Date: Tue 21 Jun 1988 17:51-EDT From: AIList Moderator Nick Papadakis Reply-To: AIList@AI.AI.MIT.EDU Us-Mail: MIT Mail Stop 38-390, Cambridge MA 02139 Phone: (617) 253-2737 Subject: AIList Digest V7 #41 To: AIList@AI.AI.MIT.EDU Status: RO AIList Digest Wednesday, 22 Jun 1988 Volume 7 : Issue 41 Today's Topics: Representation languages AI language determinism a dead issue? Ding an sich The primacy of scientific physical reality? Cognitive AI vs Expert Systems ---------------------------------------------------------------------- Date: Mon, 20 Jun 88 08:16:19 PDT From: smoliar@vaxa.isi.edu (Stephen Smoliar) Subject: Re: representation languages In article <19880615061555.7.NICK@INTERLAKEN.LCS.MIT.EDU> Ian Dickson writes: >Date: Tue, 14 Jun 88 05:42 EDT >From: Ian Dickinson >To: ailist@mc.lcs.mit.edu >Subject: Re: representation languages > >Whilst I have no doubt that these systems [KEE and ART] are useful today, _I_ as an >application developer want to see a representation system that is >maximally >small whilst giving me the power that I need. The philosophy I would like >to >see adopted is: > o define conceptual representations that allow applications to be > written at the maximum level of abstraction (eg generic tasks) > o define the intermediate representations (frames, rules, sets ..) > that are needed to implement the conceptual structures > o choose a subset of these representations that can be maximally > tightly integrated with the base language of your choice (which > would not be Lisp in my choice) > These are admirable desiderata, but they may not be sufficient to stave off the dreaded "good feature explosion." Rather, this malady is a consequence of a desire we seem to have of our representation systems which allows them to both RECORD and REASON ABOUT "units" of knowledge (whatever those units may be). (PACE, Mark Stefik; I know I have lifted the name of a knowledge representation system in my choice of words.) We take it for granted that we want both facilities. If all we were doing was recording, all we would have would be a data base; and if all we were doing was reasoning, all we would have would be a theorem prover. I would claim that our cultural expectations of a knowledge representation system has grown out of a desire to assimilate these two capabilities. Unfortunately, both capabilities turn out to be extremely demanding of computational resources. As a result, it has been demonstrated that even some of the simplest attempts to find a viable middle ground can easily lead to computational intractability (particularly if a clean semantic foundation is one of your desiderata). As a result, now may be a good time to question whether or not the sort of "homogeneous assimilation" of recording and reasoning which is to be found in many knowledge representation systems is such a good thing. Perhaps it would be more desirable to have TWO facilities which handle record keeping and reasoning as independent tasks and which communicate through a protocol which does not impede their interaction. Here at ISI we have been exploring means by which expert systems can give adequate explanatory accounts of their own behaviror. We have discovered that an important element in the service of such explanation is a TERMINOLOGICAL FOUNDATION, which amounts to a means by which all symbols which are used as part of the problem solving apparatus of the expert system also have a semantic support which links them to the text generation facilities required in explanation. Thus, for example "fever" is not treated simply as a symbolic varaible which get set to T if a patient's temperature is more than 100 degrees Farenheit but may then get set back to NIL if it is discovered that the patient had been drinking hot coffee just before the nurse took his temperature; rather it is a "word" which serves as a key to certain knowledge about patient conditions, as well as knowledge about how it may be detected and knowledge about its consequences. In keeping with the aforementioned attempt to separate the concerns of recording and reasoning, we have developed a facility (currently called HI-FI) for recording such terminological knowledge in such a way as to SUPPORT (but not necessarily PERFORM) subsequent reasoning. In pursuing this approach, we have developed as set of "terminological building blocks," which seem to be at least partially sympathetic to Ian Dickinson's philosophy. Here is a quick outline: ACTIONS are the "verbs" of the terminological foundation. They provide the basis for the expression of both the statements of problems and the statements of solution methods. While they definitely have a "generic" quality, I am not yet sure that they bear much resemblance to Chandrasekaran's generic tasks. Since this material is relatively new, that possibility remains to be investigated. TYPES are "nouns" which designate classes (i.e. categories of entities). They intentionally bear resemblance to both frame descriptions and object classes which may be found in object- oriented languages. INSTANCES are the entities which are members of classes. An instance may be a member of several classes. However, determinining whether or not a given instance is a member of a given class may often be a matter of reasoning rather than retrieval from a base of recorded facts. PROPERTIES are unary predicates applied to instances. RELATIONS are binary predicates applied to instances. ASSERTIONS are sentences in a typed predicate calculus whose types are the type classes. These sentences are "about" instances; but they may incorporate expressions of descriptions of types, properties, and relations. A major application of assertions is the representation of DEFINITIONAL FORMS. These are sentences which establish necessary and/or sufficient conditions for membership in a type class or for the satisfaction of a property or relation. These assertions are the major link to the reasoning facility. The above is a sketch of work which has just gotten under way. However, the approach has already been pursued in some test cases concerned with reasoning about digital circuits; and it appears to be promising. I plan to have more discipined documentation of our work ready in the near future; but while I am engaged in preparing such documents, I am interested in any feedback regarding similar (or contrary) approaches. ------------------------------ Date: Tue, 21 Jun 88 08:23 O From: Subject: AI language Distribution-File: AILIST@AI.AI.MIT.EDU In a recent AIList issue, Pat Hayes wished to get a list of desirable features for an AI language. My opinion is that we need new theoretical formalisms for expressing intelligence and the world of a human being for the basis of new AI languages. It seems to me that almost all successful programming languages have a good background formalism. APL has Iverson's array notation. Modern Lisp (CommonLOOPS and Zetalisp) has several ones: functional programming, the idea of the list, object-oriented programming. The Algol/Pascal family of languages has the idea of expressing the language unambiguously with the Backus-Naur notation. Andy Ylikoski ------------------------------ Date: 21 Jun 88 17:34:40 GMT From: umix!umich!eecs.umich.edu!itivax!dhw@uunet.UU.NET (David H. West) Subject: Re: determinism a dead issue? In a previous article, Bruce E. Nevin writes: > Is the notion of determinism not deeply undercut by developments in > study of nonlinearity and Chaos? No. (Or, if you prefer, "it depends".) I take it that "determinism" is for present purposes equivalent to "predictability". Then: 1) Nonlinearity is strictly irrelevent - it just makes the math more difficult, but determinism requires only the existence (*) of a solution, not that it be easy to compute, or that it be computed at all; 2) Chaos means (in the continuum view of things) that some quantity has with respect to some initial parameter a derivative the magnitude of which becomes unbounded for large times, i.e. adjacent trajectories diverge. All this means is that to predict further into the future, one needs increasingly precise knowledge of the initial conditions. Infinite precision suffices ;-) for infinite-time prediction. Remember, Laplace (or was it Lagrange?) assumed he could have the positions and velocities of every particle in the universe. Anyone who grants that would be niggardly to refuse infinite precision. Determinism *is* undercut by quantum mechanics, but that's encouraging only to those who identify randomness with freedom. (*) There are epistemological problems here. One can certainly prove things about the existence of solutions to equations, but notwithstanding that we know some equations that describe the world to some degree of approximation, it is clearly impossible for finite beings to prove or know that (P:) "equations (currently known or otherwise) describe the world exactly". Such beings (e.g. us) can believe P or not, as it suits them (or as they are determined ;-). > Is it the case that systems > involving nonlinearity always involve feedback or feedforward loops? Any system worthy of the name has loops, and linearity is only a special case, so this is a good bet, but there are counterexamples (see below). > (Isn't it mutual effect of the values > of two or more variables on one another that makes an equation > nonlinear, and isn't that a way of expressing feedback or feedforward? No. Consider the nonlinear equation y=sqrt(x). > Is it that > nonlinear systems are not error correcting? Or perhaps that they are > analog rather than digital systems? Are massively parallel systems > nonlinear, or do they tend to be? Does the distinction apply to now > familiar characterizations of brain hemisphere specialization? The answer is probably "not necessarily". > This has relevance to how an AI based on deterministic, linear systems > can do what nonlinear organisms do. Whose AI is based on *linear* systems? Logic circuits are nonlinear, semantic networks are nonlinear, connectionist networks are nonlinear... ------------------------------ Date: 20 Jun 88 0322 PDT From: John McCarthy Subject: Ding an sich I want to defend the extreme point of view that it is both meaningful and possible that the basic structure of the world is unknowable. It is also possible that it is knowable. It just depends on how much of the structure of the world happens to interact with us. This is like Kant's "Ding an sich" (the thing in itself) except that I gather that Kant considered "Ding an sich" as unknowable in principle, whereas I only consider that it might be unknowable. The basis of this position is the notion of evolution of intelligent beings in a world not created for their scientific convenience. There is no mathematical theorem requiring that if a world evolves intelligent beings, these beings must be in a position to discover all its laws. To illustrate this idea, consider the Life cellular automaton proposed by John Horton Conway and studied by him and various M.I.T. hackers and others. It's described in Winning Ways by Berlekamp, Conway and Guy. Associated with each point of the two dimensional integer lattice is a state that takes values 0 and 1. The state of a point at time t+1 is determined by its state at time t and the states at time t of its eight neighbors. Namely, if the number of neigbors in state 1 is less than two or more than 4, its state at time t+1 is 0. If it has exactly two neighbors in state 1, its state remains as it was. If it has exactly 3 neighbors in state 1, its new state is 1. There is a configuration of five cells in state 1 (with neighbors in state 0) called a glider, which reproduces itself displaced in two units of time. There is a configuration called a glider gun that emits gliders. There are configurations that thin out streams of gliders from a glider gun. There are configurations that take two streams of gliders as inputs and perform logical operations (regarding the presence of a glider at a given time in the stream as 1 and its absence as 0) on them producing a new stream. Thinned streams can cross each other and serve as wires conducting signals. This permits the construction of general purpose computers in the Life plane. The Life automaton wasn't designed to admit computers. The discovery that it did was made by hacking. Configurations that can serve as general purpose computers can be made in a variety of ways. The way indicated above and more fully described in Berlekamp, et. al. is only one. Now suppose that one or more interacting Life computers are programmed to be physicists, i.e. to attempt to discover the fundamental physics of their world. There is no reason to expect a mathematical theorem about cellular automata in general or the Life cellular automaton in particular that says that a physicist program will be able to discover that the fundamental physics of its world is the Life cellular automaton. It requires some extra attention in the design of the computer to make sure that it has any capability to observe at all, and some that can observe will be unable to observe enough detail. Of course, we could program a Life computer to simulate some other "second level" cellular automaton that admits computers, and give the "second level computer" only the ability to observe the "second level world". In that case, it surely couldn't find any evidence for the its world being the Life cellular automaton. Indeed the Life automaton could simulate exceedingly slowly any theory we like of our 3+1 dimensional world. If a Life world physicist is provided with too narrow a philosophy of science, and some of the consensual reality theories may indeed be that narrow, it might not regard the hypothesis that its physics is the Life world as meaningful. There may be Life world physicists who regard it as meaningful and Life world philosophers of science interacting with them who try to forbid it. This illustrates what I mean by metaepistemology. Metaepistemology must study what knowledge is possible for intelligent beings in a world to the structure of the world and the physical structures and computational programs that support scientific activity. The traditional methods of philosophy of science are too weak to discuss these matters, because they don't take into account how the structure of the world and the structure of its intelligences affect what science is possible. There is no more guarantee that the structure of our world is observable than that Fermat's last theorem is decidable in Peano arithmetic. Physicists are always proposing theories of fundamental physics whose testability depends on the correctness of other theories and the development of new apparatus. For example, some of the current GUTS theories predict unification of the force laws at energies of 10^15 Mev, and there is no current idea of how an accelerator producing such an energy might be physically possible. I have received messages asking me if the metaepistemology I propose is like what has been proposed by Kant and other philosophers or even by Winograd and Flores. As far as I can tell it's not, and all those mentioned are subject to the criticism of the previous paragraph. ------------------------------ Date: Mon 20 Jun 88 10:20:14-PDT From: Mike Dante Subject: Re: AIList Digest V7 #39 I can't resist replying to George McKee's insistence that "one description of the collective experience of humanity ... outranks all the alternatives ... (that is, the ) primacy of scientific physical reality". The statement is of course true only if you exclude from the "collective experience of humanity" all of history, aesthetics, human relationships, and self understanding. But I think you must also exclude George's own belief that we are only a "quantitative step away" from "telling us what we need to know." This non- scientific belief was held by Marx, Freud, etc., etc., all of whom wished to believe that the crystal purity and certainty of the scientific method had solved mankind's ills. It seems to me that the evidence necessary to support this belief would be at least some demonstrated success that we were some sort of "quantitative step away" from having any idea how to close prisons and mental hospitals, and abolish greed, fear, and war. So far I see no scientific evidence that we have more than a laundry list of things to try, and a much longer list of things that have been tried and have failed. To extrapolate from the limited (though exciting and important) successes of the scientific method in these fields to an assertion that we are only quantitatively distant from describing "the collective experience of humanity" seems to me a great deal less justified than the belief that was expressed by the Dean of American Science in the 19th Century, that all of Physics had been learned and all that was left was quantitative improvements. ------------------------------ Date: 21 Jun 88 03:02:26 GMT From: krulwich-bruce@yale-zoo.arpa (Bruce Krulwich) Subject: Re: Cognitive AI vs Expert Systems In a previous post, I claimed that there were differences between people doing "hard AI" (trying to achieve serious understanding and intelligence) and "soft AI" (trying to achieve intelligent behavior). dg1v+@ANDREW.CMU.EDU (David Greene) responds: >Since my researchs concerns developing knowledge acquisition approaches (via >machine learning) to address real world environments, I'm well aquainted with >not only the above literature, but psych, cog psych, JDM (judgement and >decision making), and BDT (behavioral decision theory). > >While I suspect AI researchers who work in Expert System might resent being >excluded from work in "serious intelligence", I think my point is that, for a >given phenomena, multiple viewpoints from different disciplines (literature) >can provide important breadth and insights. I agree fully, and I think you'll find this in the references section of alot of "hard AI" research work. (As a matter of fact, a fair number of researchers in "hard AI" are prof's in or have degrees psychology, linguistics, etc.) I'm sorry if my post seemed insulting -- it wasn't intended that way. I truly believe, however, that there are differences in the research goals, methods, and results of the two different areas. That's not a judgement, but it is a difference. Bruce Krulwich ------------------------------ End of AIList Digest ********************