[comp.research.japan] Kahaner Report: Japanese Knowledge Acquisition Workshop

rick@cs.arizona.edu (Rick Schlichting) (12/29/90)

  [Dr. David Kahaner is a numerical analyst visiting Japan for two-years
   under the auspices of the Office of Naval Research-Asia (ONR/Asia).  
   The following is the professional opinion of David Kahaner and in no 
   way has the blessing of the US Government or any agency of it.  All 
   information is dated and of limited life time.  This disclaimer should 
   be noted on ANY attribution.]

  [Copies of previous reports written by Kahaner can be obtained from
   host cs.arizona.edu using anonymous FTP.]

To: Distribution
From: David K. Kahaner, ONR Asia,        [kahaner@xroads.cc.u-tokyo.ac.jp]
      B. Chandrasekaran, Ohio State U.,  [bxc@sumex-aim.stanford.edu]
      Hiroshi Motoda, Hitachi, Ltd.      [motoda@harl.hitachi.co.jp]
Re: First Japanese Knowledge Acquisition for Knowledge-Based 
    Systems Workshop (JKAW-90), 25-31 October, 1990, Kyoto & Saitama Japan.
27 Dec 1990


ABSTRACT. The first Japanese Knowledge Acquisition for Knowledge-Based 
    Systems Workshop (JKAW-90), was held October 25-31, 1990, one (open) 
    part in Kyoto, and an invited part at Hitachi's Advanced Research 
    Lab. Professor B. Chandrasekaran and Dr. Hiroshi Motoda prepared the 
    following summaries.  

INTRODUCTION. 
The Japanese Knowledge Acquisition Workshop was held as a 2-part
workshop, an "open" session in Kyoto where more than 300 attended
gathered to listen to mostly invited talks on knowledge acquisition, and
a "closed" session held at the Hitachi Advanced Research Laboratory in
Saitama, where one invited talk, a panel and a number of submitted
papers were presented.

Professor B. Chandrasekaran,
      Laboratory for AI Research
      The Ohio State University
      Columbus, OH 43210 USA
who is currently on Leave at 
      Knowledge Systems Laboratory
      Stanford University
      Palo Alto, CA
        [bxc@sumex-aim.stanford.edu]

has provided the following summary and evaluation of the open Workshop. 

Dr. Hiroshi Motoda
      Advanced Research Laboratory
      Hitachi, Ltd.
      Hatoyama, Saitama 350-03, Japan
        [motoda@harl.hitachi.co.jp]

provided a description of the closed Workshop as well as the titles and
author's affiliations for presented papers.

OVERVIEW.
As one might expect, a subject such as knowledge acquisition (KA) can 
cover a broad territory, and the talks reflected this breadth of the 
subject matter.  Nevertheless, a few categories could be identified.  
Talks by Clancey, Chandrasekaran, Gruber and McDermott in the open 
session and numerous talks in the closed session focused on the 
relationship between knowledge, tasks, and the domains.  Knowledge 
acquisition as modeling the domain for tasks of different types was a 
theme that was made explicitly by Clancey, and implicitly by others in 
this group.  The talk by Quinlan in the open session and several talks in 
the closed session dealt with the problem of learning classification
rules from examples.  Using Explanation-Based Learning for acquiring
knowledge was also discussed.  A number of talks discussed what one
might call KA support structures, such as hypertext and natural language
text processing to produce intermediate structures which could then be
accessed by the knowledge engineer for further formalization.  Interview
techniques, and computer support for interviewing, also attracted
attention. In what follows, Chandrasekaran provides a summary of the
proceedings of the open part of the workshop.  Following that is a short
summary of the closed workshop, from H. Modota. A list of the papers
presented and authors names and addresses are appended.

CONCLUSION. (Chandrasekaran)
Part of the research in knowledge acquisition closely follows research
in the more general issues of knowledge systems.  Thus a major part of
the material covered in this workshop reflected some of the major
advances in knowledge systems research of the last several years, such
as task-specific architectures, classification problem solving in
particular and compilation from deep models.  There was also a
substantial interest in the knowledge acquisition front-end itself: how
to simplify the process of getting the knowledge into the machine.
Interviewing, analysis of natural language text, and a variety of
learning techniques  are the major dimensions of research in this
direction.  Finally, it is very clear that the Japanese have made a
major commitment to knowledge systems, and the breadth of applications,
the degree of commitment and  progress reported by Japanese industry,
government, academia in the knowledge systems area were very impressive
to this visitor.

DETAILED SUMMARY (OPEN SESSION).
The Kyoto part of the workshop started with a couple of tutorials by 
Brian Gaines of the University of Calgary and John Boose of Boeing AI 
Center, both of whom have been very active in putting together a series 
of international knowledge acquisition workshops.  Gaines has developed a 
comprehensive model of the social and organizational aspects of knowledge 
acquisition.  He used this framework to motivate the issues, as he saw 
them, in knowledge acquisition.  In his view, the knowledge acquisition 
bottleneck is still very much with us: we still lack the variety of tools 
needed to help knowledge engineers transfer knowledge to shells, shells 
are still quite limited in the kinds of tasks they can help us with, and 
we still have a variety of problems in validating, maitaining and 
upgrading knowledge bases.  Nevertheless there has been considerable 
progress as well. The first generation of KA tools were based on human 
interviewing.  The second generation saw the emergence of a number of 
computer-based tools that helped in the acquisition process.  In his 
view, the so-called third generation tools will provide KA environments 
supporting a wide range of complementry tools and techniques.  

John Boose's tutorial surveyed the development of KA technology by using 
the concept of mediating representations.  By mediating representations 
he means problem modeling languages that help bridge the gap between 
experts and computer implementations.  In a sense most of the research in 
expert systems and KA in the last decade can be said to be in this area.  
He surveyed manual and computer based methods and tools.  In the manual 
category are a variety of ideas such as brainstorming, interviewing, 
protocol analysis techniques.  Semi-automated methods include shells or 
KA interfaces that use the structure of the generic problem (e.g, 
Heracles for heuristic classification, CSRL for hierarchical 
classification, Acquinas for general analysis problems, and so on) to 
help structure and represent knowledge.  In addition to the tools that 
support KA for various types of problems are various semi-automated tools 
that support knowledge acquisition more generically: tools to support 
interviewing, textual reprsentation (e.g., hypercards), modeling tools, 
tools to support acquisition of knowledge from multiple experts, and so 
on. Finally there is the body of research and tools that help in fully 
automated KA: tools based on machine induction for concept learning 
(important for classification problems).  

The invited talks started with B. Chandrasekaran's presentation on 
knowlege acquisition for real-time (RT) problems.  His work has focused 
on the close relation between the structure of the task and the knowledge 
needed for it.  He has been part of the movement in knowlege-based 
systems that has emphasized the task-structure as the mediating agency in 
knowledge acquisition and knowledge representation.  He applied that 
point of view to the design of knowledge systems for RT problems.  He 
started by noting that all problem solving works by using models of the 
world, and to the extent that these models cannot be guaranteed complete 
or correct, no problem solving process can guarantee that solutions 
generated by it will be correct.  In RT problems, the goal of problem 
solving is to generate action sequences that will keep the environment in 
desired states in the face of changes and disturbances.  Chandrasekaran 
noted that since success in this enterprise cannot be guaranteed, the 
best that can be done is to know how to abandon goals that are not being 
achieved and replace them with goals that are more likely to be 
achieved.  Thus all realistic RT problems have a built-in goal structure 
which helps in making goal abandonment and substitution decisions 
rapidly.  Chandrasekaran also noted that power in successful RT problem 
solving in humans comes about, not from complex reasoning at run-time 
about resources or by various forms of "meta-reasoning", but by avoiding 
complex run-time reasoning altogether.  Instead, power arises from four 
specific sources: good design of sensor and action systems so that direct 
mappings from observations to internal states of importance, and from 
internal states to needed ations can be made as directly as possible; a 
pre-compiled body of knowledge that helps to order goals in terms of 
priorities and preferences, so that goals can be abandoned and modified 
as needed; and another body of knowledge that helps to synthesize action 
sequences, i.e., plans.  Chandrasekaran proposed that this form of task 
analysis is very useful for knowledge acquisition since it identifies 
what types of knowledge to seek in the domain.  

William Clancey proposed a very general framework for qualitative 
modeling of domains and tasks.  Expert systems can be viewed as programs 
that use and construct models of some system in the world so that it can 
be assembled, repaired, controlled, etc. In his view, the major difference 
between conventional computer programs and expert systems is that the 
models in latter systems represent processes and structures by relational 
networks.  Control knowledge for constructing such a model can be 
described as operators that construct a graph linking processes and 
structures causally, temporally, spatially, by subtype, etc.  Clancey did 
not think that expert systems are models of intelligence as much as they 
provide a new and powerful way to model systems and processes.  He also 
proposed that "recurrent macrostructures" in such models-- both of 
objects and processes -- can be shared.  He related a wide body of work -
- Heracles-DX, Blackboard-ACCORD, generic tasks, and role-limiting 
methods in a unified framework.  

Hiroshi Motoda presented a joint research program whose aim is to build 
interviewing systems for KA.  Their system supports KA in the domain of 
logical design of databases.  The task description is used by the authors 
to give the interview system substantial capabilities for knowledge 
analysis.  In particular, they can identify which knowledge is lacking or 
erroneous by actually solving a problem, and refine the knowledge base as 
it goes through a problem solving process.  It helps in making ambiguous 
knowledge explicit by forcing the user to think about the way a problem 
is being solved, and is able to handle requirements that were not thought 
of in the beginning.  The authors then proposed a general architecture 
for KA based on their experience with the above KA system.  The proposed 
architecture is intended to incorporate machine learning techniques, 
especially by failure analysis.  

Quinlan is known as the originator of ID3, a very successful inductive 
learning system for learning classification rules from examples.  In his 
talk he reviewed the progress of such approaches, and the wide range of 
applications in which they have been used.  One of the deficiencies of 
ID3 and related algorithms is that they are not applicable for structured 
data, i.e., data with complex relationship to other data.  Instead of a 
data set where each object has a fixed number of attributes, in the more 
complex example, each object may have an open-ended number of attributes 
and also may be related to other objects by means of an open-ended number 
of relations.  He proposed the use of first-order logic to describe the 
relations.  He described some experiments on learning such relations.  

Shigenobu Kobayashi presented a jointly developed survey of the work on 
KA in Japan.  Not surprisingly, Japanese research in this area covered a 
wide territory.  It was clear to me that knowledge systems are a major 
focus of research in Japan.  They reported on a recent study by Japanese 
Information Processing Center of a number of expert system projects in 
Japan, and the lessons for KA that resulted from this study.  Task 
analysis and identification of the problem solving functions involved in 
the task are a hard first step.  Domain models can be acquired in the 
form of generic prototypical libraries without waiting for the system 
development stage.  No uniform formal methods emerged as the most 
appropriate.  This report proposed a general architecture for KA with 
components to support different requirements of the different stages in 
the Life Cycle.  

It appears that a number of systems are being developed to support the 
interview process.  Quite a bit of interest seems to exist in guiding 
knowledge acquisition by task-oriented methods, i.e., by methods that 
take advantage of the relation between knowledge and the structure of the 
task.  Examples of these methods are the Generic Task methods of this 
author, and the Heuristic Classification work of Clancey.  

Another idea that has been discussed in the US for a decade or more is 
knowledge compilation.  (My work with Sanjay Mittal was one of the 
earliest in this area, which I followed up with work on compiling 
diagnostic knowledge from deep models. The DART project of Genesereth is 
another fore-runner.)  It appears that in Japan quite a few projects are 
investigating this approach to KA.  Qualitative simulation and 
explanation-based learning are being investigated for compiling high-
level rules from structural models of devices.  

They reported on a substantial body of research on KA from text.  A 
system called hmU performs automatic model construction from hardware 
manuals for CAD systems.  Research on acquiring concepts from natural 
language texts was also reported.  

There is extensive research on machine learning techniques.  Inductive 
and similarity learning techniques were reported.  A variety of research 
on Explanation-Based Learning was discussed.  A new technique called 
Frustration-Based Learning by Motoda was discussed.  FBL uses frustrated 
states during problem solving processes as motive forces for learning.  
It associates a frustration element with information relevant to it.  
Case-based reasoning as a way of acquiring knowledge also seems to be an 
active area of research.  

A major initiative is worth mentioning.  Advanced Software Technology and 
Mechatronics Research Institute of Kyoto is setting up a project on a 
large-scale engineering knowledge base under collaboration with academic 
and industrial research organizations.  

John McDermott reported on the research by him and his colleagues at DEC 
on making applications programming easier.  They are building three 
tools: Spark, Burn and Firefighter.  To quote from McDermott: "Spark 
interviews a developer and on the basis of what the developer says about 
the information available for and the results desired from a task, Spark 
will select, from a library of pre-defined computational mechanisms, one 
or more mechanisms that can collectively produce the results desired on 
the basis of the information available.  Spark will then configure those 
mechanisms into a problem solving method.  
  
"Burn will retrieve, from a library of KA tools, the tool associated with 
each of the mechanisms selected by Spark.  It will use those tools to 
elicit expertise from the developers and produce an application program 
by encoding that expertise in a way that  will allow it to be used by the 
problem-solving method previously configured by Spark.  

"Whenever the application program is used by one of the developers, 
Firefighter will ask whether the program has performed correctly and if 
it has not, will invoke either Burn or Spark to add knowledge to the 
program or modify the problem solving method.." 

According to McDermott, Spark takes its inspiration from the generic task 
work of Chandrasekaran and from Bennet's Roget program.  He expects that 
the insights that will emerge from Spark are: 

Based on the characteristics of a task (at the non-programmer level), 
appropriate pre-defined computational mechanisms can be identified and 
configured to perform the task, and the number of such mechanisms need 
not be large to cover a large number of interesting applications.  

Burn is based on the idea that KA can be guided by the problem solving 
methods and mechanisms.  Firefighter is a relatively simple 
"diagnostician" of the problem solving method or knowledge, and uses a 
variety of techniques to "fix" the knowledge or the method.  

Tom Gruber described two knowledge acquisition problems that require the 
domain expert to formalize and operationalize knowledge that is currently 
expressed only informally.  One is the problem of acquiring strategy: how 
to elicit strategic knowledge from experts without requiring them to 
design procedures.  The other is the problem of device teleology: how to 
elicit explanations of the "purpose" of a device (i.e.  what functions 
are served by aspects of its design).  A technique called justification-
based knowledge acquisition, developed for the first problem, was used by 
the knowledge acquisition tool ASK.  Gruber discussed how ASK acquired 
strategies for diagnostic test selection in a medical domain by asking 
experts to justify specific choices in a diagnostic context.  The 
justification language is a user-extensible set of features that abstract 
states of the diagnosis in a running knowledge system.  Then he described 
a generalization of the technique that could be used to elicit design 
rationale for electromechanical devices.  In the proposed system, the 
user justifies a design decision by demonstrating the behaviors enabled 
or prevented by the designed structure using a qualitative simulation, 
comparing alternative designs in the same simulation context.  Again, 
features that justify one design decision over another are states of a 
running computer model of the process--in this case, a simulation instead 
of a diagnosis.  

The human uses his access to the external world and his richer knowledge 
base by selecting what is relevant among the possibilities afforded by 
the computational model, and providing new vocabulary terms in terms of 
old terms.  Since the knowledge is stated in the language of the machine, 
and the machine is made to formulate it in the context of its task, the 
additional knowledge is always operational.  

The final talk of the open conference was by Alain Rapaport of Neuron 
Data who outlined an architecture that his company is developing to 
integrate a large number of different types of problem solving and KA 
modules.  

BRIEF SUMMARY (CLOSED SESSION). 
Approximately forty people attended the closed meeting.  Papers
presented appear in the proceedings listing below.  Discussions centered
around two main topics - "What is knowledge?" and " Knowledge Reuse."

Many metaphors were used to describe knowledge and knowledge
acquisition.  All of them generated controversy.  Examples include the
"chicken soup," "filtering process," and "bottleneck" metaphors for
knowledge acquisition.  A comment was that it might be better to avoid
distracting metaphor discussions and concentrate on interesting work.

The theme of knowledge reuse raised during the open meeting dominated
many of the presentations and discussion.  Many systems are attempting
to reuse knowledge at different levels to gain leverage during knowledge
acquisition (task knowledge, problem- solving knowledge, task
ontologies).  A comment was that it might not be realistic to expect the
work to get very far very quickly since most success in knowledge-based
systems seems to come from domain-specific knowledge, not from
high-level reusable domain templates or procedures.


----------------------------------------------------------------
JKAW-90 Proceedings

Proceedings are available as Proceedings of the First Japanese
Knowledge Acquisition for Knowledge-Based Systems Workshop (JKAW '
90), Motoda, H., Mizoguchi, R., Boose, J., and Gaines, B. (Eds.),
Ohmsha Ltd., 3-1 Kanda Nishiki-cho, Chiyoda-ku, Tokyo 101, Japan.

Papers in the proceedings include:

Invited Papers

Gaines, B., Shaw, M. (1990).  Foundations of knowledge acquisition,
Proceedings of the First Japanese Knowledge Acquisition for
Knowledge-Based Systems Workshop (JKAW '90), Ohmsha Ltd.: Tokyo, 
pp. 3-24.

Boose, J. (1990).  Knowledge acquisition tools, methods, and
mediating representations, Proceedings of the First Japanese
Knowledge Acquisition for Knowledge-Based Systems Workshop (JKAW '
90), Ohmsha Ltd.: Tokyo, pp. 25-64.

Clancey, W. (1990).  Implications of the system-model-operator
metaphor for knowledge acquisition, Proceedings of the First
Japanese Knowledge Acquisition for Knowledge-Based Systems Workshop
(JKAW '90), Ohmsha Ltd.: Tokyo, pp. 65-80.

Kawaguchi, A., Motoda, H., Mizoguchi, R. (1990).  An architecture
of knowledge acquisition by interview based on dynamic analysis,
Proceedings of the First Japanese Knowledge Acquisition for
Knowledge-Based Systems Workshop (JKAW '90), Ohmsha Ltd.: Tokyo, 
pp. 81-96.

Quinlan, J. R. (1990).  Inductive knowledge acquisition from
structured data, Proceedings of the First Japanese Knowledge
Acquisition for Knowledge-Based Systems Workshop (JKAW '90), Ohmsha
Ltd.: Tokyo, pp. 97-112.

Kobayashi, S., Terano, T., Motoda, H., Mizoguchi, R. (1990).
Research activities of knowledge acquisition and learning in Japan,
Proceedings of the First Japanese Knowledge Acquisition for
Knowledge-Based Systems Workshop (JKAW '90), Ohmsha Ltd.: Tokyo, 
pp. 113-133.

McDermott, J., Dallemagne, D., Klinker, G., Marques, D., Tung,
Proceedings of the First Japanese Knowledge Acquisition for
Knowledge-Based Systems Workshop (JKAW '90), Ohmsha Ltd.: Tokyo, 
pp. 134-147.

Gruber, T. (1990).  Justification-Based Knowledge Acquisition,
Proceedings of the First Japanese Knowledge Acquisition for
Knowledge-Based Systems Workshop (JKAW '90), Ohmsha Ltd.: Tokyo, 
pp. 148-158.



Rule Induction

Leung, K., Wong, M. (1990).  AKARS-1: An automatic knowledge
acquisition and refinement system, Proceedings of the First
Japanese Knowledge Acquisition for Knowledge-Based Systems Workshop
(JKAW '90), Ohmsha Ltd.: Tokyo, pp. 161-174.

Tsujino, K., Takegaki, M., Nishida, S. (1990).  A knowledge
acquisition system that aims at integrating inductive learning and
explanation-based learning, Proceedings of the First Japanese
Knowledge Acquisition for Knowledge-Based Systems Workshop (JKAW '
90), Ohmsha Ltd.: Tokyo, pp. 175-190.

Medow, M., Travis, L. (1990).  Novice: Getting textbook knowledge
into expert systems, Proceedings of the First Japanese Knowledge
Acquisition for Knowledge-Based Systems Workshop (JKAW '90), Ohmsha
Ltd.: Tokyo, pp. 191-206.

Hagiwara, K. (1990).  A new approach to classification knowledge
refinement, Proceedings of the First Japanese Knowledge Acquisition
for Knowledge-Based Systems Workshop (JKAW '90), Ohmsha Ltd.: Tokyo, 
pp. 207-218.

Webb, G. (1990).  Rule optimization and theory optimization:
Heuristic search strategies for data driven machine learning,
Proceedings of the First Japanese Knowledge Acquisition for
Knowledge-Based Systems Workshop (JKAW '90), Ohmsha Ltd.: Tokyo, 
pp. 219-232.



Hypertext

Cunningham, P., Fujisawa, H., Hederman, L., Cummins, F. (1990).  A
combined approach to text retrieval using concept networks linked
to hypertext, Proceedings of the First Japanese Knowledge
Acquisition for Knowledge-Based Systems Workshop (JKAW '90), Ohmsha
Ltd.: Tokyo, pp. 235-248.

Langendorfer, H., Schreiweis, U., Hofmann, M. (1990).  Knowledge
acquisition with a special hypertext system, Proceedings of the
First Japanese Knowledge Acquisition for Knowledge-Based Systems
Workshop (JKAW '90), Ohmsha Ltd.: Tokyo, pp. 249-258.



EBL/CBR

Nakamura, K., Kobayashi, s. (1990).  Knowledge acquisition from
machine adjustment cases using causal model and operationality
criteria, Proceedings of the First Japanese Knowledge Acquisition
for Knowledge-Based Systems Workshop (JKAW '90), Ohmsha Ltd.: Tokyo, 
pp. 261-276.

Mizoguchi, R., Matsuda, K., Nomura, Y. (1990).  ISAK: Interview
system for acquiring design knowledge - A new architecture of
interview systems using examples, Proceedings of the First Japanese
Knowledge Acquisition for Knowledge-Based Systems Workshop (JKAW '
90), Ohmsha Ltd.: Tokyo, pp. 277-286.



Interview

Hori, K., Ohsuga, S. (1990).  Assisting the articulation of the
mental world, Proceedings of the First Japanese Knowledge
Acquisition for Knowledge-Based Systems Workshop (JKAW '90), Ohmsha
Ltd.: Tokyo, pp. 289-300.

Bergadano, F., Giordana, A., Saitta, L. (1990).  Automated versus
manual knowledge acquisition: A comparison in a real domain,
Proceedings of the First Japanese Knowledge Acquisition for
Knowledge-Based Systems Workshop (JKAW '90), Ohmsha Ltd.: Tokyo, 
pp. 301-314.

Taki, H., Tersaki, S. (1990).  A proposal guided knowledge
acquisition support system, Proceedings of the First Japanese
Knowledge Acquisition for Knowledge-Based Systems Workshop (JKAW '
90), Ohmsha Ltd.: Tokyo, pp. 315-330.

Tijerino, Y., Kitahashi, T., Mizoguchi, R. (1990).  A task analysis
interview system that uses a problem-solving model, Proceedings of
the First Japanese Knowledge Acquisition for Knowledge-Based
Systems Workshop (JKAW '90), Ohmsha Ltd.: Tokyo, pp. 331-344.

Yoshida, K., Motoda, H. (1990).  Hierarchical knowledge
representation based on approximations, Proceedings of the First
Japanese Knowledge Acquisition for Knowledge-Based Systems Workshop
(JKAW '90), Ohmsha Ltd.: Tokyo, pp. 345-360.



Miscellaneous

Barfoursh, A., Rogers, M. (1990).  A knowledge acquisition method
for distributed information systems: A multi-database approach,
Proceedings of the First Japanese Knowledge Acquisition for
Knowledge-Based Systems Workshop (JKAW '90), Ohmsha Ltd.: Tokyo, 
pp. 363-381.

Spirgi, S., Probst, A., Wenger, D. (1990).  Knowledge acquisition
in a development methodology for knowledge-based applications,
Proceedings of the First Japanese Knowledge Acquisition for
Knowledge-Based Systems Workshop (JKAW '90), Ohmsha Ltd.: Tokyo, 
pp. 382-397.

Tajima, M. (1990).  A framework of learning system RLS, Proceedings
of the First Japanese Knowledge Acquisition for Knowledge-Based
Systems Workshop (JKAW '90), Ohmsha Ltd.: Tokyo, pp. 398-409.



Papers not presented at the workshop

Gaines, B. (1990).  Knowledge representation servers: A generic
technology for knowledge acquisition and knowledge-based systems,
Proceedings of the First Japanese Knowledge Acquisition for
Knowledge-Based Systems Workshop (JKAW '90), Ohmsha Ltd.: Tokyo, 
pp. 413-430.

Mo, D. (1990).  Acquiring text editing programs from examples,
Proceedings of the First Japanese Knowledge Acquisition for
Knowledge-Based Systems Workshop (JKAW '90), Ohmsha Ltd.: Tokyo, 
pp. 431-447.


Names and addresses of authors in JKAW-90. 
--------------------
Ahmad A. Barfoursh 
 University of Bristol
 Department of Computer Science
 Queen's Building, Room 0.65
 University Walk, Bristol BS8 1TR
 England
 ahmad@compsci.bristol.ac.uk

John Boose 
 Advance Technology Center
 Boeing Computer Services, 
 Bldg. 33.07
 2760 160th Ave. SE
 Bellevue, Washington,
 USA 98008
 john@atc.boeing.com

B. Chandrasekaran 
 Computer and Information Science Department 
 Ohio State University
 Columbus, OH 43210-1277
 USA
 chandra@cis.ohio-state.edu
 bxc@sumex-aim.stanford.edu

William J. Clancey 
 Institute for Research on Learning
 2550 Hanover St., Palo Alto,
 CA 94304
 USA
 clancey.pa@xerox.com

Padraig Cunningham 
 Hitachi Dublin Laboratory
 O'Reilly Institute
 Trinity College
 Dublin 2
 Ireland
 cnnnghmp@vax1.tcd.ie

Hiromichi Fujisawa 
 Central research Laboratory 
 Hitachi, Ltd.
 1-280 Higashi-Koigakubo,
 Kokubunji, Tokyo 185
 Japan
 fujisawa@crl.hitachi.co.jp

Brian Gaines
 Department of Computer Science
 University of Calgary
 2500 University Dr. NW
 Calgary, Alberta, 
 Canada T2N 1N4        
 gaines@cpsc.ucalgary.ca

Attilio Giordana 
 Dipartimento di Informatica
 Universita' di Torino
 C.so Svizzera 185, 10149
 Torino
 Italy
 saitta@itoinfo.bitnet

Thomas Gruber 
 Knowledge Systems Lab
 Deaprtment of Computer Science
 Stanford Unioversity
 701 Welch Road, Building C
 Palo Alto, CA 94304
 USA
 gruber@sumex-aim.stanford.edu

Kenichi Hagiwara 
 The 2nd System Development Section
 The 1st System Development Department
 Systems Engineering Group
 Fujifacom Corporation
 1, Fuji-machi, Hino-shi,
 Tokyo 191
 Japan

Koichi Hori 
 Research Center for Advanced
 Science and Technology
 The University of Tokyo
 4-6-1 Komaba, Meguro-ku, Tokyo 153
 Japan
 hori@ohsuga.u-tokyo.ac.jp

Atsuo Kawaguchi 
 Advanced Research Laboratory
 Hitachi, Ltd.
 Hatoyama, Saitama 350-03
 Japan
 atsuo@harl.hitachi.co.jp

Shigenobu Kobayashi 
 Department of Systems Science
 Graduate School of Science and
 Engineering at Nagatsuda
 Tokyo Institute of Technology
 4259 Nagatsuda, Midori-ku,
 Yokohama 227
 Japan
 kobayasi@sys.titech.ac.jp

K. S. Leung 
 Computer Science Department
 The Chinese University of Hong Kong
 Shatin, N.T.
 Hong Kong
 ksleung%cucsd.cuhk.hk@uunet.uu.net
       
John McDermott 
 DEC, Dlb 5-3/E2, 
 290 Donald Lynch Blvd.,
 Marlboro, MA 01752
 USA
 mcdermott@airg.enet.dec.com,
 jmd@cs.cmu.edu

Mitchell Medow 
 Department of Computer Science
 University of Wisconsin-Madison
 1210 W. Dayton Street
 Madison, WI 53706
 U.S.A
 Mailing Adress
 604K Eagle Heights
 Madison, WI 53705
 U.S.A
 medow@cs.wisc.edu

Riichiro Mizoguchi 
 Institute of Scientific and
 Industrial Research
 Osaka University 
 8-1 Mihogaoka, Ibaraki, Osaka 567
 Japan
 miz@ei.sanken.osaka-u.ac.jp

Dan H. Mo 
 ACTC Technologies Inc.
 350, 6715-8 Street, N.E.
 Calgary, Alberta, 
 Canada T2E 7H7

Hiroshi Motoda
 Advanced Research Laboratory
 Hitachi, Ltd.
 Hatoyama, Saitama 350-03
 Japan
 motoda@harl.hitachi.co.jp

Kotaro Nakamura 
 Research Team of Information Science
 Engineering Research Laboratory
 Japan Tobacco Inc.
 1-31 Kurobegaoka, Hiratsuka, 
 Kanagawa 254
 Japan

John Ross Quinlan 
 Department of Computer Science
 University of Sydney, Basser
 Sydney, NSW 2006
 Australia
 quinlan@basser.cs.su.oz.au

Alain Rappaport 
 Neuron data
 444 High Street
 Palo Alto, California 94301
 USA
 atr@ml.ri.cmu.edu

Lorenza Saitta 
 Dipartimento di Informatica
 Universita' di Torino
 C.so Svizzera 185, 10149
 Torino
 Italy
 saitta@itoinfo.bitnet

Uwe Schreiweis 
 Technische Universit"{at
 Braunschweig Institut f"{ur
 Betriebssysteme und
 Rechnerverbund B "{ultenweg 
 74/75 D-3300 Braunschweig 
 Fed. Rep. Germany
 schrei@tubsibr.uucp

Susan Spirgi 
 IBM Basel 
 Hirschg"{asslein 11 
 CH-4010 Basel 
 Switzerland 

Morihiko Tajima 
 Inference Laboratory
 Intelligent Information Department
 Electrotechnical Laboratory
 1-1-4 Umezono, Tsukuba-City, 305
 Japan
 tazima@etl.go.jp

Hirokazu Taki 
 Mitsubishi Electric Corp.
 Information Systems and Electronics 
 Dev. Lab.
 ESG
 5-1-1 Ofuna, Kamakura, Kanagawa
 Japan
 taki@isl.melco.co.jp

Takao Terano 
 Graduate School of Systems Management
 The University of Tsukuba, Tokyo
 3-29-1 Otsuka, Bunkyo-ku, Tokyo 112
 Japan
 terano@gssm.otsuka.tsukuba.ac.jp

Yuri A. Tijerino 
 Institute of Scientific and
 Industrial Research, 
 Osaka University 
 8-1 Mihogaoka, Ibaraki, Osaka 567
 Japan
 yuri@ei.sanken.osaka-u.ac.jp

Katsuhiko Tsujino 
 System 4G, 
 Central Research Laboratoary,
 Mitsubishi elec. corp.,           
 8-1-1, Tsukaguchi-Honmachi,
 Amagasaki, Hyugo 661 
 Japan
 tsujino@sys.crl.melco.co.jp

Geoffrey I. Webb 
 Department of Computing and
 Mathematics 
 Deakin University Geelong,
 Victoria 3217 
 Australia
 webb@cm.deakin.oz.au

Dieter Wenger 
 Swiss Bank Coporation
 Hochstrasse, 16
 CH-4053 Basel
 Switzerland 

Kenichi Yoshida 
 Advanced Research Laboratory
 Hitachi, Ltd.
 Hatoyama, Saitama 350-03
 Japan
 yoshida@harl.hitachi.co.jp

---------------END OF REPORT----------------------------------------