LAWS@SRI-AI.ARPA (06/21/85)
From: AIList Moderator Kenneth Laws <AIList-REQUEST@SRI-AI> AIList Digest Friday, 21 Jun 1985 Volume 3 : Issue 80 Today's Topics: Seminars - A General Machine-Learning Mechanism (GM) & Distributed Decision Procedures (IBM-SJ) & Organisms' Internal Models (CSLI) & Design Expert Systems (CMU) & Unification of Logic, Function and Frames (MIT) & Automatic Example Generation (UTexas) & Qualitative Process Theory (CSLI) & Architectures for Logic Programming (GE) & Reasoning about Programs via Constraints (MIT) & Planning as Debugging (SRI) ---------------------------------------------------------------------- Date: Wed, 19 Jun 85 15:47 EST From: "S. Holland" <holland%gmr.csnet@csnet-relay.arpa> Subject: Seminar - A General Machine-Learning Mechanism (GM) Towards a General Machine-Learning Mechanism Paul Rosenbloom Stanford University Thursday, June 27, 1985, 10:00 a.m. General Motors Research Laboratories Computer Science Department Warren, Michigan Machine learning is the process by which a computer can bring about improvements in its own performance. A general machine-learning mechanism is a single mechanism that can bring about a wide variety of performance improvements (ultimately all required types). In this talk I will present some recent progress in building such a mechanism. This work shows that the combination of a simple learning mechanism (chunking) with a sophisticated problem-solver (SOAR) can yield: (1) practice speed-ups, (2) transfer of learning between related tasks, (3) strategy acquisition, (4) automatic knowledge-acquisition, and (5) the learning of general macro-operators of the type used by Korf (1983) to solve Rubik's cube. These types of learning are demonstrated for traditional search-based tasks, such as tic-tac-toe and the eight puzzle, and for R1-SOAR (a reformulation of a portion of the R1 expert system in SOAR). This work has been pursued in collaboration with John Laird (Xerox PARC) and Allen Newell (Carnegie-Mellon University). -Steve Holland ------------------------------ Date: Wed, 19 Jun 85 16:25:22 PDT From: IBM San Jose Research Laboratory Calendar <calendar%ibm-sj.csnet@csnet-relay.arpa> Reply-to: IBM-SJ Calendar <CALENDAR%ibm-sj.csnet@csnet-relay.arpa> Subject: Seminar - Distributed Decision Procedures (IBM-SJ) [Excerpted from the IBM-SJ Calendar by Laws@SRI-AI.] IBM San Jose Research Lab 5600 Cottle Road San Jose, CA 95193 Tues., June 25 Computer Science Seminar 11:15 A.M. DECISION PROCEDURES Aud. A Distributed artificial intelligence is the study of how a group of individual intelligent agents can combine to solve a difficult global problem. This talk discusses in very general terms the problems of achieving this global goal by considering simpler, local subproblems; we drop the usual requirement that the agents working on the subproblems do not interact. We are led to a single assumption, which we call common rationality, that is provably optimal (in a formal sense) and which enables us to characterize precisely the communication needs of the participants in multi-agent interactions. An example of a distributed computation using these ideas is presented. M. Ginsberg, Stanford University Host: J. Halpern (HALPERN@IBM-SJ) ------------------------------ Date: Wed 19 Jun 85 17:02:36-PDT From: Emma Pease <Emma@SU-CSLI.ARPA> Subject: Seminar - Organisms' Internal Models (CSLI) [Excerpted from the CSLI Newsletter by Laws@SRI-AI.] *NEXT* THURSDAY, June 27, 1985 2:15 p.m. CSLI Seminar Redwood Hall ``An Organism and Its Internal Model of the World'' Room G-19 Pentti Kanerva, CSLI Discussion led by Alex Pentland ABSTRACT OF NEXT WEEK'S SEMINAR ``An Organism and Its Internal Model of the World'' There is a glaring disparity in how children and computers learn things. By and large, children are not instructed explicitly but learn by observation, imitation, and trial and error. What kind of computer architecture would allow a machine to learn the way children do? In the model I have been studying, an organism is coupled to the world by its sensors and effectors. The organism's mind-ware consists of a relatively small focus and a large memory. The sensors feed information into the focus, the effectors are driven from the focus, the memory is addressed by the contents of the focus, the contents of the focus are stored in memory, and the memory feeds information into the focus. The contents of the focus at a moment account for the subjective experience of the organism at that moment. The function of the memory is to store a model of the world for later reference. The memory is sensitive to similarity in that approximate retrieval cues can be used to retrieve exact information. It is dynamic in that the present situation (its encoding) brings to focus the consequences of similar past situations. The model sheds light on the frame problem of robotics, and it appears that a robot built according to this principle would learn by trial and error and would be able to plan actions and to perform planned sequences of actions. Reading: ``Parallel Structures in Human and Computer Memory,'' available from Susi Parker at the Ventura Hall receptionist desk and on line as <PKANERVA>COGNITIVA.PAPER at SU-CSLI.ARPA. --Pentti Kanerva ------------------------------ Date: 20 Jun 85 11:16:01 EDT From: Mary.Lou.Maher@CMU-RI-CIVE Subject: Seminar - Design Expert Systems (CMU) DESIGN RESEARCH CENTER BI-WEEKLY SEMINAR SERIES COPS - A Concurrent Production System BY Luiz Alberto Villaca Leao Department Of Electrical and Computer Engineering Wednesday, June 26 at 1:30 pm in the Adamson Wing, Baker Hall Existing tools for writing expert systems are most helpful when one wants to emulate a single human expert working alone and without the aid of large number crunching programs. Few engineering problems fit this template. Rather, they tend to require multiple experts, working concurrently and supported by numbers of CAD, CAM and other tools. COPS has been designed with these requirements in mind. It is an interpreter of a superset of the OPS5 language. It provides the means for implementing multiple blackboards that integrate cooperating, concurrent expert systems, running in a distributed network of processors. ------- Refereshments will be served at 1:15 ------------------------------ Date: Thu 20 Jun 85 10:49:56-EDT From: Monica M. Strauss <MONICA%MIT-OZ@MIT-MC.ARPA> Subject: Seminar - Unification of Logic, Function and Frames (MIT) Date: Friday 21 June, 1985 Time: 11:00AM Place: 8th Floor Playroom The Uranus System -- Unification of Logic, Function and Frames Hideyuki Nakashima Electrotecnical Laboratory Tsukuba, Japan Abstract Uranus is a knowledge representation system based on the concept of logic programming. The basic computational mechanism is the same as that of the famous (or infamous!) logic programming language, Prolog, with several extensions. One important extension is the introduction of a multiple world mechanism. Uranus consists of several independent definition spaces called worlds. Worlds are combined at execution time to form a context for predicate definitions. Regarding a given world as a frame for a given concept, and predicates as slots, you have a frame-like system in logic programming. Another extension is along the lines of functional notations within the semantics of logic. Uranus has only one semantics, that of logic programming. At the same time, it has the expressive power, or convenience, of functional programming. Lazy execution of functional forms follows naturally, since portions are computed only when they are necessary for unification. A brief demonstration of the system is scheduled following the talk. Host: Gerald J. Sussman REFRESHMENTS will be served. ------------------------------ Date: Thu, 20 Jun 85 15:11:31 cdt From: briggs@ut-sally.ARPA (Ted Briggs) Subject: Seminar - Automatic Example Generation (UTexas) EGS: A Transformational Approach to Automatic Example Generation by Myung W. Kim noon Friday June 28 PAI 5.60 In the light of the important roles of examples in AI, methods for automatic example generation have been investigated. A system (EGS) has been built which automatically generates examples given a constraint specified in the Boyer-Moore logic. In EGS examples are generated by successively transforming the constraint formula into the form of an example representation scheme. Several strategies have been incorporated: testing stored examples, solving equations, doing case-analysis, and expanding definitions. Global simplification checks inconsistency and rewrites formulas to be easy to handle. EGS has been tested for the problems of controlling backward chaining and conjecture checking in the Boyer-Moore theorem prover. It has proven to be powerful -- its power is mainly due to combining efficient procedural knowledge and general formal reasoning capacity. In this talk I will present the operational aspect of EGS and some underlying principles of its design. ------------------------------ Date: Wed 19 Jun 85 17:02:36-PDT From: Emma Pease <Emma@SU-CSLI.ARPA> Subject: Seminar - Qualitative Process Theory (CSLI) [Excerpted from the CSLI Newsletter by Laws@SRI-AI.] *NEXT* THURSDAY, June 27, 1985 4:15 p.m. CSLI Colloquium Redwood Hall ``Qualitative Process Theory'' Room G-19 Ken Forbus, University of Illinois, Computer Science ``Qualitative Process Theory'' Things move, collide, flow, bend, stretch, break, cool down, heat up, and boil. Intuitively we think of the things that cause changes in physical situations as processes. Qualitative Process Theory defines simple notions of quantity, function, and process that allow interesting common-sense inferences to be drawn about dynamical systems. This talk will describe the basics of the Qualitative Process Theory, illustrate how it can be used to capture certain aspects of different models of physical phenomena, and discuss the claims it makes about causal reasoning. --Ken Forbus ------------------------------ Date: Tue, 18 Jun 85 10:04:54 EDT From: coopercc@GE-CRD Subject: Seminar - Architectures for Logic Programming (GE) Computer Science Seminar General Electric R & D Center Schenectady, N.Y. Experimental Computer Architectures for Logic Programming Prof. John Oldfield Syracuse University Tuesday, June 25 10:30 AM, Conference Room 2, Bldg. K1 (Refreshments at 10:15) ABSTRACT: Syracuse University is an established center for research in logic programming languages and their applications. In the last few years research has com- menced on ways of speeding-up the execution of logic programs by special-purpose computer architectures and the incorporation of custom VLSI components. The Syracuse Unification Machine (SUM) is a co- processor for a host computer executing LOGLISP. Unifi- cation is a fundamental and common operation in the execution of logic programs, and is highly recursive in nature. SUM speeds up unification by the combination of separate functional units operating concurrently, high- speed pattern matching and the use of content- addressable memory (CAM) techniques. Unification fre- quently requires a variable to be bound to something else, such as an expression, a constant or even another variable. The Binding Agent of SUM holds the set of current bindings in the form of a segmented stack, and with the aid of a CAM made up of custom nMOS circuits it is possible to check if a variable is already bound in under 150 nS. Binding is an operation which may be carried out concurrently in most situations, and an extra Binding Agent may be used to advantage. The Analysis Agent is another custom nMOS component which implements the pattern-matching and case-by-case analysis required. It is organized as a pipeline, and uses a state machine implemented as a PLA. (Note: the June issue of Byte Magazine contains an informative article on this work by Phillip Robinson) Notice to Non-GE attendees: It is necessary that we ask you to notify Marion White ((518) 387-6138 or WHITEMM@GE-CRD) at least two days in advance of the seminar. ------------------------------ Date: Thu, 13 Jun 1985 11:58 EDT From: DICK%MIT-OZ@MIT-MC.ARPA Subject: Seminar - Reasoning via Constraints (MIT) [Forwarded from the MIT bboard by SASW@MIT-MC.] Tuesday, June 18 8th Floor Playroom 4:00PM REASONING ABOUT PROGRAMS VIA CONSTRAINT PROPAGATION Thomas G. Dietterich Department of Computer Science Oregon State University This talk describes a program reasoning system (PRE) and its application to problems of incremental program development in machine learning. PRE solves the following problem: given a program (in a modified Programmer's Apprentice notation) with tokens labeling some of the ports in the program, find the set of possible "interpretations" of the program. That is, compute the set of executions consistent with the given information. The characterization of these executions should be succinct and their computation should be efficient. To perform this task, PRE applies constraint propagation methods. The talk will focus on (a) modifications made to the P.A. notation, (b) techniques introduced to handle failure of local propagation, and (c) strategies for resolving the frame problem. PRE is part of a larger system, EG, whose task is to form (procedural) theories of the UNIX operating system through experimentation and observation. EG's theories take the form of programs, and PRE is applied to perform tasks of data interpretation and goal regression. Refreshments will be served. Host: Richard C. Waters ------------------------------ Date: Tue 18 Jun 85 19:27:09-PDT From: LANSKY@SRI-AI.ARPA Subject: Seminar - Planning as Debugging (SRI) PLANNING AS DEBUGGING Reid Simmons -- MIT AI Lab / SPAR 11:00 am, Monday, June 24 Room EJ232, SRI International We are currently building a domain independent planner which can represent and reason about fairly complex domains. The first part of the talk will focus on the representations used and the rationale for choosing them. The planner uses explicit temporal representations, based on time points and the notion of "histories". It also extends the traditional precondition/postcondition representation of actions to include quantification, conditionals and the ability to reason about cumulative changes. The second part of the talk will focus on techniques to organize and control the search for a plan. We view planning as "debugging a blank sheet of paper". We correct a bug (ie. unachieved goal) by changing one of the underlying assumptions in the plan which are responsible for the bug. This problem solving approach combines backtracking with traditional planning techniques, giving the planner the potential for finding a solution with much less search. We also present a simple, but effective, technique for choosing which plan modification to pursue, based on maintaining a complete goal structure of the plan. This planner has been partially implemented and tested on traditional blocks-world and register-transfer examples. It is currently being applied to the problem of geologic interpretation and to diagnosis of chip manufacturing problems. ------- ------------------------------ End of AIList Digest ********************