LAWS@SRI-AI.ARPA (11/12/84)
From: AIList Moderator Kenneth Laws <AIList-REQUEST@SRI-AI> AIList Digest Wednesday, 24 Oct 1984 Volume 2 : Issue 144 Today's Topics: Courses - Decision Systems & Introductory AI, Journals - Annotated AI Journal List, Automatic Programming - Query, AI Tools - TI Lisp Machines & TEK AI Machine, Administrivia - Reformatting AIList Digest for UNIX, Humor - Request for Worst Algorithms, Seminars - Metaphor & Learning in Expert Systems & Representing Programs for Understanding ---------------------------------------------------------------------- Date: Tue 23 Oct 84 13:33:06-PDT From: Samuel Holtzman <HOLTZMAN@SUMEX-AIM.ARPA> Subject: Responses to Decision Systems course. Several individuals have requested further information on the course in decision systems I teach at Stanford (advertised in AILIST a few weeks ago). Some of the messages I received came from non-ARPANET sites, and I have had trouble replying electronically. I would appreciate getting a message from anyone who has requested information from me and has not yet received it. Please include a US (paper) mail address for my reply. Thanks, Sam Holtzman (HOLTZMAN@SUMEX or P.O. Box 5405, Stanford, CA 94305) ------------------------------ Date: 22 Oct 1984 22:45:40 EDT From: Lockheed Advanced Software Laboratory@USC-ISI.ARPA Subject: Request for information A local community college is considering adding an introductory course in AI to its curriculum. Evening courses would be of benefit to a large community of technical people interested in the subject. The question is what will be the benefit to first and second year students. If anyone knows of any lower division AI courses taught anywhere, could you please drop me a line over the net. Also, course descriptions on introductory AI classes, either lower or upper division, would be appreciated. Comments on the usefulness or practicality of such a course at this level are also welcome. Thank You, Michael A. Moran Lockheed Advanced Software Laboratory address: HARTUNG@USC-ISI ------------------------------ Date: Tue, 23 Oct 84 11:34 CDT From: Joseph_Hollingsworth <jeh%ti-eg.csnet@csnet-relay.arpa> Subject: annotated ai journal list I am interested in creating an annotated list of the AI related journals list that was published in AIList V1 N43. I feel that this annotated list would be beneficial for those persons who do not have easy access to the journals mentioned in the previously published list, but who feel that some of them may apply to their work. I solicit information about each journal in the following form, (which I will compile and release to the AIList if there is enough interest shown). 1) Journal Name 2) Subjective opinion of the type of articles that frequently appear in that journal (short paragraph or so). 3) Keywords and phrases that characterize the articles/journal, (don't let formalized keyword lists hinder your imagination). 4) The type of scientist, engineer, technician, etc. that the journal would benefit. 5) Address of journal for subscription correspondence, (include price too, if possible). Please send this information to Joe Hollingsworth at jeh%ti-eg@csnet-relay (if you are on the ARPANET) jeh@ti-eg (if you are on the CSNET; I am on the CSNET) The following is the aforementioned list of journals: AI Magazine AISB Newsletter Annual Review in Automatic Programming Artificial Intelligence Artificial Intelligence Report Behavioral and Brain Sciences Brain and Cognition Brain and Language Cognition Cognition and Brain Theory Cognitive Pshchology Cognitive Science Communications of the ACM Computational Linguistics Computational Linguistics and Computer Languages Computer Vision, Graphics, and Image Processing Computing Reviews Human Intelligence IEEE Computer IEEE Transactions on Pattern Analysis and Machine Intelligence Intelligence International Journal of Man Machine Studies Journal of the ACM Journal of the Assocation for the Study of Perception New Generation Computing Pattern Recognition Robotics Age Robotics Today SIGART Newsletter Speech Technology ------------------------------ Date: 23 October 1984 22:28-EDT From: Herb Lin <LIN @ MIT-MC> Subject: help needed on automatic programming information I need some information on automatic programming. 1. How complex a problem can current automatic programming systems handle? The preferred metric would be complexity as measured by the number of lines of code that a good human programmer would use to solve the same problem. 2. How complex a problem will future automatic programming systems be able to handle? Same metric, please. Of course, who can predict the future? More precisely, what do the most optimistic estimates predict, and for what time scale? 3. In 30 years (if anyone is brave enough to look that far ahead), what will automatic programming be able to do? Please provide citable sources if possible. Many thanks. ------------------------------ Date: 22 Oct 1984 12:07:39-PDT From: William Spears <spears@NRL-AIC> Subject: TI Lisp machines The AI group at the Naval Surface Weapons Center is interested in the new TI Lisp Machine. Does anyone have any detailed information about it? Thanks. "Always a Cosmic Cyclist" William Spears Code N35 Naval Surface Weapons Center Dahlgren, VA 22448 ------------------------------ Date: 22 Oct 84 08:10:32 EDT From: Robert.Thibadeau@CMU-RI-VI Subject: TEK AI Machine [Forwarded from the CMU bboard by Laws@SRI-AI.] I have good product literature on the Tektronix 4404 Artificial Intelligence System (the workbook for their people). This appears to be a reasonable system which supports Franz Lisp, Prolog, and Smalltalk-80. It uses a 68010 with floating point hardware and comes standard with a 1024^2 bit map, 20mb disk, floppy, centronics 16 bit port, RS232, 3-button mouse, ethernet interface, 1 mbyte RAM, and a Unix OS. The RAM upgrades at least 1 more mbyte and you can have a larger disk and streaming tape. The major thing is that the price (retail without negotiation) is $14,950 complete. It is apparently real, but I don't know this system first hand. The product description is all I have. ------------------------------ Date: Sat, 20 Oct 84 23:10:53 edt From: Douglas Stumberger <des%bostonu.csnet@csnet-relay.arpa> Subject: reformatting AILIST digest for UNIX For those of you on Berkeley UNIX installations, there is a program available which does the slight modifications to ailist digest necessary so it is in the correct format for a "mail -f ...". This allows using the UNIX mail system functionality to maintain your ailist digest files. For a copy of the program, net to: douglas stumberger csnet: des@bostonu ------------------------------ Date: Mon 22 Oct 84 10:30:00-PDT From: Jean-Luc Bonnetain <BONNETAIN@SUMEX-AIM.ARPA> Subject: worst algorithms as programming jokes [Forwarded from the Stanford bboard by Laws@SRI-AI.] After reading the recent complaint(s) about those people who slow down the system with their silly programs to sort a 150-element list, and talking with a friend, I came up with the following dumb idea : A lot of emphasis is understandably put on good, efficient algorithms, but couldn't we learn also from bad, terrible algorithms ? I have heard that Dan Friedman at Indiana collects elegant LISP programs that he calls LISP poems. To turn things upside down, how about LISP jokes (more generally, programming jokes) ? I'm pretty sure most if not all of programmers have some day (night) burst into laughter when encountering an algorithm that is particularly dumb, and funny for the same reason. I don't know whether anyone ever collected badgorithms (sorry, that was the worst name I could find), so I suggest that you bright guys send me your favorite entries. To qualify as a badgorithm, the following conditions should be met: (if you don't like them, send me your suggestions for a better definition) 1. It *is* an algorithm in the sense described by Knuth Vol 1. 2. It *does* solve the problem it addresses. Entering the Knuth-Bendix algorithm as a badgorithm for binary addition is illegal (though I admit it is somewhat funny). 3. Though it solves the problem, it must do so in an essentially clumsy way. Adding loops to slow down the algorithm is cheating. In some sense a badgorithm should totally miss the right structure to approach the problem. 4. The hopeless off-the-track-ness of a badgorithm should be humorous for someone knowledgeable with the problem addressed. We are not interested in alborithms, right ? Just being the second or third best algorithm for a problem is not enough to qualify (think of the "common sense" algorithm for finding a word in a text as opposed to the Boyer-Moore algorithm, or of the numerous ways to evaluate a polynomial as opposed to Horner's rule; there is nothing to laugh at in those cases). There is nothing funny in just being a O(n^(3/(pi^3)-1/e)) algorithm, I think. 5. It should be described in a simple, clear way. Remember that the best jokes are the shortest ones. I'm sure there are enough badgorithms for well-known problems (classical list manipulation, graph theory, arithmetic, cryptography, sorting, searching, etc). Please don't enter algorithms to solve NP problems unless you have good reasons to think they are interesting in our sense. If anyone out there is willing to send me an entry, please send the following: * a simple description of the problem (the name is enough if it's a well-known problem). * a verbal description of the badgorithm if possible. * a programmed version of the badgorithm (in LISP preferably). this is not necessary if your verbal description makes it clears enough how to write such a program, but still it would be nice. * a description of a good algorithm for the same problem in case most people are not expected to be familiar with one. Comparing this to the badgorithm should help us in seeing what's wrong with the latter, and I would say that this could have good educational value. To start things, let me enter my favorite badgorithm (I call it "stupid-sort"): * the problem is to sort a list, according to some "order" predicate. * well, that's easy. just generate all permutations of the list, and then check whether they are "order"ed. would you bet that someone in CS105 does actually use this one ? [I once had to debug an early version of the BMD nonparametric package. It found the min and max of a vector by sorting the elements ... (Presumably most users would also request the median and other sort-related statistics.) For a particularly slow sort routine see the Hacker's Dictionary definition of JOCK, quoted in Jon Bentley's April Programming Pearls in CACM. -- KIL] I understand perfectly that some people/organizations do not wish to have their names associated with badgorithms, but please don't refrain from entering something because of that. I swear that if you request it there will be no trace of the origin of the entry if I ever compile a list of them for personal or public use (you know, "name withheld by request" is the usual trick). jean-luc ------------------------------ Date: 17 Oct 1984 16:25-EDT From: Andrew Haas at BBNG.ARPA Subject: Seminar - Metaphor [Forwarded from the MIT bboard by SASW@MIT-MC.] Next week's BBN AI seminar is on Thursday, October 25th at 10:30 AM in the 3rd floor large conference room. Bipin Indurkhya of the University of Massachusetts at Amherst will speak on "A Computational Theory of Metaphor Comprehension and Analogical Reasoning". Abstract follows. Though the pervasiveness and importance of metaphors in natural languages is widely recognised, not much attention has been given to them in the fields of Artificial Intelligence and Computational Linguistics. Broadly speaking, a metaphor can be characterized as application of terms belonging to source domain in describing target domain. A large class of such metaphors are based on structural analogy between the two domains. A computational model of metaphor comprehension was proposed by Carbonell which required an explicit representation of a mapping which maps terms of the source domain to the terms of the target domain. In our research we address ourselves to the question of how one can characterize this mapping in terms of the knowledge of the source and the target domains. In order to answer this question, we start from Gentner's theory of Structure-Mapping. We show limitations of Gentner's theory and propose a theory of Constrained Semantic Transference [CST] that allows part of the structure of the source domain to be transferred to the target domain coherently. We will then introduce two recursive operators, called Augmentation and Positing Symbols, that make it possible to create new structure in the target domain constrained by the structure of the source domain. We will show how CST captures several cognitive properties of metaphors and then discuss its limitations with regard to computability and finite representability. If time permits, we will use CST as a basis to develop a theory of Approximate Semantic Transference which can be used to develop computational models of the cognitive processes involved in metaphor comprehension, metaphor generation, and analogical reasoning. ------------------------------ Date: Tue 23 Oct 84 10:45:51-PDT From: Paula Edmisten <Edmisten@SUMEX-AIM.ARPA> Subject: Seminar - Learning in Expert Systems [Forwarded from the Stanford SIGLUNCH distribution by Laws@SRI-AI.] DATE: Friday, October 26, 1984 LOCATION: Chemistry Gazebo, between Physical and Organic Chemistry TIME: 12:05 SPEAKER: Li-Min Fu Electrical Engineering ABSTRACT: LEARNING OBJECT-LEVEL AND META-LEVEL KNOWLEDGE IN EXPERT SYSTEMS A high performance expert system can be built by exploiting machine learning techniques. A learning method has been developed that is capable of acquiring new diagnostic knowledge, in the form of rules, from a case library. The rules are designed to be used in a MYCIN-like diagnostic system in which there is uncertainty about data as well as about the strength of inference and in which the rules chain together to infer complex hypotheses. These features greatly complicate the learning problem. In machine learning, two issues that can't be overlooked are efficiency and noise. A subprogram, called "Condenser," is designed to remove irrelevant features during learning and improve the efficiency. It works well when the number of features used to characterize training instances is large. One way of removing noise associated with a learned rule is seeking a state with minimal prediction error. Another subprogram has been developed to learn meta-rules which guide the invocation of object-level rules and thus enhance the performance of the expert system using the object-level rules. By embodying all the ideas developed in this work, an expert program called JAUNDICE is built, which can diagnose the likely cause and mechanisms of a patient with jaundice. Experiments with JAUNDICE show the developed theory and method of learning are effective in a complex and noisy environment where data may be inconsistent, incomplete, and erroneous. Paula ------------------------------ Date: Tue, 23 Oct 84 00:08:10 cdt From: rajive@ut-sally.ARPA (Rajive Bagrodia) Subject: Seminar - Representing Programs for Understanding [Forwarded from the UTexas-20 bboard by Laws@SRI-AI.] Graduate Brown Bag Seminar: Representing Programs For Understanding by Aaron Temin noon Friday Oct. 26 PAI 3.36 Automatic help systems would be much easier to generate than they are now if the same code used to create the executable version of a program could be used as the major database for the help system. The desirable properties of such a program representation will be discussed. An overview of MIRROR, our implementation of those properties, will be presented with an explanation of why MIRROR works. It will also be argued that functional program representations are inadequate for the task. If you are interested in receiving mail notifications of graduate brown bag seminars in addition to the bboard notices, please send a note to briggs@ut-sally ------------------------------ End of AIList Digest ********************