LAWS@SRI-AI.ARPA (12/31/84)
From: AIList Moderator Kenneth Laws <AIList-REQUEST@SRI-AI.ARPA> AIList Digest Monday, 31 Dec 1984 Volume 2 : Issue 183 Today's Topics: Projects - Cognitive Science Dictionary, AI Tools - Cheap Lisp Machines & Xerox, News - Recent Articles & Thinking Machines Corporation & Space Shuttle, Courses - Massively Parallel Models of Intelligence (CMU) & Reasoning about the Physical World (UIUC) ---------------------------------------------------------------------- Date: Sun, 30 Dec 84 21:53:17 est From: 20568%vax1@cc.delaware (FRAWLEY) Subject: Cognitive Science Dictionary I recently spoke with a publisher about the possibility of compiling a Dictionary of Cognitive Science. I'm sending out this preliminary inquiry to you all to see what you think of the idea. I'd appreciate responses to any or all of the following: 1. Is the idea of such a dictionary good, bad, ridiculous...? 2. Is such a dictionary a feasible project? 3. If the project is feasible, what areas of Cognitive Science ought to be covered? 4. What do you think of the marketability of such a dictionary? 5. If the project is feasible, what form should the dictionary take (i.e., standard dictionary form, encyclopedic form, etc.)? You can send your responses via the AIList or to me directly. Thanks, Bill Frawley Linguistics U. of Delaware 20568.ccvax1@udel ------------------------------ Date: Thu, 27 Dec 84 17:07:33 pst From: hplabs!sdcrdcf!darrelj@Berkeley (Darrel VanBuer) Subject: A Very Cheap Lisp Machine To be slightly partisan toward the machines I use, Xerox Dandelions can be had for under $19,000 in some configurations. For not much over the high end of the proposal in V2 #182, you GET the high end machine (except addition of Ethernet and a display with 6 times the pixels of the Macintosh). About a third of the cost of a Dandelion is for the Interlisp software (inferred from the unbundled Star price list). This is a reasonable cost given the complexity of a full-blown display-oriented Lisp environment and the (relatively) small market for Lisp machines. Darrel J. Van Buer, PhD System Development Corp. 2500 Colorado Ave Santa Monica, CA 90406 (213)820-4111 x5449 ...{allegra,burdvax,cbosgd,hplabs,ihnp4,orstcs,sdcsvax,ucla-cs,akgua} !sdcrdcf!darrelj VANBUER@USC-ECL.ARPA ------------------------------ Date: 26 Dec 1984 1757 PST From: Larry Carroll <LARRY@JPL-VLSI.ARPA> Reply-to: LARRY@JPL-VLSI.ARPA Subject: Xerox AI Paul Erler's message reminds me: the latest Computerworld has a full-page ad with the banner XEROX ANNOUNCES A 15-YEAR HEADSTART IN ARTIFICIAL INTELLIGENCE. It seems they're now selling and supporting what they call the Xerox AI System. It includes a combination of 1108 or 1132 workstations, Interlisp D and LOOPS, and training as well as support. Added info can be gotten from attn: AI Marketing, MS 1245 Xerox Special Information Systems Artificial Intelligence Business Unit 250 N. Halstead St., PO Box 7018 Pasadena, CA 91109 ------------------------------ Date: Sat, 29 Dec 84 06:24:05 cst From: Laurence Leff <leff%smu.csnet@csnet-relay.arpa> Subject: Recent AI Articles New Scientist November 8, 1984 Volume 104 No. 1429 pp 10 Japan unveils its fifth generation New Scientist November 15, 1984 Volume 104 No. 1430 AI is Stark Naked from the Ankles Up. [An entertaining article claiming that the emperor's new clothes (AI) consist only of sneakers (20-year-old expert systems technology). -- KIL] Distributing Computing APIC studies in Data Processing Volume 20 Edited by F. B. Chambers D. A. Dune G. P. Jones Academic Press $22.50 The following titles in this compendium might be of interest: Using Algebra for Concurrency Reasoning about Concurrent Systems Functional Programming Logic Programming and Prolog ------------------------------ Date: Mon 31 Dec 84 11:40:57-PST From: Ken Laws <Laws@SRI-AI.ARPA> Subject: Thinking Machines Corporation From the January, 1985, issue of Omni, p. 33, by Edward Rosenfeld: [...] The latest fusion of acadame and venture capital is Thinking Machines Corporation (TMC), a Cambridge, Mass., company that boasts Marvin Minsky, cofounder of MIT's AI Laboratory and one of the pioneers of AI, as one of its founders. A group of investors headed by CBS founder William Paley has reportedly put up a $10 million stake to get TMC off the ground. AI insiders refer to the company as the Marv and Marv Show because, in addition to Marvin Minsky, TMC has also acquired the services of Marvin Denicoff, who formerly guided the AI programs at the Office of Naval Research. The company's first product, currently in prototype development, will be the connection machine, a parallel-processing supercomputer designed by W. Daniel Hillis, of MIT. [...] -- Ken Laws ------------------------------ Date: Fri, 28 Dec 84 14:04:56 est From: nikhil@mit-fla (Rishiyur S. Nikhil) Subject: AI and the Shuttle Here are some items of interest from Aviation Week and Space Technology: ++++ AWST Sep 17, 1984, page 79 Johnson Space Center (Houston) officials expect to use AI techniques in future Shuttle missions, beginning late 1984 or in 1985. The first use will be Navex, a "navigational expert system". Currently, the navigation console position is manned in 4 shifts, with 3 controllers per shift. Each person needs 2 years of training to make high-speed decisions about shuttle velocity and trajectory. JSC officials expect to man it with one controller per shift in conjunction with Navex. Navex is built on ART (Automatic Reasoning Tool), which is written in Lisp. ART is a product of Inference Corp. of Los Angeles. Navex was developed by Inference Corp. and LinCom Corp. of Houston. ++++ AWST Dec10, 1984, page 24 NASA will test Navex along with its human counterparts in Jan 1985. A Symbolics computer will run in a lab near Mission Control at Johnson Space Center, Houston, and will be wired to the navigator console position. They expect it to make decisions about Shuttle velocity and trajectory six times faster than humans. By March, an AI program will perform Shuttle electrical system checks during pre-launch ground preparations. The actual program is finished, but documentation to explain it will take 3 months. (!!) By late summer 1985, Johnson Space Center wil complete an expert system that captures the expertise of a person whose job would be to talk the shuttle down during re-entry if it were to emerge from a radio blackout with malfunctioning navigation instruments. It will take 2 months to build, and will run in Mission Control as an advisor to flight controllers. ------------------------------ Date: 22 Dec 1984 1152-EST From: Geoff Hinton <HINTON@CMU-CS-C.ARPA> Subject: Course - Massively Parallel Models of Intelligence [Forwarded from the CMU bboard by Laws@SRI-AI.] Advanced Course on: MASSIVELY PARALLEL MODELS OF NATURAL INTELLIGENCE Geoffrey Hinton & Scott Fahlman This is a 7 week advanced course. It meets from 11.30 - 12.50 on Wednesdays and Fridays in 5409, starting on Wednesday Jan 16. A reading list and a brief description of each lecture will be available from Geoff Hinton on Jan 15th. The course covers models of @b(search, representation,) and @b(learning) in networks of simple processing elements that are richly interconnected. The emphasis will be on the computational properties of these networks, but we will also cover the psychological and neurophysiological evidence for and against various models. SEARCH The main search technique used in these networks is iterative relaxation. Five different models of relaxation will be presented and their performance will be compared on a variety of tasks including stereo-fusion, surface-interpolation, shape-recognition, and figure-ground segmentation. Other search methods will also be covered. REPRESENTATION To make efficient use of the representational capacity of massively parallel networks, it is often necessary to use novel kinds of representation in which individual processing elements do not have a simple relationship to the concepts being represented. We will cover methods of representing continuous variables, high-dimensional feature spaces, spatial transformations, simple associations, schemas, trees, production systems, and Clyde. We will discuss the interaction between representational efficiency and ease of search for each kind of representation. LEARNING We will cover the history of attempts to make networks that learn by modifying connection strengths, and show why these attempts generally failed or worked only for very circumscribed domains. The difficult problem in learning is to construct @i(new) representations. We will compare three different models that create representations by modifying connection strengths. We will also compare these connectionist models with more conventional AI learning methods. ------------------------------ Date: Thu, 27 Dec 84 20:57:01 cst From: Kenneth Forbus <forbus%uiucdcsp@uiuc.ARPA> Subject: Course - Reasoning about the Physical World (UIUC) Course Announcement - U. of Illinois at Urbana CS 497, Spring 1985 Title: Reasoning about the Physical World Instructor: Ken Forbus This graduate seminar will examine principles and methods developed in Artificial Intelligence for reasoning about problems involving space, time, processes, and action. Topics include: solving word problems; qualitative physics; planning actions, experiments, assemblies, and routes; analysis, design, troubleshooting, and control of engineered systems. A solid AI background will be assumed. Outline: 1. Solving Textbook Physics Problems Survey of programs: Charniak's CARPS, Novak, Larkin, Bundy, de Kleer. Transformation from natural language to equations Symbolic algebra 2. Qualitative Physics Qualitative State representation: ontology, making predictions, correlating qualitative results with quantitative results, using qualitative reasoning to guide search for quantitative solutions. Qualitative Process theory: processes as mechanisms of change, influences as representation of equations, basic deductions sanctioned by QP theory, prediction, measurement interpretation. Qualitative System Dynamics: breakdown of processes when system connectivity becomes high, device-centered model for physics. Confluences as representation of equations, constraint-satisfaction and propagation techniques for solving confluences. 3. Planning "Classical" AI planning: GPS, STRIPS, NOAH, MOLGEN. Limitations due to inadequate models of time, space, and action. Modelling time: Histories and Chronicles. Allen's interval-based formulation. Vere's DEVISER. Theories of action. Modelling space: symbolic, metric, and analog representations of space. The "visual routines" model of human spatial competence. Robot planning (routes): Configuration space approach and related computational problems. Quantizing free space into "freeways". Robot planning (assembly): Symbolic analysis of errors. Automatic insertion of inspection steps into assembly plans. 4. Engineering Problem Solving Analysis: Propagation of constraints, EL. Qualitative analysis for functional recognition. Design: SYN, the role of causality in circuit design, circuit grammars. Troubleshooting: Digital electronics: Davis' group and the DART project. Continuous systems: SOPHIE. Control: Temporal logic for synthesizing control strategies. ------------------------------ End of AIList Digest ********************