LAWS@SRI-AI.ARPA (08/09/85)
From: AIList Moderator Kenneth Laws <AIList-REQUEST@SRI-AI> AIList Digest Thursday, 8 Aug 1985 Volume 3 : Issue 104 Today's Topics: Seminars - Flexible Planning (SRI) & KADS Methodology for Knowledge Acquisition (BBN) & Parallelism in Logic Programs (SU) & Expert System Toolkit (SU) & Experiments with Belief Resolution (SRI) & Purpose-Directed Analogy (Rutgers) ---------------------------------------------------------------------- Date: Wed 31 Jul 85 16:55:51-PDT From: LANSKY@SRI-AI.ARPA Subject: Seminar - Flexible Planning (SRI) How to Plan an Action When You Don't Know What to Do: A Logic of Knowledge, Action, and Communication Leora Morgenstern New York University, SRI-AIC 11:00 AM, Monday, August 5 SRI International, Building E, Room EJ232 Most AI planners work on the assumption that they have complete knowledge of their problem domain and situation, so that formulating a plan really consists of searching through some pre-packaged list of action operators for an action sequence that achieves some desired goal. Real life planning rarely works this way, because we usually don't have enough information to map out a detailed plan of action when we start out. Instead, we initially draw up a sketchy plan and fill in details as we proceed and gain more exact information about the world. This talk will present a formalism that is expressive enough to describe this flexible planning process. The talk will consist of 5 (hopefully short) parts: 1. Motivation for a flexible logic of knowledge, action, and communication, 2. Discussion of Bob Moore's modal logic of knowledge and action, its advantages, and its limitations with respect to a robust theory of planning, 3. A move towards a syntactic theory of knowledge, and a discussion of the resulting paradoxes (esp. the Knower Paradox), 4. A solution to the Knower Paradox based on Kripke's solution to the Liar Paradox, 5. A solution to the problem of knowledge preconditions. ------------------------------ Date: 5 Aug 1985 11:09-EDT From: AHAAS at BBNG.ARPA Subject: Seminar - KADS Methodology for Knowledge Acquisition (BBN) [Forwarded from the MIT bboard by SASW@MIT-MC.] BBN-AI Seminar, 9 August 1985, 10.30 a.m. 10 Moulton St., Large Conference Room 2nd floor KADS: a structured methodology for knowledge acquisition Bob Wielinga University of Amsterdam Current Expert System technology lacks a methodology and tools which support a structured development of commercial Expert Systems. This is particularly the case for the knowledge acquisition stage in E.S. development. KADS is the result of an attempt to develop a structured methodology for knowledge acquisition for E.S., and includes some preliminary support tools. The KADS methodology is based on the following principles: 1) a decomposition of the knowledge acquisition task, 2) the use of a number of techniques for elicitation and interpretation of verbal data, 3) the formalization of verbal data in terms of epistemological models, independent of implementation details, and 4) the use of generic models for expert problem solving behaviour to guide the knowledge analysis. The KADS methodology is being implemented in a system that will support the knowledge engineer, both in performing the analysis task and in the production of documentation. In a number of case studies KADS has been used in designing and implementing expert systems. A qualitative evaluation of these studies will be presented. ------------------------------ Date: Fri, 2 Aug 85 16:33:25 pdt From: Moshe Vardi <vardi@diablo> Subject: Seminar - Parallelism in Logic Programs (SU) AND/OR PARALLELISM IN LOGIC PROGRAMS by Simon Kasif Department of Computer Science University of Maryland, College Park MJH 352 Aug. 14, 1:00pm The separation of logic and control in logic programs has been shown to allow the programmer to write declara- tively lucid programs whose execution is determined by the interpreter. This appealing characteristic of logic program- ming spurred research directed towards diversifying the means for controlling the execution of logic programs. In particular, parallelism in logic programs may be exploited even when it is impossible to state a priori that two goals may be executed concurrently, but such an opportunity may be detected during the course of the execution. This talk will address the problem of AND/OR parallel- ism in logic programming. We describe a computational model for AND/OR parallel execution of logic programs. The model provides the primitives to describe and analyze parallel interpreters, emphasizing the data-flow among conjunctive goals. The effectiveness of our computational model is esta- blished through its ability to support both old and new com- munication paradigms for the parallel interpretation of logic programs. Several methods to implement AND/OR parallelism in logic programs are investigated based on notation developed in the model. The methods are shown to define a spectrum of communication schemes, ranging from the set intersection method where communication is eliminated altogether, through methods based on producers-consumers, where communi- cation is uni-directional and finally ending at a flexible bi-directional scheme introduced in the paper, called the Intelligent Channel. The primitives that comprise the model are used to syn- thesize two new parallel interpreters: Disjunctive System (DS) interpreters and Intelligent Channel Interpreters. The Intelligent Channel is a scheme we propose to constrain the combinatorial explosion of active processes, and to elim- inate the need to maintain a separate binding environment for every active OR-branch. ------------------------------ Date: Tue 6 Aug 85 16:07:30-PDT From: Carol Wright/Susie Barnes <SBARNES@SUMEX-AIM.ARPA> Subject: Seminar - Expert System Toolkit (SU) SIGLUNCH DATE: Friday, August 9, 1985 LOCATION: Chemistry Gazebo, betweeen Physical and Organic Chemistry TIME: 12:05 SPEAKER: Peter Jackson, Department of Artificial Intelligence, University of Edinburgh TITLE: A Flexible Toolkit for Building Expert Systems. This presentation describes the progress to date of a three-year Alvey-funded project to study, design and implement tools for building knowledge-based systems. The parties involved are Edinburgh University's Department of Artificial Intelligence and GEC's Artificial Intelligence Group at Great Baddow. The aim of the seminar is not to present finished work (the project is only six months old!), but rather to air our ideas and prejudices in the hope of attracting criticism and other kinds of feedback from the expert systems community. Our survey of current expert systems technology has led us to believe that neither shells such as EMYCIN (and derivatives) nor high-level programming languages (such as LOOPS) represent the last word in expert system building tools. The former are generally restrictive with respect to both the representation of knowledge and the specification of control, while the latter present the average programmer with a bewildering array of possibilities with little indication of how one combines different programming styles in the construction of an expert systems architecture. Thus, although there are groups of users for whom shells and AI programming languages are well-suited, we feel that there is a substantial gap in the market between the relative beginner or non-programmer, for whom the majority of commercial shells are intended, and the veteran hacker, for and by whom systems like LOOPS were developed. The alternative approach that we are currently exploring can be summed up by a number of slogans: (1) The process of choosing a logically adequate representation language and a heuristically adequate control regime and embedding these into a suitable architecture should be guided by some analysis of the task one wants one's system to perform. (2) It is worth attempting to establish a correlation between a taxonomy of expert systems tasks and representational schemes based on logical languages, with respect to both the expressiveness required by the task (e.g. modal and temporal notions, fuzziness, etc) and the control of inference (e.g. different problem solving strategies). (3) It is worth attempting to establish a similar correlation between tasks and problem solving paradigms (such as ends-means analysis, hypothesize and test, etc), with a view to helping the user decide on an architecture within which he can embed the interpretation of this chosen logical language. The problems we are currently considering include the following: (1) Is it possible to provide, along with the toolkit, an abstract architecture that can be instantiated in different ways to implement different problem solving paradigms? (2) Could one then embed different interpretations of different logics in this architecture as part of the instantiation process? (3) Could one get the behaviour associated with different problem solving paradigms from this instantiated architecture by running the logical language under different meta-level regimes? (4) Will knowledge bases created for use with one instantiation of the architecture have to be 'recompiled' for use with another instantiation? (5) How can we help a user to make the 'right' design decisions (assuming we know what these are)? We feel that this research program raises a number of very difficult issues, many of which will not be solved within the scope of the present project. Nevertheless, we also feel that practical advances in expert systems development ultimately depend upon theoretical issues of this kind being addressed, however inadequately. We still have open minds with regard to the kinds of utility that a toolkit should provide, and are always ready to talk to both the builders and users of tools in order to try and gain new insights into the problem. ------------------------------ Date: Wed 7 Aug 85 11:26:38-PDT From: LANSKY@SRI-AI.ARPA Subject: Seminar - Experiments with Belief Resolution (SRI) Experiments with Belief Resolution Kurt Konolige and Christophe Geissler SRI AI Center 11:00 AM, Monday, August 12 SRI International, Building E, Room EJ232 In recent work, Konolige developed a resolution rule for a quantified modal logic of belief. However, the rule is difficult to apply in practice, because it takes an arbitrary number of input clauses, and some instances of the rule may subsume others. In this talk we describe a solution to this problem based on a generalization of semantic attachment, that controls the growth of the search space. We have implemented the resulting version of belief resolution with Stickel's first-order connection-graph theorem prover. We present several examples of automatic reasoning about belief using this system, including a solution to the wise man problem. ------------------------------ Date: 7 Aug 85 09:37:41 EDT From: PRASAD@RUTGERS.ARPA Subject: Seminar - Purpose-Directed Analogy (Rutgers) MACHINE LEARNING SEMINAR Title: Purpose-Directed Analogy** Speaker: Smadar Kedar-Cabelli Date: Monday, August 12, 11:00 AM Place: Hill Center, room 423 Recent artificial intelligence models of analogical reasoning are based on mapping some underlying causal network of relations between analogous situations. However, causal relations relevant for the purpose of one analogy may be irrelevant for another. In this talk, I will introduce a technique which uses an explicit representation of the purpose of the analogy to automatically create the relevant causal network. I will illustrate the technique with two case studies in which concepts of everyday artifacts are learned by analogy. ** This is a dry-run for a talk being presented at the Cognitive Science Society Conference in Irvine, CA, August 15-17. ------------------------------ End of AIList Digest ********************