic; but nonmonotonic inheritance is a topic in its own right. Although there are strong affinities to nonmonotonic logic, nonmonotonic logic relies more heavily on graph-based representations than on traditional logical ideas, and seems to provide a much finer-grained approach to nonmonotonic reasoning that raises entirely new issues, and which quickly becomes problematic. For this reason, systems of nonmonotonic inheritance tend to be expressively weak, and their relations to the more powerful nonmonotonic logic has never been fully clarified. For background on this topic, see Thomason 1992, Horty 1994.
11. Contextual Reasoning

In the tradition in philosophical logic dealing with contextual effects on the interpretation of expressions, as well as in the more recent tradition in dynamic logic, context is primarily formalized as an assignment of values to variables, and the language is designed to make explicit reasoning about context either very limited or outright impossible.

Concern in AI about the representation of large and apparently heterogeneous domains and about the integration of disparate knowledge sources, as well as interests in formalizing common sense of the sort discussed in Section 2.2, above, have led to interest in the AI community in formalizing languages that take context into account more explicitly.

In McCarthy 1993b, McCarthy recommends the study of languages containing a construct


where ist is read “is-true.” This is analogous to the Holds construct of the situation calculus—but now c stands for a context, and f is a possibly complex propositional representation, which many (including McCarthy) take to refer to a sentence.

There are analogies here both to modal logic and to languages with an explicit truth predicate. But the applications that are envisioned for a logic of context create opportunities and problems that are in many ways new. For more about the logic of context, see McCarthy & Buvac 1998, Guha 1991, and the papers in Akman et al. 2001, Bouquet et al. 1999.
12. Prospects for a Logical Theory of Practical Reason

I believe there is reason to hope that the combination of logical methods with planning applications in AI can enable the development of a far more comprehensive and adequate theory of practical reasoning than has heretofore been possible. As with many problems having to do with common sense reasoning, the scale and complexity of the formalizations that are required are beyond the traditional techniques of philosophical logic. However, with computational methods of implementing and testing the formalizations and with areas such as cognitive robotics to serve as laboratories for developing and testing ideas, we can hope to radically advance a problem that has seen little progress since it was first proposed by Aristotle: the problem of devising a formalization of practical reasoning that is genuinely applicable to common sense reasoning problems.

The classical work in deontic logic that was begun by von Wright (see von Wright 1983) is one source of ideas; see (Horty 2001, van der Torre 1997). In fact, as the more recent work in deontic logic shows, nonmonotonic logic provides a natural and useful way to modify the classical deontic logic.

An even more robust account of practical reasoning begins to emerge when these ideas are supplemented with work on the foundations of planning and reasoning about action that were discussed in Section 4, above. But this development can be pursued even further, by extending the formalism to include preferences and intentions.[41]

Ultimately, what is needed is a model of an intelligent reasoning and acting agent. Developing such a model need not be entirely a matter of logic, but according to one school of thought, logic has a central role to play in it; see, for instance, Baral & Gelfond 2000, Wobcke et al. 1998, Rao & Georgeff 1991, Burkhard et al. 1998).

Adams, Ernest W., 1975, The Logic of Conditionals, Dordrecht: D. Reidel Publishing Co.
Aiello, Luigia Carlucci, Doyle, Jon, and Shapiro, Stuart (eds.), 1996, KR'96: Principles of Knowledge Representation and Reasoning, San Francisco: Morgan Kaufmann.
Akman, Varol, Bouquet, Paolo, Thomason, Richmond, and Young, Roger A. (eds.), 2001, Modeling and Using Context, Berlin: Springer-Verlag, ISBN 3-540-42379-6.
Alcourrón, Carlos E., 1995, “Defeasible logics: Demarcation and affinities”, in Conditionals: From Philosophy to Computer Science, Gabriella Crocco, Luis Fari nas del Cerro, and A. Herzig, eds., Oxford: Oxford University Press, 67-102.
Allen, James F., Fikes, Richard, and Sandewall, Erik (eds.), 1991, KR'91: Principles of Knowledge Representation and Reasoning, San Mateo, California: Morgan Kaufmann.
Allwein, Gerard and Barwise, Jon (eds.), 1996, Logical Reasoning With Diagrams, Oxford: Oxford University Press.
Amarel, Saul, 1968, “On representations of problems of reasoning about actions”, in Machine Intelligence 3, D. Mitchie, ed., Chichester, England: Ellis Horwood, 131-171.
Antoniou, Grigoris, 1997, Nonmonotonic Reasoning, Cambridge, Massachusetts: The MIT Press.
Arlo-Costa, Horacio and Shapiro, Scott, 1992, “Maps between nonmonotonic logic and conditional logic”, in KR'92. Principles of Knowledge Representation and Reasoning: Proceedings of the Third International Conference, Bernhard Nebel, Charles Rich, and William Swartout, eds., San Mateo, California: Morgan Kaufmann, 553-564.
Asher, Nicholas, 1995, “Commonsense entailment: A conditional logic for some generics”, in Conditionals: From Philosophy to Computer Science, Gabriella Crocco, Luis Fari nas del Cerro, and A. Herzig, eds., Oxford: Oxford University Press, 103-145.
Asher, Nicholas and Lascarides, Alex, 1994, “Intentions and information in discourse”, in Proceedings of the Thirty-Second Meeting of the Association for Computational Linguistics, James Pustejovsky, ed., Association for Computational Linguistics, San Francisco: Morgan Kaufmann, 35-41.
Asher, Nicholas and Lascarides, Alex, 1997, “Lexical disambiguation in a discourse context”, in Lexical Semantics: The Problem of Polysemy, James Pustejovsky and Brian Boguraev, eds., Oxford: Oxford University Press, 69-108.
Asher, Nicholas and Morreau, Michael, 1991, “Commonsense entailment: a modal theory of nonmonotonic reasoning”, in Proceedings of the Twelfth International Joint Conference on Artificial Intelligence, J. Mylopoulos and R. Reiter, eds., Los Altos, California: Morgan Kaufmann, 387-392.
Austin, John L., 1961, “A plea for excuses”, in Philosophical Papers, J.O. Urmson and G.J. Warnock, eds., Oxford: Oxford University Press.
Bacchus, Fahiem, Halpern, Joseph Y., and Levesque, Hector J., 1999, “Reasoning about noisy sensors and effectors in the situation calculus”, Artificial Intelligence, 111(1-2): 171-208.
Bach, Emmon, 1989, Informal Lectures on Formal Semantics, Albany, NY: State University of New York Press.
Baker, Andrew B., 1989, “A simple solution to the Yale shooting problem”, in KR'89: Principles of Knowledge Representation and Reasoning, Ronald J. Brachman, Hector J. Levesque, and Raymond Reiter, eds., San Mateo, California: Morgan Kaufmann, 11-20.
Baral, Chitta, 1995, “Reasoning about actions: Non-deterministic effects, constraints, and qualification”, in Proceedings of the Fourteenth International Joint Conference on Artificial Intelligence, Chris Mellish, ed., San Francisco: Morgan Kaufmann, 2017-2023.
Baral, Chitta and Gelfond, Michael, 2000, “Reasoning agents in dynamic domains”, in Logic-Based Artificial Intelligence, Jack Minker, ed., Dordrecht: Kluwer Academic Publishers, 257-279.
Baral, Chitta, McIlraith, Sheila, and San, Tran Cao, 2000, “Formulating diagnostic reasoning using an action language with narratives and sensing”, in KR2000: Principles of Knowledge Representation and Reasoning, Anthony G. Cohn, Fausto Giunchiglia, and Bart Selman, eds., San Francisco: Morgan Kaufmann, 311-322.
Belnap, Jr., Nuel D., 1996, “Agents in branching time”, in Logic and Reality: Essays on the Legacy of Arthur Prior, Jack Copeland, ed., Oxford: Oxford University Press, 239-271.
Benferat, Salem, Dubois, Didier, and Prade, Henri, 1997, “Nonmonotonic reasoning, conditional objects, and possibility theory”, Artificial Intelligence, 92(1-2): 259-276.
Besnard, Philippe, 1992, Default Logic, Berlin: Springer-Verlag.
Birnbaum, Lawrence, 1991, “Rigor mortis: A response to Nilsson's ‘Logic and artificial intelligence’”, Artificial Intelligence, 47(1-3): 57-77.
Boolos, George, 1993, The Logic of Provability, Cambridge, England: Cambridge Universoti Press.
Bouquet, Paolo, Serafini, Luigi, Brézillon, Patrick, Benerecetti, Massimo, and Castellani, Francesca (eds.), 1999, Modeling and Using Contexts: Proceedings of the Second International and Interdisciplinary Conference, CONTEXT'99, Berlin: Springer-Verlag.
Boutilier, Craig, 1992, “Conditional logics for default reasoning and belief revision”, Tech. Rep. KRR-TR-92-1, Computer Science Department, University of Toronto, Toronto, Ontario.
Boutilier, Craig, 1996, “Iterated revision and minimal change of conditional beliefs”, Journal of Philosophical Logic, 25(3): 263-305.
Boutilier, Craig, Dean, Thomas, and Hanks, Steve, 1996, “Planning under uncertainty: Structural assumptions and computational leverage”, in New Directions in AI Planning, Malik Ghallab and Alfredo Milani, eds., Amsterdam: IOS Press, 157-171.
Brachman, Ronald J., Levesque, Hector J., and Reiter, Raymond (eds.), 1989, KR'89: Principles of Knowledge Representation and Reasoning, San Mateo, California: Morgan Kaufmann.
Brachman, Ronald J., McGuinness, Deborah L., Patel-Schneider, Peter F., and Resnik, Lori A., 1991, “Living with CLASSIC: When and how to use a KL-ONE-like language”, in Principles of Semantic Networks, John F. Sowa, ed., San Mateo, California: Morgan Kaufmann, 401-456.
Brewka, Gerhard, 1991, Nonmonotonic Reasoning: Logical Foundations of Commonsense, Cambridge, England: Cambridge University Press.
Brewka, Gerhard, Dix, Jürgen, and Konolige, Kurt, 1997, Nonmonotonic Reasoning: An Overview, Stanford: CSLI Publications.
Buchanan, Bruce and Shortliffe, Edward H., 1984, Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project, Reading, Massachusetts: Addison Wesley.
Burger, Wilhelm and Bhanu, Bir, 1992, Qualitative Motion Planning, Dordrecht: Kluwer Academic Publishers, ISBN 0792392515.
Burgess, John P., 1984, “Basic tense logic”, in Handbook of Philosophical Logic, Volume II: Extensions of Classical Logic, Dov Gabbay and Franz Guenther, eds., Dordrecht: D. Reidel Publishing Co., 89-133.
Burkhard, Hans-Dieter, Hannebauer, Markus, and Wendler, Jan, 1998, “Belief-desire-intention deliberation in artificial soccer”, The AI Magazine, 1998(3): 87-93.
Buszkowski, Wojciech, 1996, “Mathematical linguistics and proof theory”, in Handbook of Logic and Language, Johan van Benthem and Alice ter Meulen, eds., Amsterdam: Elsevier, 683-736.
Carlson, Greg N. and Pelletier, Francis Jeffrey (eds.), 1995, The Generic Book, Chicago, IL: Chicago University Press.
Carnap, Rudolph, 1955, “Meaning and synonymy in natural languages”, Philosophical Studies, 7: 33-47. Reprinted in Carnap 1956, pp. 233-247.
Carnap, Rudolph, 1956, Meaning and Necessity, Chicago: Chicago University Press, 2 ed. (First edition published in 1947.).
Casati, Roberto and Varzi, Achille C., 1996, Holes and Other Superficialities, Cambridge, Massachusetts: The MIT Press, ISBN 0262032112.
Casati, Roberto and Varzi, Achille C., 1999, Parts and Places: The Structures of Spatial Representation, Cambridge, Massachusetts: The MIT Press, ISBN 026203266X (alk. paper).
Chellas, Brian, 1975, “Basic conditional logic”, Journal of Philosophical Logic, 4(2): 133-154.
Chen, Su-Shing (ed.), 1990, Advances in Spatial Reasoning, Volume 1, Norwood, New Jersey: Ablex.
Clancey, William J., 1983, “The epistemology of a rule-based expert system: a framework for explanation”, Artificial Intelligence, 20: 215-251.
Clark, Keith L., 1978, “Negation as failure”, in Logic and Data Bases, H. Gallaire and Jack Minker, eds., New York: Plenum Press, 293-322.
Clarke, Bowman L., 1981, “A calculus of individuals based on ‘connection’”, Notre Dame Journal of Formal Logic, 22(3): 204-218.
Clarke, Bowman L., 1985, “Individuals and points”, Notre Dame Journal of Formal Logic, 26(1): 61-75.
Cohen, Philip R. and Levesque, Hector J., 1990, “Intention is choice with commitment”, Artificial Intelligence, 42(3): 213-261.
Cohn, Anthony G., 1996, “Qualitative spatial representation and reasoning techniques”, in KI-97, Advances in Artificial Intelligence, Gerhard Brewka, Christopher Habel, and Bernhard Nebel, eds., Berin: Springer-Verlag, 1-30.
Cohn, Anthony G., Bennett, Brandon, Gooday, John, and Gotts, Nicholas M., 1997, “Representing and reasoning with qualitative spatial relations”, in Spatial and Temporal Reasoning, Oliviero Stock, ed., Dordrecht: Kluwer Academic Publishers, 97-134.
Cohn, Anthony G., Giunchiglia, Fausto, and Selman, Bart (eds.), 2000, KR2000: Principles of Knowledge Representation and Reasoning, San Francisco: Morgan Kaufmann.
Cohn, Anthony G., Schubert, Lenhart, and Shapiro, Stuart C. (eds.), 1998, KR'89: Principles of Knowledge Representation and Reasoning, San Francisco: Morgan Kaufmann.
Copeland, B. Jack, 1996, “Arthur Prior's life and legacy”, in Logic and Reality: Essays on the Legacy of Arthur Prior, Jack Copeland, ed., Oxford: Oxford University Press, 1-40.
Davis, Ernest, 1991, Common Sense Reasoning, San Francisco: Morgan Kaufmann.
Davis, Ernest and Morgenstern, Leora, 2004, “Introduction: Progress in formal commonsense reasoning”, Artificial Intelligence, 153(1-2): 1-12.
Davis, Martin, 1988, “Mathematical logic and the origin of modern computers”, in The Universal Turing Machine: A Half-Century Survey, Rolf Herkin, ed., Oxford: Oxford University Press, 149-174.
DeJong, Gerald D. and Bennett, Scott W., 1989, “Permissive planning: Extending classical planning to uncertain task domains”, Artificial Intelligence, 89(1-2): 173-217.
Delgrande, James P., 1998, “Conditional logics for defeasible logics”, in Handbook of Defeasible Reasoning and Uncertainty Management Systems, Volume 2, Dov M. Gabbay and Philippe Smets, eds., Dordrecht: Kluwer Academic Publishers, 135-174.
Dennett, Daniel, 1987, “Cognitive wheels: The frame problem of AI”, in The Robot's Dilemma: The Frame Problem in Artificial Intelligence, Zenon Pylyshyn, ed., Norwood, New Jersey: Ablex Publishing Co., 41-64.
Dowty, David R., 1977, “Toward a semantic analysis of verb aspect and the English ‘imperfective’ progressive”, Linguistics and Philosophy, 1(1): 45-77.
Doyle, Jon, 1979, “A truth maintenance system”, Artificial Intelligence, 12(3): 231-272.
Doyle, Jon, Sandewall, Erik, and Torasso, Pietro (eds.), 1994, KR'94: Principles of Knowledge Representation and Reasoning, San Francisco: Morgan Kaufmann.
Doyle, Jon and Thomason, Richmond H., 1999, “Background to qualitative decision theory”, AI Magazine, 20(2): 55-68.
Elkan, Charles, 1991, “Reasoning about action in first-order logic”, in Proceedings of the Conference of the Canadian Society for Computational Studies of Intelligence (CSCSI), Canadian Society for Computational Studies of Intelligence, San Francisco: Morgan Kaufman, 221-227.
Elkan, Charles, 1995, “On solving the qualification problem”, in Working Notes of the AAAI Spring Symposium on Extending Theories of Action: Formal Theories and Applications, Menlo Park, California: American Association for Artificial Intelligence.
Fagin, Ronald, Halpern, Joseph Y., Moses, Yoram, and Vardi, Moshe Y., 1995, Reasoning about Knowledge, Cambridge, Massachusetts: The MIT Press.
Fagin, Ronald, Halpern, Joseph Y., and Vardi, Moshe Y., 1984, “A model theoretic analysis of knowledge”, in Proceedings of the Twenty-Fifth Annual Symposium on the Foundations of Computer Science, 268-278.
Fensel, Dieter, Giunchiglia, Fausto, McGuinness, Deborah, and Williams, Mary-Anne (eds.), 2002, KR2002: Principles of Knowledge Representation and Reasoning, San Francisco, California: Morgan Kaufmann.
Fikes, Richard, 1996, “Ontologies: What are they, and where's the research?”, in KR'96: Principles of Knowledge Representation and Reasoning, Luigia Carlucci Aiello, Jon Doyle, and Stuart Shapiro, eds., San Francisco, California: Morgan Kaufmann, 652-654.
Finger, Jeffrey J., 1987, Exploiting Constraints in Design Synthesis, Ph.D. dissertation, Department of Computer Science, Stanford University, Stanford, California.
Fodor, Jerry A., 1987, “Modules, frames, fridgeons, sleeping dogs, and the music of the spheres”, in The Robot's Dilemma: The Frame Problem in Artificial Intelligence, Zenon Pylyshyn, ed., Norwood, New Jersey: Ablex Publishing Co., 139-149.
Forbus, Kenneth D., 1988, “Qualitative physics: Past, present, and future”, in Exploring Artificial Intelligence, Howard E. Shrobe, ed., San Mateo, California: Morgan Kaufmann, 239-296.
Forbus, Kenneth D., Nielsen, Paul, and Faltings, Boi, 1991, “Qualitative spatial reasoning: The CLOCK project”, Artificial Intelligence, 51(1-3): 417-471.
Ford, Kenneth M. and Pylyshyn, Zenon (eds.), 1996, The Robot's Dilemma Revisited: The Frame Problem in Artificial Intelligence, Norwood, New Jersey: Ablex Publishing Co., ISBN 1567501435 (pbk).
Gabbay, Dov, Hogger, Christopher J., and Robinson, J. A. (eds.), 1994, Handbook of Logic in Artificial Intelligence and Logic Programming, Volume 3: Nonmonotonic Reasoning and Uncertain Reasoning, Oxford: Oxford University Press.
Gabbay, Dov M., 1995, “Conditional implications and non-monotonic consequence”, in Conditionals: From Philosophy to Computer Science, Gabriella Crocco, Luis Fari nas del Cerro, and A. Herzig, eds., Oxford: Oxford University Press, 337-359.
Galton, Anthony, 1997, “Space, time, and movement”, in Spatial and Temporal Reasoning, Oliviero Stock, ed., Dordrecht: Kluwer Academic Publishers, 321-352.
Gärdenfors, Peter and Makinson, David, 1994, “Nonmonotonic inferences based on expectations”, Artificial Intelligence, 65(2): 197-245. Uses selection functions and “expectation relations” to model expectations. Relations between theory revision and nonmonotonic logic.
Geanakopolos, John, 1994, “Common knowledge”, in Handbook of Game Theory, Volume 2, Robert Aumann and S. Hart, eds., Amsterdam: Elsevier Science Publishers, chap. 40.
Geffner, Hector, 1990, “Causal theories of nonmonotonic reasoning”, in Proceedings of the Eighth National Conference on Artificial Intelligence, Thomas Dietterich and William Swartout, eds., American Association for Artificial Intelligence, Menlo Park, CA: AAAI Press, 524-530.
Geffner, Hector, 1992, Default Reasoning: Causal and Conditional Theories, Cambridge, Massachusetts: MIT Press.
Gelfond, Michael and Lifschitz, Vladimir, 1998, “Action languages”, Electronic Transactions on AI, 3. Available at
Genesereth, Michael and Nilsson, Nils J., 1987, Logical Foundations of Artificial Intelligence, San Mateo, California: Morgan Kaufmann.
Ginsberg, Matthew L. (ed.), 1987, Readings in Nonmonotonic Reasoning, Los Altos, California: Morgan Kaufmann. (Out of print.).
Giunchiglia, Enrico, Kartha, G. Neelakantan, and Lifschitz, Vladimir, 1997, “Representing action: Indeterminacy and ramifications”, Artificial Intelligence, 95(2): 409-438.
Giunchiglia, Enrico and Lifschitz, Vladimir, 1998, “An action language based on causal explanation”, in Proceedings of the Fourteenth National Conference on Artificial Intelligence and the Ninth Innovative Applications of Artificial Intelligence Conference, Ted Senator and Bruce Buchanan, eds., American Association for Artificial Intelligence, Menlo Park, California: AAAI Press, 623-628.
Glasgow, Janice, Narayanan, N. Hari, and Chandrasekaran, B. (eds.), 1995, Diagrammatic Reasoning, Cambridge, Massachusetts: The MIT Press.
Golden, Keith and Weld, Daniel, 1996, “Representing sensing actions: The middle ground revisited”, in KR'96: Principles of Knowledge Representation and Reasoning, Luigia Carlucci Aiello, Jon Doyle, and Stuart Shapiro, eds., San Francisco, California: Morgan Kaufmann, 174-185.
Goodman, Nelson, 1946, Fact, Fiction and Forecast, Harvard University Press, fourth ed.
Gotts, N.M., 1994, “How far can we ‘C’? defining a doughnut using connection alone”, in KR'94: Principles of Knowledge Representation and Reasoning, Jon Doyle, Erik Sandewall, and Pietro Torasso, eds., San Francisco, California: Morgan Kaufmann, 246-257.
Gotts, N.M., 1996, “Topology from a single primitive relation: Defining topological properties and relations in terms of connection”, Tech. Rep. 96.24, School of Computer Studies, University of Leeds, Leeds.
Guha, Ramanathan V., 1991, “Contexts: a formalization and some applications”, Tech. Rep. STAN-CS-91-1399, Stanford Computer Science Department, Stanford, California.
Gustaffson, Joakim and Doherty, Patrick, 1996, “Embracing occlusion in specifying the indirect effects of actions”, in KR'96: Principles of Knowledge Representation and Reasoning, Luigia Carlucci Aiello, Jon Doyle, and Stuart Shapiro, eds., San Francisco, California: Morgan Kaufmann, 87-98.
Halpern, Joseph Y. (ed.), 1986, Theoretical Aspects of Reasoning about Knowledge: Proceedings of the First Conference (TARK 1986), Los Altos, California: Morgan Kaufmann Publishers, Inc.
Halpern, Joseph Y. and Moses, Yoram, 1985, “Towards a theory of knowledge and ignorance”, in Logics and Models of Concurrent Systems, Krzysztof R. Apt, ed., Berlin: Springer-Verlag, 459-476.
Hammer, Eric M., 1995, Logic and Visual Information, Stanford, California: CSLI Publications.
Hanks, Steven and McDermott, Drew, 1985, “Temporal reasoning and default logics”, Tech. Rep. YALEU/CSD/RR#430, Department of Computer Science, Yale University, New Haven, Connecticut.
Hanks, Steven and McDermott, Drew, 1986, “Default reasoning, nonmonotonic logics and the frame problem”, in Proceedings of the Fifth National Conference on Artificial Intelligence, Tom Kehler and Stan Rosenschein, eds., American Association for Artificial Intelligence, Los Altos, California: Morgan Kaufmann, 328-333.
Hanks, Steven and McDermott, Drew, 1987, “Non-monotonic logics and temporal projection”, Artificial Intelligence, 33(3): 379-412.
Haugeland, John, 1981, “Semantic engines: An introduction to mind design”, in Mind Design, John Haugeland, ed., Cambridge, Massachusetts: The MIT Press, 1-34.
Haugh, Brian, 1987, “Simple causal minimization for temporal persistence and projection”, in Proceedings of the Seventh National Conference on Artificial Intelligence, Kenneth Forbus and Howard Shrobe, eds., American Association for Artificial Intelligence, Menlo Park, California: AAAI Press, 218-223.
Hintikka, Jaakko, 1962, Knowledge and Belief, Ithaca, New York: Cornell University Press.
Hintikka, Jaakko, 1970, “Objects of knowledge and belief: Aquaintances and public figures”, The Journal of Philosophy, 67(21): 869-883.
Hintikka, Jaakko, 1986, “Reasoning about knowledge in philosophy: The paradigm of epistemic logic”, in Theoretical Aspects of Reasoning about Knowledge: Proceedings of the First Conference, Joseph Y. Halpern, ed., Los Altos, California: Morgan Kaufmann Publishers, Inc., 63-80.
Hobbs, Jerry, et al., Mark, Appelt, Douglas, and Martin, Paul, 1993, “Interpretation as abduction”, Artificial Intelligence, 63(1-2): 69-142.
Hobbs and Gordon, Hobbs, Jerry and Gordon, Andrew (2005) Encoding Knowledge of Commonsense Psychology. 7th International Symposium on Logical Formalizations of Commonsense Reasoning. May 22-24, 2005, Corfu, Greece.
Horty, John F., 1994, “Some direct theories of nonmonotonic inheritance”, in Handbook of Logic in Artificial Intelligence and Logic Programming, Volume 3: Nonmonotonic Reasoning and Uncertain Reasoning, Dov Gabbay, Christopher J. Hogger, and J. A. Robinson, eds., Oxford University Press, 111-187.
Horty, John F., 2001, Agency and Deontic Logic, Oxford: Oxford University Press, ISBN 0195134613.
Israel, David J., 1991, “A short sketch of the life and career of john McCarthy”, in Artificial Intelligence and Mathematical Theory of Computation: Papers in Honor of John McCarthy, Vladimir Lifschitz, ed., San Diego, California: Academic Press.
Iwasaki, Yumi and Simon, Herbert, 1986, “Causality in device behavior”, Artificial Intelligence, 29(1): 3-32.
Johnston, Benjamin and Williams, Mary-Anne, “A Generic Framework for Approximate Simulation in Commonsense Reasoning Systems”, AAAI 2007 Spring Symposium on Commonsense Reasoning, American Association for Artificial Intelligence, Menlo Park, 2007.
Kapur, Deepak and Mundy, Joseph L., 1988, “Geometric reasoning and artificial intelligence: Introduction to the special volume”, Artificial Intelligence, 37(1-3): 1-11.
Konolige, Kurt, 1986, “What awareness isn't: A sentential view of implicit and explicit belief”, in Theoretical Aspects of Reasoning about Knowledge: Proceedings of the First Conference, Joseph Y. Halpern, ed., Los Altos, California: Morgan Kaufmann Publishers, Inc., 241-250.
Konolige, Kurt, 1988, “On the relation between default and autoepistemic logic”, Artificial Intelligence, 35(3): 343-382. (See also errata, Artificial Intelligence 41(1): 115.).
Konolige, Kurt, 1994, “Autoepistemic logic”, in Handbook of Logic in Artificial Intelligence and Logic Programming, Volume 3: Nonmonotonic Reasoning and Uncertain Reasoning, Dov Gabbay, Christopher J. Hogger, and J. A. Robinson, eds., Oxford: Oxford University Press, 217-295.
Konolige, Kurt and Pollack, Martha, 1993, “A representationalist theory of intention”, in Proceedings of the Thirteenth International Joint Conference on Artificial Intelligence, Ruzena Bajcsy, ed., San Mateo, California: Morgan Kaufmann.
Kosslyn, Stephen Michael, 1990, “Visual cognition: Introduction”, in An Invitation to Cognitive Science. Volume 2: Visual Cognition and Action, Daniel N. Osherson and Howard Lasnik, eds., Cambridge, Massachusetts: The MIT Press, 3-4.
Kowalski, Robert A. and Sergot, Marek J., 1986, “A logic-based calculus of events”, New Generation Computing, 4: 67-95.
Krifka, Manfred, Pelletier, Francis Jeffrey, Carlson, Gregory, ter Meulen, Alice, Chierchia, Gennaro, and Link, Godehard, 1995, “Genericity: An introduction”, in The Generic Book, Gregory Carlson and Francis Jeffrey Pelletier, eds., Chicago, IL: Chicago University Press, 1-124.
Kuipers, Benjamin, 1993, “Reasoning with qualitative models”, Artificial Intelligence, 59(1-2): 125-132.
Lakemeyer, Gerhard, 1997, “Limited reasoning in first-order knowledge bases”, Artificial Intelligence, 71(2): 213-255.
Lascarides, Alex, 1992, “The progressive and the imperfective paradox”, Synthese, 87(6): 401-447.
Laux, Armin and Wansing, Heinrich (eds.), 1995, Knowledge and Belief in Philosophy and Artificial Intelligence, Berlin: Akedemie Verlag.
Lehmann, Daniel and Magidor, Menachem, 1992, “What does a conditional knowledge base entail?”, Artificial intelligence, 55(1): 1-60.
Lenat, Douglas B. and Guha, R.V., 1989, Building Large Knowledge-Based Systems: Representation and Inference in the CYC Project., Reading, Massachusetts: Addison-Wesley Publishing Company.
Lesniewski, Stanisaw, 1916, “Podstawy ogólnej teorii mnogosci I”, English Title: “Foundations of a general set theory I.”.
Levesque, Hector and Lakemeyer, Gerhard, 2000, The Logic of Knowledge Bases, Cambridge, Massachusetts: The MIT Press, ISBN 0-262-12232-4 (hardback).
Levesque, Hector J., 1984, “A logic of implicit and explicit belief”, in Proceedings of the Fourth National Conference on Artificial Intelligence, American Association for Artificial Intelligence, 198-202.
Levesque, Hector J., 1987, “Taking issue: Guest editor's introduction”, Computational Intelligence, 3(3): 149-150.
Levy, Alon Y., 2000, “Logic-based techniques in data integration”, in Logic-Based Artificial Intelligence, Jack Minker, ed., Dordrecht: Kluwer Academic Publishers, 575-595.
Lifschitz, Vladimir, 1987, “Formal theories of action: Preliminary report”, in Proceedings of the Tenth International Joint Conference on Artificial Intelligence, John McDermott, ed., Los Altos, California: Morgan Kaufmann.
Lifschitz, Vladimir (ed.), 1990a, Formalizing Common Sense: Papers by John McCarthy, Norwood, New Jersey: Ablex Publishing Corporation.
Lifschitz, Vladimir, 1990b, “Understanding common sense: McCarthy's research in artificial intelligence”, in Formalizing Common Sense: Papers by John McCarthy, Vladimir Lifschitz, ed., Norwood, New Jersey: Ablex Publishing Corporation, 1-8.
Lifschitz, Vladimir, 1997, “On the logic of causal explanation”, Artificial Intelligence, 96(2): 451-465.
Lifschitz, Vladimir, 1998, “Situation calculus and causal logic”, in KR'98: Principles of Knowledge Representation and Reasoning, Anthony G. Cohn, Lenhart Schubert, and Stuart C. Shapiro, eds., San Francisco, California: Morgan Kaufmann, 536-546.
Lin, Fangzhen, 1995, “Embracing causality in specifying the indirect effects of actions”, in Proceedings of the Fourteenth International Joint Conference on Artificial Intelligence, Chris Mellish, ed., San Francisco: Morgan Kaufmann, 1985-1991.
Lormand, Eric, 1996, “The holorobophobe's dilemma”, in The Robot's Dilemma Revisited: The Frame Problem in Artificial Intelligence, Kenneth M. Ford and Zenon Pylyshyn, eds., Norwood, New Jersey: Ablex Publishing Co., 61-88.
Lukasiewicz, Jan, 1970, Jan Lukasiewicz, Selected Writings, North-Holland. Edited by Ludwik Borowski.
Lukaszewicz, Witold, 1990, Non-Monotonic Reasoning: Formalization of Commonsense Reasoning, New York: Ellis Horwood, ISBN 0136244467.
Marek, Victor and Truszczynski, Mirosaw, 1991, “Autoepistemic logic”, Journal of the Association for Computing Machinery, 38(3): 588-619.
Marek, Wictor and Truszczynski, Mirosaw, 1989, “Relating autoepistemic and default logics”, in KR'89: Principles of Knowledge Representation and Reasoning, Ronald J. Brachman, Hector J. Levesque, and Raymond Reiter, eds., San Mateo, California: Morgan Kaufmann, 276-288.
Marek, Wictor and Truszczynski, Mirosaw, 1994, Nonmonotonic Logic: Context-Dependent Reasoning, Berlin: Springer-Verlag.
McCain, Norman and Turner, Hudson, 1995, “A causal theory of ramifications and qualifications”, in Proceedings of the Fourteenth International Joint Conference on Artificial Intelligence, Chris Mellish, ed., San Francisco: Morgan Kaufmann, 1978-1984.
McCain, Norman and Turner, Hudson, 1997, “Causal theories of action and change”, in Proceedings of the Thirteenth National Conference on Artificial Intelligence and the Eighth Innovative Applications of Artificial Intelligence Conference, Howard Shrobe and Ted Senator, eds., American Association for Artificial Intelligence, Menlo Park, California: AAAI Press, 460-465.
McCarthy, John, 1959, “Programs with common sense”, in Proceedings of the Teddington Conference on the Mechanization of Thought Processes, London: Her Majesty's Stationary Office, 75-91.
McCarthy, John, 1979, “First order theories of individual concepts and propositions”, in Machine Intelligence 9, J.E. Hayes, D. Mitchie, and L.I. Mikulich, eds., Chichester, England: Ellis Horwood, 129-148.
McCarthy, John, 1980, “Circumscription: A form of non-monotonic reasoning”, Artificial Intelligence, 13: 27-39.
McCarthy, John, 1983, “Situations, actions, and causal laws”, Tech. Rep. Memo 2, Stanford Artificial Intelligence Project, Stanford University.
McCarthy, John, 1986, “Applications of circumscription to formalizing common sense knowledge”, Artificial Intelligence, 13: 27-39.
McCarthy, John, 1987, “Epistemological problems of artificial intelligence”, in Readings in Nonmonotonic Reasoning, Matthew L. Ginsberg, ed., Los Altos, California: Morgan Kaufmann, 46-55.
McCarthy, John, 1993a, “History of circumscription”, Artificial Intelligence, 59: 23-26.
McCarthy, John, 1993b, “Notes on formalizing contexts”, in Proceedings of the Thirteenth International Joint Conference on Artificial Intelligence, Ruzena Bajcsy, ed., San Mateo, California: Morgan Kaufmann, 555-560.
McCarthy, John, 1999, “Elaboration tolerance”, Http://
McCarthy, John and Buvac, Saša, 1998, “Formalizing context (expanded notes)”, in Computing Natural Language, Atocha Aliseda, Rob van Glabbeek, and Dag Westerståhl, eds., Stanford, California: CSLI Publications, 13-50.
McCarthy, John and Hayes, Patrick J., 1969, “Some philosophical problems from the standpoint of artificial intelligence”, in Machine Intelligence 4, B. Meltzer and D. Michie, eds., Edinburgh: Edinburgh University Press, 463-502.
McDermott, Drew, 1982, “Nonmonotonic logic II: Nonmonotonic modal theories”, Journal of the Association for Computing Machinery, 29(1): 33-57.
McDermott, Drew, 1987, “Critique of pure reason”, Computational Intelligence, 3(3): 149-160.
McDermott, Drew and Doyle, Jon, 1980, “Non-monotonic logic I”, Artificial Intelligence, 13: 41-72.
Meyer, John-Jules Ch. and van der Hoek, Wiebe, 1995, Epistemic Logic for AI and Computer Science, Cambridge: Cambridge University Press, ISBN 052146014X (hardback).
Minker, Jack, 1997, “Logic and databases: Past, present and future”, AI Magazine, 18(3): 21-47.
Minker, Jack, 2000a, “Introduction to logic-based artificial intelligence”, in Logic-Based Artificial Intelligence, Jack Minker, ed., Dordrecht: Kluwer Academic Publishers, 3-33.
Minker, Jack (ed.), 2000, Logic-Based Artificial Intelligence, Dordrecht: Kluwer Academic Publishers.
Minsky, Marvin, 1974, “A framework for representing knowledge”, Tech. Rep. 306, Artificial Intelligence Laboratory, MIT. Republished in several places, including Haugeland 1981.
Moore, Johanna, 1995, Participating in Explanatory Dialogues, The MIT Press.
Moore, Robert C., 1985, “A formal theory of knowledge and action”, in Formal Theories of the Commonsense World, Jerry R. Hobbs and Robert C. Moore, eds., Norwood, New Jersey: Ablex Publishing Corporation, 319-358.
Moore, Robert C., 1993, “Autoepistemic logic revisited”, Artificial Intelligence, 59(1-2): 27-30.
Moore, Robert C., 1995b, Logic and Representation, Cambridge, England: Cambridge University Press.
Morgenstern, Leora, 1996, “The problem with solutions to the frame problem”, in The Robot's Dilemma Revisited: The Frame Problem in Artificial Intelligence, Kenneth M. Ford and Zenon Pylyshyn, eds., Norwood, New Jersey: Ablex Publishing Co., 99-133.
Morgenstern, Leora, 2001, “Mid-Sized Axiomatizations of Commonsense Problems: A Case Study in Egg Cracking”, Studia Logica, 67(3):333-384.
Morgenstern, Leora and Stein, Lynn, 1994, “Motivated action theory: a formal theory of causal reasoning”, Artificial Intelligence, 71(1): 1-42.
Mueller, Erik T., 2006, Common Sense Reasoning, Elsevier.
Nakashima, Hideyuki, Matsubara, Hitoshi, and Osawa, Ichiro, 1997, “Causality as a key to the frame problem”, Artificial Intelligence, 91(1): 37-50.
Nebel, Bernhard, Rich, Charles, and Swartout, William (eds.), 1992, KR’: Principles of Knowledge Representation and Reasoning, San Francisco: Morgan Kaufmann.
Nilsson, Nils J., 1991, “Logic and artificial intelligence”, Artificial Intelligence, 47(1-3): 31-56.
Nilsson, Nils J., 1995, “Eye on the prize”, Available at http://robotics.Stanford.EDU/~nilsson/.
Ohrstrom, Peter and Hasle, Per F.V., 1995, Temporal Logic from Ancient Ideas to Artificial Intelligence, Dordrecht: Kluwer Academic Publishers.
Osborne, Martin J. and Rubenstein, Ariel, 1994, A Course in Game Theory, Cambridge, Massachusetts: The MIT Press.
Osherson, Daniel N. and Lasnik, Howard (eds.), 1990, An Invitation to Cognitive Science. Volume 2: Visual Cognition and Action, Cambridge, Massachusetts: The MIT Press.
Pearl, Judea, 1994, “From Adams' conditionals to default expressions, causal conditionals, and counterfactuals”, in Probability and Conditionals: Belief Revision and Rational Decision, Ellery Eells and Brian Skyrms, eds., Cambridge, England: Cambridge University Press, 47-74.
Pearl, Judea, 2000, Causality: Models, Reasoning, and Inference, Cambridge, England: Cambridge University Press, ISBN 0-521-77362-8.
Perlis, Donald, 1985, “Languages with self-reference I: Foundations”, Artificial Intelligence, 25: 301-322.
Pirri, Fiora and Finzi, Alberto, 1999, “A preliminary approach to perception in theory of agents”, in Proceedings of the IJCAI-99 Workshop on Nonmonotonic Reasoning, Action and Change, Michael Thielscher, ed., IJCAI, Murray Hill, New Jersey: International Joint Conference on Artificial Intelligence, 49-56.
Pollack, Martha, 1992, “The uses of plans”, Artificial Intelligence, 57(1): 43-68.
Pollock, John L., 1995, Cognitive Carpentry: A Manual for How to Build a Person, Cambridge, Massachusetts: The MIT Press.
Prior, Arthur, 1956, Time and Modality, Oxford: Oxford University Press.
Prior, Arthur, 1967, Past, Present and Future, Oxford: Oxford University Press.
Prior, Arthur, 1968, Papers on Time and Tense, Oxford: Oxford University Press.
Pylyshyn, Zenon (ed.), 1987, The Robot's Dilemma: The Frame Problem in Artificial Intelligence, Norwood, New Jersey: Ablex Publishing Co., ISBN 0893913715.
Quine, Willard V.O, 1960, “Variables explained away”, in Selected Logic Papers, Willard V. Quine, ed., Cambridge, Massachusetts: Harvard University Press, 227-235.
Rao, Anand S. and Georgeff, Michael P., 1991, “Modeling rational agents within a BDI-architecture”, in KR'91: Principles of Knowledge Representation and Reasoning, James F. Allen, Richard Fikes, and Erik Sandewall, eds., San Mateo, California: Morgan Kaufmann, 473-484.
Reiter, Raymond, 1978, “On closed world data bases”, in Logic and Data Bases, H. Gallaire and J. Minker, eds., New York: Plenum Press, 55-76.
Reiter, Raymond, 1980, “A logic for default reasoning”, Artificial Intelligence, 13: 81-32.
Reiter, Raymond, 1993, “Proving properties of states in the situation calculus”, Artificial Intelligence, 64: 337-351.
Reiter, Raymond, 2001, Knowledge in Action: Logical Foundations for Specifying and Implementing Dynamical Systems, Cambridge, Massachusetts: The MIT Press.
Renz, Jochen and Nebel, Bernhard, 1999, “On the complexity of qualitative spatial reasoning: A maximal tractable fragment of the region connection calculus”, Artificial Intelligence, 108(1-2): 69-123.
Rosenschein, Stanley J., 1989, “Synthesizing information-tracking automata from environment descriptions”, in KR'89: Principles of Knowledge Representation and Reasoning, Ronald J. Brachman, Hector J. Levesque, and Raymond Reiter, eds., San Mateo, California: Morgan Kaufmann, 386-393.
Rosenschein, Stanley J. and Kaelbling, Leslie Pack, 1995, “A situated view of representation and control”, Artificial Intelligence, 73(1-2): 149-173.
Rounds, William C., 1996, “Feature logics”, in Handbook of Logic and Language, Johan van Benthem and Alice ter Meulen, eds., Amsterdam: Elsevier, 475-533.
Russell, Bertrand, 1957, Mysticism and Logic, Garden City, New York: Doubleday.
Russell, Stuart and Norvig, Peter, 2003, Artificial Intelligence: A Modern Approach, Englewood Cliffs, New Jersey: Prentice Hall, 2 ed., ISBN 0137903952.
Russell, Stuart J. and Wefald, Eric, 1991, Do the Right Thing, Cambridge, Massachusetts: The MIT Press.
Sadek, M.D., 1992, “A study in the logic of intention”, in KR'92. Principles of Knowledge Representation and Reasoning: Proceedings of the Third International Conference, Bernhard Nebel, Charles Rich, and William Swartout, eds., San Mateo, California: Morgan Kaufmann, 462-473.
Sandewall, Eric, 1972, “An approach to the frame problem, and its implementation”, in Machine Intelligence 7, D. Michie and B. Meltzer, eds., Edinburgh University Press, 195-204.
Sandewall, Erik, 1994, Features and Fluents: A Systematic Approach to the Representation of Knowledge About Dynamical Systems, Oxford: Oxford University Press.
Schlechta, Karl, 1997, Nonmonotonic Logics, Berin: Springer-Verlag, ISBN 3-540-62482-1.
Schubert, Lenhart, 1990, “Monotonic solution of the frame problem in the situation calculus; an efficient method for worlds with fully specified actions”, in Knowledge Representation and Defeasible Reasoning, Henry Kyburg, Ronald Loui, and Greg Carlson, eds., Dordrecht: Kluwer Academic Publishers, 23-67.
Seligman, Jerry and Moss, Lawrence S., 1996, “Situation theory”, in Handbook of Logic and Language, Johan van Benthem and Alice ter Meulen, eds., Amsterdam: Elsevier, 239-307.
Shanahan, Murray, 1997, Solving the Frame Problem, Cambridge, Massachusetts: The MIT Press.
Shanahan, Murray, 2004, “An attempt to formalise a non-trivial benchmark problem in common sense reasoning”, Artificial Intelligence, 153(1-2): 141-165.
Shieber, Stuart, 1992, Constraint-Based Grammar Formalisms: Parsing and Type Inference for Natural and Computer Languages, Cambridge, Massachusetts: The MIT Press.
Shoham, Yoav, 1988, Reasoning About Change: Time and Causation From the Standpoint of Artificial Intelligence, Cambridge, Massachusetts: The MIT Press.
Simon, Herbert, 1952, “On the definition of the causal relation”, The Journal of Philosophy, 49: 517-528.
Simon, Herbert, 1966, “On reasoning about action”, Tech. Rep. Complex Information Processing Paper #87, Carnegie Institute of Technology, Pittsburgh, Pennsylvania.
Simon, Herbert A., 1977, Models of Discovery, Dordrecht: D. Reidel Publishing Co.
Simon, Herbert A., 1982a, Models of Bounded Rationality, Volume 1, Cambridge, Massachusetts: The MIT Press.
Simon, Herbert A., 1982b, Models of Bounded Rationality, Volume 2, Cambridge, Massachusetts: The MIT Press.
Simons, Peter, 1987, Parts: A Study in Ontology, Oxford: Oxford University Press.
Stalnaker, Robert C., 1993, “A note on non-monotonic modal logic”, Artificial Intelligence, 64(2): 183-196. Widely circulated in manuscipt form, 1980 to 1992.
Steedman, Mark, 1995, “Dynamic semantics for tense and aspect”, in Proceedings of the Fourteenth International Joint Conference on Artificial Intelligence, Chris Mellish, ed., San Francisco: Morgan Kaufmann, 1292-1298.
Stefik, Mark J., 1995, An Introduction to Knowledge Systems, San Francisco: Morgan Kaufmann.
Stock, Oliviero (ed.), 1997, Spatial and Temporal Reasoning, Dordrecht: Kluwer Academic Publishers, ISBN 0792346440 (paper).
Stone, Matthew, 1998, Modality in Dialogue: Planning, Pragmatics and Computation, Ph.D. dissertation, Computer Science Department, University of Pennsylvania, Philadelphia, Pennsylvania.
Thielscher, Michael, 1989, “Ramification and causality”, Artificial Intelligence, 89(1-2): 317-364.
Thielscher, Michael, 1996, “Causality and the qualification problem”, in KR'96: Principles of Knowledge Representation and Reasoning, Luigia Carlucci Aiello, Jon Doyle, and Stuart Shapiro, eds., San Francisco, California: Morgan Kaufmann, 51-62.
Thielscher, Michael, 2000, “Representing the knowledge of a robot”, in KR2000: Principles of Knowledge Representation and Reasoning, Anthony G. Cohn, Fausto Giunchiglia, and Bart Selman, eds., San Francisco: Morgan Kaufmann, 109-120.
Thomason, Richmond H., 1992, “NETL and subsequent path-based inheritance theories”, in Semantic Networks in Artificial Intelligence, Fritz Lehmann, ed., Oxford: Pergamon Press, 179-204.
Turner, Hudson, 1999, “A logic of universal causation”, Artificial Intelligence, 113(1-2): 87-123.
van Benthem, Johan, 1983, The Logic of Time, Dordrecht: D. Reidel Publishing Company.
van Benthem, Johan and ter Meulen, Alice, 1996, Handbook of Logic and Language, Amsterdam: Elsevier Science Publishers.
van der Torre, Leendert W.N., 1997, Reasoning about Obligations: Defeasibility in Preference-Based Deontic Logic, Amsterdam: Thesis Publishers.
von Wright, Georg Henrik, 1983, Practical Reason: Philosophical Papers, Volume 1, Ithaca: Cornell University Press, ISBN 0801416736.
Weld, Daniel S. and de Kleer, Johan (eds.), 1990, Qualitative Reasoning about Physical Systems, San Mateo, California: Morgan Kaufmann.
Wilson, Randall H., 1998, “Geometric reasoning about assembly tools”, Artificial Intelligence, 98(1-2): 237-279.
Wobcke, Wayne, Pagnucco, Maurice, and Zhang, C. (eds.), 1998, Agents and Multi-Agent Systems—Formalisms, Methodologies, and Applications, Berlin: Springer-Verlag.
Wolter, Frank and Zakharyaschev, Michael, 2000, “Spatio-temporal representation and reasoning based on RCC-8”, in KR2000: Principles of Knowledge Representation and Reasoning, Anthony G. Cohn, Fausto Giunchiglia, and Bart Selman, eds., San Francisco: Morgan Kaufmann, 3-14.
Woods, William A. and Schmolze, James G., 1992, “The KL-ONE family”, in Semantic Networks in Artificial Intelligence, Fritz Lehmann, ed., Oxford: Pergamon Press, 133-177.
Yeap, Wai K. and Jeffries, Margaret E., 1999, “Computing a representation of the local environment”, Artificial Intelligence, 107(2): 265-301.

Other Internet Resources

Davis, Ernest, Home Page for Common Sense 2001.
Leora Morgenstern, Common Sense Problem Page.
Forbus, Kenneth, 1996, “Qualitative reasoning”. CRC Hand-book of Computer Science and Engineering. CRC Press.
(Next-to final drafts, with more citations than final version.) Available in PDF.
Steedman, Mark 2000, “The productions of time”. Available online in PDF.

Related Entries

conditionals | frame problem | reasoning: automated

I am grateful to John McCarthy, who read a draft of this article and provided extensive and helpful comments.
Copyright © 2008 by
Richmond Thomason


...I AM CONVINCED that most of humans and the WORLD, including governments, Academia and so called "think tanks" ahve reached a "limit in their logic" bur not a limit in logic as such !
WHY ? Because my Book 1 : BOOK OF PURE LOGIC, establishes basis for better and more progress in logic, which they have shown themselves incapable of accepting, much less perceiving their limits in their logic !

AS TO PROGRESS IN ARTIFICIAL INTELLIGENCE and the big hard question of MIND and MATER, it is clear there are too many AMBIGUITIES ! And How do you make a machine AMBIGUIS "ENOUGH" ! The UNIVERSE of possibilities is too great !And the Great Universe of the human MIND-BRAIN, does not require " the antropic principle". Hence we encounter a clear limitation in the logic thinking capacity of today's humans, Science and Academia !