AIList-REQUEST@SRI-AI.ARPA (AIList Moderator Kenneth Laws) (11/19/85)
AIList Digest Tuesday, 19 Nov 1985 Volume 3 : Issue 173 Today's Topics: Seminars - Adaptive Planning (UCB) & Sparse Distributed Memory (BBN) & Explanation-Based Learning (BBN) & Learning Search Control Knowledge (CMU) & MED2 Diagnostic Expert (MIT) & Truth Maintenance, Multiple Worlds in KEE (SU) & Representation of Natural Form (SU) & Setting Tables and Illustrations With Style (CSLI), Course - Connectionist Models (CMU) ---------------------------------------------------------------------- Date: Thu, 14 Nov 85 16:57:53 PST From: admin%cogsci@BERKELEY.EDU (Cognitive Science Program) Subject: Seminar - Adaptive Planning (UCB) BERKELEY COGNITIVE SCIENCE PROGRAM Fall 1985 Cognitive Science Seminar - IDS 237A Tuesday, November 19, 11:00 - 12:30 240 Bechtel Engineering Center Discussion: 12:30 - 1:30 in 200 Building T-4 ``Adaptive Planning is Commonsense Planning'' Richard Alterman Computer Science Division, U.C.B. A characteristic of commonsense planning is that it is knowledge intensive. For most mundane sorts of situations human planners have access to, and are capable of exploiting, large quantities of knowledge. Commonsense planners re-use old plans under their normal circumstances. Moreover, commonsense planners are capable of refitting old plans to novel cir- cumstances. A commonsense planner can plan about a wide range of phenomena, not so much because his/her depth of knowledge is consistent throughout that range, but because s/he can re-fit old plans to novel contexts. This talk is about an approach to commonsense planning called adaptive planning. An adaptive planner plans by exploit- ing planning knowledge in a manner that delays the reduction of commonsense planning to problem-solving. Key elements in the theory of adaptive planning are its treatment of background knowledge and the introduction of a notion of planning by situation matching. This talk will describe adaptive planning as it applies to a number of commonsense planning situations, including: riding the NYC subway, trading books, transferring planes at JFK airport, and driving a rented car. ------------------------------ Date: 14 Nov 1985 11:48-EST From: BGOODMAN at BBNG.ARPA Subject: Seminar - Sparse Distributed Memory (BBN) [Forwarded from the MIT bboard by SASW@MIT-MC.] BBN Labs SDP AI Seminar Speaker: Dr. Michael R. Raugh Research Institute for Advanced Computer Science NASA Ames Research Center Title: Kanerva's Sparse Distributed Memory: A RIACS Project Date: Friday, November 22nd, 2:00pm Location: Main Seminar Room (2nd floor) Bolt Beranek and Newman Inc. 50 Moulton Street Cambridge, MA. An exciting new concept in which information is stored in a large number of neighboring addresses determined by "content," produces a memory that retrieves causal relationships as well as sequences of episodes and is sensitive to similarity. It is also forgetful and reinforcable: a memory much like yours and mine. ------------------------------ Date: 14 Nov 1985 11:48-EST From: BGOODMAN at BBNG.ARPA Subject: Seminar - Explanation-Based Learning (BBN) [Forwarded from the MIT bboard by SASW@MIT-MC.] BBN Labs SDP AI Seminar Speaker: Professor Gerald DeJong Coordinated Science Laboratory University of Illinois at Urbana-Champaign Title: Explanation Based Learning Date: Monday, November 25th, 10:30am Location: 2nd Floor Large Conference room BBN Laboratories Inc. 10 Moulton Street Cambridge, MA. Machine learning is one of the most important current areas of artificial intelligence. With the trend away from "weak methods" and toward a more knowledge intensive approach to intelligence, the lack of knowledge in an AI system becomes one of the most serious limitations. This talk advances a technique called explanation based learning. It is a method of learning from observation. Basically, it involves endowing a system with sufficient knowledge so that intelligent planning behavior of others can be recognized. Once recognized, these observed plans are generalized as far as possible while preserving the underlying explanation of their success. The approach supports one-trial learning. A new general concept can be acquired from an observation of just one observed example. The approach has been applied to three diverse areas: natural language processing, robot task planning, and proof of propositional calculus theorems. The approach holds promise for solving the knowledge collection bottleneck in the construction of current knowledge-based systems. ------------------------------ Date: 14 Nov 85 23:39:59 EST From: Steven.Minton@CAD.CS.CMU.EDU Subject: Seminar - Learning Search Control Knowledge (CMU) [Forwarded from the CMU bboard by Laws@SRI-AI.] On Wednesday, November 20, at 12:00 I will present my thesis proposal in 5409. My thesis is concerned with the use of explanation-based generalization in the PRODIGY system, a learning apprentice that (among other things) acquires search control rules. The title is: "Analytic Techniques for Learning Search Control Knowledge". Copies are in the lounge. ABSTRACT Compression analysis, the subject of the proposed thesis, is a method for analyzing search spaces to produce effective search control rules. As with previous explanation-based learning techniques, an example problem focuses the analysis process so that the entire search space need not be analyzed. The key idea behind compression analysis is that many alternative explanations can be produced to justify a search control decision; therefore it is appropriate to search for an explanation that produces the most effective generalized control rule. In practice this is achieved by proposing an initial explanation which is then improved using a set of heuristic transformation strategies. ------------------------------ Date: Sun, 17 Nov 85 16:28:06 EST From: "Steven A. Swernofsky" <SASW@MIT-MC.ARPA> Subject: Seminar - MED2 Diagnostic Expert (MIT) Wednesday 20, November 4: 00pm (4:15 Refreshments) Room: NE43-512A "MED2: An Expert System Shell for Diagnosis and Therapy in Complex Domains" Frank Puppe Kaiserlautern University Germany Concentrating on the medical domain, MED2 is a shell combining a wide variety of important aspects of clinical reasoning. It's "Working-Memory" control structure involves investigating a set of hypotheses simultaneously, avoiding the shortcomings of focussing on the top-hypothesis only. This concept allows using differential diagnosis techniques and exploiting relationships among patho-concepts in an efficient manner. Other interesting features of MED2 include separation of database and diagnostic reasoning, temporal reasoning, and belief revision. HOST: Prof. Peter Szolovits ------------------------------ Date: Mon 18 Nov 85 08:32:01-PST From: Anne Richardson <RICHARDSON@SU-SCORE.ARPA> Subject: Seminar - Truth Maintenance, Multiple Worlds in KEE (SU) DAY December 3, 1985 EVENT Computer Science Colloquium PLACE Skilling Auditorium TIME 4:15 TITLE "Truth Maintenance and Multiple Worlds in KEE" PERSON Paul Morris, Robert Nado, Richard Fikes FROM IntelliCorp TRUTH, MAINTENANCE AND MULTIPLE WORLDS IN KEE We describe the integration of an assumption-based truth maintenance system (ATMS) into the frame-based representation facilities of the KEE system, and the use of the ATMS to implement a multiple-world context graph system for KEE. Integration into the frame system involves associating with potential slot values ATMS nodes that are used to determine in which worlds (contexts) the slot values are believed. Built-in inferences provided by the frame system, such as inheritance and the checking of value class and cardinality constraints, are recorded, when needed, as explicit justifications in the ATMS. In addition, the default reasoning capabilities of KEE have been refined and extended to take advantage of the ATMS. Tradeoffs in the integration between flexibility of use and run-time efficiency are examined. We describe the multiple-world context graph system with particular attention to an interpretation of the graph as a network of actions. In this framework, the semantics of graph merges are investigated and restrictions to ensure valid action sequences are discussed. ------------------------------ Date: Mon 18 Nov 85 08:29:15-PST From: Anne Richardson <RICHARDSON@SU-SCORE.ARPA> Subject: Seminar - Representation of Natural Form (SU) DAY November 19, 1985 EVENT Computer Science Colloquium PLACE Skilling Auditorium TIME 4:15 TITLE Perceptual Organization and the Representation of Natural Form PERSON Alex P. Pentland FROM AI Center, SRI Int'l and CSLI, Stanford PERCEPTUAL ORGANIZATION AND THE REPRESENTATION OF NATURAL FORM To understand both perception and commonsense reasoning we need a representation that captures important physical regularities and that correctly describes the people's perceptual organization of the stimulus. Unfortunately, the current representations were originally developed for other purposes (e.g., physics, engineering) and are therefore often unsuitable. We have developed a new representation and used it to make accurate descriptions of an extensive variety of natural forms including people, mountains, clouds and trees. The descriptions are amazingly compact. The approach of this representation is to describe scene structure in a manner similar to people's notion of ``a part,'' using descriptions that reflect a possible formative history of the object, e.g., how the object might have been constructed from lumps of clay. For this representation to be useful it must be possible to recover such descriptions from image data; we show that the primitive elements of such descriptions may be recovered in an overconstrained and therefore reliable manner. An interactive ``real-time'' 3-D graphics modeling system based on this representation will be shown, together with short animated sequences demonstrating the descriptive power of the representation. ------------------------------ Date: Mon 18 Nov 85 11:46:58-PST From: Fred Lakin <LAKIN@SU-CSLI.ARPA> Subject: Seminar - Setting Tables and Illustrations With Style (CSLI) Pixels and Predicates: SETTING TABLES AND ILLUSTRATIONS WITH STYLE Who: Rick Beach, Xerox PARC Where: CSLI trailers When: 1:00pm - Wednesday, November 20, 1985 Abstract: Two difficult examples of incorporating complex information within electronic documents are illustrations and tables. The notion of style, a way of maintaining consistency, helps manage the complexities of formatting both tables and illustrations. The concept of graphical style extends document style to illustrations. Observing that graphical style does not adequately deal with the layout of information leads to the study of formatting tabular material. A grid system for describing the arrangement of information in a table, and a constraint solver for determining the layout of the table are key components of this research. These ideas appear to extend to formatting other complex material, including mathematical typesetting and page layout. Several typographic issues for illustrations and tables will be highlighted during the talk. ------------------------------ Date: 18 Nov 85 23:29 EST From: Dave.Touretzky@A.CS.CMU.EDU Subject: Course - Connectionist Models (CMU) CONNECTIONIST MODELS: A SUMMER SCHOOL Sponsored by the Sloan Foundation ORGANIZERS: Geoffrey Hinton (Carnegie-Mellon University) Terrence Sejnowski (The Johns Hopkins University) David Touretzky (Carnegie-Mellon University) DATE: June 20 through 29, 1986 PLACE: Carnegie-Mellon University, Pittsburgh, Pennsylvania PURPOSE OF THE PROGRAM: The purpose of the summer school is to familiarize young researchers with current techniques in the area of connectionist models of intelligence. This includes search procedures, learning procedures, and methods for representing knowledge in massively parallel networks of neuron-like units. Application areas include vision, speech, associative memory, natural language and motor control. FACULTY: There will be six full time Tutors plus several Guest Lecturers. Tutors: Guest Lecturers: James Anderson, Brown University Jerome Feldman, U. of Rochester Dana Ballard, U. of Rochester Christof Koch, MIT Andrew Barto, U. Mass. Amherst David Rumelhart, UCSD Geoffrey Hinton, CMU David Touretzky, CMU James McClelland, CMU others to be announced Terrence Sejnowski, Johns Hopkins WHO MAY ATTEND: Participation is limited to graduate students and recent PhD's who are or will be working on connectionist models. About 40 students will be accepted. Persons who have already completed a Ph.D. degree must have done so no earlier than January 1985 to be eligible to attend. EXPENSES: There is no tuition charge. Funding from the Sloan Foundation will provide dormitory accommodations and breakfast and lunch for each attendee, plus reimbursement for a substantial portion of travel expenses. HOW TO APPLY: By March 1, 1986, send your curriculum vitae and a copy of one relevant paper, technical report, or research proposal to: Dr. David Touretzky, Computer Science Department, Carnegie-Mellon University, Pittsburgh, PA, 15213. Applicants will be notified of acceptance by April 15, 1986. ------------------------------ End of AIList Digest ********************