LAWS@SRI-AI.ARPA (07/01/85)
From: AIList Moderator Kenneth Laws <AIList-REQUEST@SRI-AI> AIList Digest Monday, 1 Jul 1985 Volume 3 : Issue 84 Today's Topics: Queries - Expert System Validation & LISP Productivity, Psychology - Predation/Cooperation & Common Sense, Business - TI and Sperry Join Forces, Games - Chess Programs and Cheating, Seminars - Learning in Expert Systems (Rutgers) & How to Clear a Block (SRI) ---------------------------------------------------------------------- Date: Sat, 29 Jun 85 01:58:21 edt From: Walter Maner <maner%bgsu.csnet@csnet-relay.arpa> Subject: Expert System Validation I would appreciate pointers to research dealing with answers to questions about expert system bugs, e.g., How can expert-system advice be validated? Are there failure modes specific to expert systems? What classes of error can be prevented by consistency enforcers? I am primarily interested in how these answers would apply to very large rule-based systems which have evolved under multiple authorship. Walter Maner CSNet maner@bgsuvax UseNet ...cbosgd!osu-eddie!bgsuvax!maner SnailMail Department of Computer Scinece Bowling Green State University Bowling Green, OH 43403 ------------------------------ Date: Wed, 26 Jun 85 13:02 CDT From: Patrick_Duff <pduff%ti-eg.csnet@csnet-relay.arpa> Subject: requested: papers concerning LISP programmer man-hours I am trying to locate articles which discuss the differences between LISP and non-AI languages in terms of the time and effort required to create prototype systems, to make additions or revisions to a design after much of the programming is completed, total programming time from start to finish, etc.. My opinion is that in general, it takes fewer man-hours to create a LISP program than to create a program to do the same task using languages such as Ada, Pascal, FORTRAN, or an assembly language. Note that I am *not* claiming that the program will also be "better", more efficient, or faster--just that most relatively large programs will take less time to write in LISP. I have been asked to come up with justification for using LISP based upon the total man-hours required. Does anyone know of a paper which would support or undercut my opinion? Has there been a convincing demonstration or test of the power of LISP (and its powerful programming environment) versus more traditional languages? regards, Patrick Patrick S. Duff, ***CR 5621*** pduff.ti-eg@csnet-relay 5049 Walker Dr. #91103 214/480-1659 (work) The Colony, TX 75056-1120 214/370-5363 (home) (a suburb of Dallas, TX) ------------------------------ Date: Saturday, 29 Jun 1985 22:22-EST From: munnari!psych.uq.oz!ross@seismo Subject: Predation/Cooperation (AIL v3 #78) David West (AIL v3 #82) mentioned the work of Robert Axelrod on the evolution of cooperation. Another good summary of Axelrod's work can be found in Douglas Hofstadter's Metamagical Themas column in Scientific American, May 1983, v248 #5, pp 14-20. UUCP: {decvax,vax135,eagle,pesnta}!mulga!psych.uq.oz!ross ARPA: ross%psych.uq.oz@seismo.arpa CSNET: ross@psych.uq.oz ACSnet: ross@psych.uq.oz Mail: Ross Gayler Phone: +61 7 224 7060 Division of Research & Planning Queensland Department of Health GPO Box 48 Brisbane 4001 AUSTRALIA ------------------------------ Date: Thu, 27 Jun 85 14:08:05 pdt From: Evan C. Evans <evans%cod@Nosc> Subject: Common Sense Common sense = conclusions reached thru the processes of natural reasoning (or behaviors resulting from such). I borrow heavily from Julian Jaynes, The Origin of Consciousness in the Breakdown of the Bicameral Mind. Natural reasoning is neither conscious nor rigorous in the sense of formal logic. For in- stance, upon observing a piece of wood floating on a given pond one will conclude directly that ANOTHER piece of wood will float on ANOTHER pond. This is sometimes called reasoning from partic- ulars. More simply, it's expectation based on subliminal gen- eralization. A baby quickly concludes that objects will fall without being AWARE of that conclusion. We're constantly ex- ercising natural reasoning to reach conclusions about others' feelings or motives based on their expressions or actions. Such reasoning was early recognized as unconscious & called automatic inference or COMMON SENSE, see John Steward Mill or James Sully. Pu's elaboration on Pratt stands, but it is well to remember that natural reasoning is usually unconscious & does not necessarily proceed by logical means. In fact, automatic infer- ence sometimes achieves correct conclusions by demonstrably il- logical means. evans@nosc-cc ------------------------------ Date: Thu 27 Jun 85 15:54:15-CDT From: Werner Uhrig <CMP.WERNER@UTEXAS-20.ARPA> Subject: news: TI and SPERRY join forces to sell AI [ from Austin American Statesman - June 26, 1985 ] TI captures computer deal with Sperry ===================================== (Kirk Ladendorf - Statesman staff) - TI has landed what it calls its biggest ever sales contract in the still-infant artificial intelligence industry - a three-year, $42 million deal to supply computers and related equipment to Sperry Corp. Sperry, one of the largest computer makers with $5.7 billion in sales last year, plans a large-scale campaign to develop specialized, salable uses for the TI machine, which Sperry will call the Knowledge Workstation. For TI the contract gives credibility that its well-regarded artificial intelligence system, called Explorer, is more than just an esoteric product with limited sales potential. ..... Sperry will combine the TI machine with a software system called the Knowledge Engineering Environment software developed by Intellicorp of Menlo Park, Calif. Intellicorp software is regarded as a very sophisticated tool for building specialized AI-programs. The new system can be used to create so-called "expert systems" which ... Such programs have been used on a demonstration basis to perform such tasks as the running of an electrical power-plant and experimental weather forecasting. Sperry's 26 specialized applications programs will be aimed at areas that have been difficult to serve with traditional computers. Those areas include software development; testing and debugging; navigation; communications sognal processing; CAD/CAM; and scheduling and resource allocation. Sperry chose TI over 2 principal competitors in the field, Symbolics and Lisp Machine Inc, because TI "has the best AI hardware available," a Sperry spokesman said. ..... Sales of AI Lisp-machines totaled only about $85 million last year, but Sperry projects the AI market will mushroom to more than $4 billion by 1990. ..... TI has announced no major additions to its 3,000 person Austin staff because of the new contract, but ... it has already begun to build the staff it needs to support the Explorer and the Business Pro. TI is already at work developing new features for the Explorer. They include developing computer communications links so that the AI machine can interact with Sperry and other IBM-compatible mainframes. .... ------------------------------ Date: Fri 21 Jun 85 21:44:18-EDT From: Andrew M. Liao <WESALUM.A-LIAO-85@KLA.WESLYN> Reply-to: LIAO%Weslyn.Bitnet@WISCVM.ARPA Subject: Chess, Programs And Cheating A Consideration Of "Do Computers Cheat At Chess?" I've been giving some thought to the question, "Do computers actually cheat at chess?". To start, I'm going to assume that what is at issue in the first objection is a chess program's use of a game tree whose nodes are representations of potential board/piece/move configurations. I think the objection that computers cheat because they use "external boards" (albeit represented internally) can be answered by saying, "No - there is no cheating involved because humans 'look ahead' in some way and since no physical external boards are allowed, the only way to 'look ahead' is to represent an 'image' of potential board positions in one's mind [though in a real limited way]. But isn't this just what a program does - only better?" I think that, in some sense, the argument that programs cheat at chess by virtue of having "internally represented 'external boards'" is just wrong. What a program tries to do on one aspect is to simulate what is going on inside a person's mind and, in a limited sense, this is actually achieved (albeit by brute force game trees). The second objection concerns the problems of "moves-made- by-reference". The objection, if I understand it correctly, is that (1) one cannot refer to moves that have been pre- recorded for the player's use during a match and that (2) such moves are encoded into a program (we disallow an external database file of moves since it is, in some way, a set of moves that have been pre-recorded for future use), and without these encoded moves, a program does not know what opening move(s)/strategy(ies) is optimal. Presumably, the reason for this rule is to force a player to rely on his experience and no one else's (i.e. no outside help) and at the same time, prevent any player being put at an unfair disadvantage. But I think it cannot be denied that encoding any move into a chess program is tantamount to making the program dependent upon the author's experience and not its own - a clear violation of the spirit of the rule. The question remains - Is it cheating? I am of the opinion that such a program is cheating on the basis that the program cannot decide during the opening of the game what strategy is optimal for it and hence must rely on outside help, in the form of stored data, given to it by its author. Although I feel the first objection is easily answered, I am still not happy with my reply to the second, although my intuition tells me that my reply to the second objection is, at least in spirit, on the right track. The motivation for my second reply is due (in great part) to J.R. Searle's conception of the Background which directly relates to the problem of "experience" and the like. ------------------------------ Date: Wed 26 Jun 85 09:58:03-PDT From: Ken Laws <Laws@SRI-AI.ARPA> Subject: Re: Chess, Programs And Cheating Reply to Andrew Liao: When I open with a king pawn, I am relying on the experience and knowledge of others -- that doesn't seem to be cheating. I prefer an interpretation of the rules as "you run what you brung" -- namely that you cannot access external help >>during the match<<. I do admit that stored book openings seem questionable (although chess masters certainly memorize such material), but to say that a computer's superior memory gives it an advantage is no more damning than to say that its superior speed gives it an advantage. In just a few years it will be obvious that computers are inherently better "chess machines" than people are, and people will stop quibbling about handicaping the computer in one way or another to make the contest "fair". -- Ken Laws ------------------------------ Date: 28 Jun 85 11:07:37 EDT From: PRASAD@RUTGERS.ARPA Subject: Seminar - Learning in Expert Systems (Rutgers) LEARNING IN SECOND GENERATION EXPERT SYSTEMS Walter Van De Velde AI Laboratory, Vrije Universiteit Brussel This talk discusses a learning mechanism for second generation expert systems: rule-learning by progressive refinement. Second generation expert systems not only use heuristic rules, but also have a model of the domain of expertise so that deeper reasoning is possible whenever the rules are deficient. A learning component is described that abstracts new rules out of the results of deep reasoning. Gradually, the rule set is refined and restructured so that the expert system can solve more problems in a more effecient way. The approach is illustrated with concrete implemented examples. Date: Friday, June 28, 1985 Time: 11 AM Place: Hill Center, Room 423 ------------------------------ Date: Thu 27 Jun 85 12:21:03-PDT From: LANSKY@SRI-AI.ARPA Subject: Seminar - How to Clear a Block (SRI) "HOW TO CLEAR A BLOCK" or Unsolved Problems in the Blocks World #17 Richard Waldinger -- SRI AI Center 11:00 am, WEDNESDAY, July 3 Room EJ232, SRI International ABSTRACT: Apparently simple problems in the blocks world get more complicated when we look at them closely. Take the problem of clearing a block. In general, it requires forming conditionals and loops and even strengthening the specifications; no planner has solved it. We consider how such problems might be approached by bending a theorem prover a little bit. ------------------------------ End of AIList Digest ********************