Copy Link
Add to Bookmark
Report
Machine Learning List Vol. 6 No. 17
Machine Learning List: Vol. 6 No. 17
Monday, June 27, 1994
Contents:
MLnetNEWS 2.3
The Machine Learning List is moderated. Contributions should be relevant to
the scientific study of machine learning. Mail contributions to ml@ics.uci.edu.
Mail requests to be added or deleted to ml-request@ics.uci.edu. Back issues
may be FTP'd from ics.uci.edu in pub/ml-list/V<X>/<N> or N.Z where X and N are
the volume and number of the issue; ID: anonymous PASSWORD: <your mail address>
----------------------------------------------------------------------
Date: Fri, 24 Jun 94 14:16:19 0000
From: MLnet Admin <mlnet@computing-science.aberdeen.ac.uk>
Subject: MLnetNEWS 2.3
MLnet NEWS 2.3 May 1994
Text-only version
**********************************
(If you want to be added to the mailing list
of the "printed" newsletter please contact us at :
mlnet@csd.abdn.ac.uk
Tel: +44 224 272304
Fax: +44 224 273422
***********************************
CONTENTS:
- The 2nd Familiarization Workshop Series at Catania
- Decisions and News from the Catania MB/TC Meetings
- Policy Statement from the Network of excellence in Machine Learning
- ECML-94 Community Meeting
- ECML-95 First Announcement and Call for Papers
- MLnet Sponsored Familiarization Workshops, Heraclion, Crete
- MLnet Workshop on Industrial Applications of Machine Learning
- MLnet summer School on Machine Learning and Knowledge Acquisition
- Technical Reports on the Catania Familiarization Workshops
- Theory Revision and Restructuring
- Knowledge Level Models of Machine Learning
- Declarative Bias
- Machine Learning and Statistics
- Focus on Machine Learning research at Daimler-Benz
- Funding for Workshops/Meetings
- Guidelines for Proposals to MLnet
- Putting Together a Machine Discoverer: Basic Building Blocks
(J.Zytkow)
- Mobal 3.0 released
- News from the University of Dortmund
- 1996 International Machine Learning Conference - Call for Proposals
- Technical Meetings between MLnet and other Networks
- News from our "data-base" of Industrial applications
- Procedures for Joining MLnet
- List of Main and Associate Nodes
- Documents available from Aberdeen
***********************************
The 2nd Familiarization Workshop Series at Catania
FAMILIARIZATION WORKSHOPS
COMMUNITY MEETING
Each person attending the Workshops had been provided with an Annual
Report, and a Policy Document; both of which were presented in
MLnet's new folder! (For the convenience of other readers who have
not seen the Policy Document it is reproduced on page 4 of this
newsletter).
Derek Sleeman introduced the session, which was organised as 4 parts:
% A brief review of the Annual Report with comments and discussion;
% A discussion of the Policy Document;
% A summary of the relevant parts of the Fourth Framework Program;
% Highlights of the Catania Management Board and Technical Board
Meetings.
The first three items are discussed here; the fourth is reported
independently on page 3.
Annual Report
The Convener Derek Sleeman, summarised the organisational structure
of MLnet with its five Technical Committees: Electronic
Communication, Industrial Liaison, Research, Training, and Written
Communication, and reported that the Management Board and Technical
Committees had met 3 times in 1992/93 at Leuven, Vienna and Blanes.
MLnet had been involved indirectly in the organisation of the 1993
European Conference on Machine Learning (ECML-93), and the associated
Workshops held in Vienna in April 1993.
MLnet organized a Familiarization workshop in Blanes (ES) in
September 1993; there were four subworkshops: "Learning and Problem
Solving" (coordinator: Maarten van Someren, Amsterdam); "Multi-
strategy Learning" (coordinator: Lorenza Saitta, Torino); "Machine
Discovery" (coordinator: Peter Edwards, Aberdeen) and "Learning in
Autonomous Agents" (coordinator: Walter van de Velde, Brussels).
Additionally at Blanes we held a Community meeting to discuss MLnet's
activities and its plans; and further we held a final overview
session where each of the workshop coordinators presented the key
issues which had arisen in their sessions. This led to a very
important exchange of ideas, where technical details, issues of
research methodology, and industrial relevance were intermingled.
For more details, see MLnet News 2,1.
Newsletter
MLnet published three Newsletters, in 1993. Regular features
include:
- reports on MLnet events
- details of future events
- conference/meeting reviews
- profile of one of the network's groups
- list of forthcoming events worldwide in ML/KA
- outline of recent (PhD) theses
We hope to include a feature shortly on industrially available ML/KA
tools.
The Newsletter is now sent to 467 persons, many of whom receive
additional copies for local/national distribution. The last two
issues have also been distributed electronically in Europe and
internationally via the MLlist at UC Irvine.
Copies of our Newsletters have been available at several of the major
AI/ML events in 1993 including ECML93 (Vienna), ML93(Amherst, Mass),
IJCAI93 (Chambery, France) and AAAI-93 (Washington)
Collection of information by Technical Comittees
The Industrial Liaison, Research and Technical Committees all sent
out questionnaires to collect information relevant to their goals.
Thus, the Industrial Liaison Technical Committee sought to determine
the ML & KA tools which are commercially available, and to identify
Industrial projects current in European industry. The Research
Committee sought to acquire information about ongoing Research
projects in European HEIs and industrial Research Laboratories.
Similarly, the Training Technical Committee sought to identify
courses taught in ML and KA in European HEIs. (These surveys were
carried out essentially by the Conveners of the corresponding
Technical Committees with some support from the other members of the
committees and the network.)
FTP and Amsterdam facilities
A Machine Learning FTP Archive has been installed at GMD (Bonn) which
provides the users with electronic access to ML related papers,
technical reports, data and software. The archive is continuously
updated and extended by the MLnet partners.
An email server is now installed at the University of Amsterdam which
provides the members of MLnet with an automated emailing list for
distributing materials and announcements to the subscribers.
For more details please see MLnet Newsletter 2,1.
Contacts with other networks and broader dissemination
ESPRIT week 1992 featured a session on Basic Research where each of
the Networks was asked via a telelink to highlight its plans and
activities; at the end of the formal session delegates visited the
NoE booths for more information. Booths were, in fact, manned for
the whole of ESPRIT week.
There was an Inter-Networks meeting in May 1993, which focused on the
provision of computer networking infrastructure for the NoEs and
specifically on a plan proposed by CABERNET that all the NoEs should
use the Andrew File system. MLnet agreed to actively investigate the
facilities which its nodes require, and the facilities provided by
the Andrew system.
There was a further meeting of Networks held in Brussels in late
September 1993 where the major topics for discussion were:
- the provision of a computer-network infrastructure for all Networks
(a continuation of the discussion held in May),
- the roles of NoEs both within their communities and as potential
policy-influencers within the EC.
- funding in Framework-4
Bob Wielinga represented MLnet at both these meetings; Bob has
produced a report on the last meeting which was presented to the
Management Board in December, 1993, and published in MLnews 2,2.
Later the same day, the several representatives of NoEs met the
Energy subcommittee of the European Parliament to informally discuss
Networks and their activities and plans. Again Bob Wielinga
represented MLnet at this meeting.
The Academic Coordinator, Derek Sleeman, has had considerable
informal contact with several of the Networks' Coordinators through
the year by email, phone and several face-to-face meetings.
Membership Applications in 1992/93
We received a considerable number (32) of initial enquiries about
MLnet. Many of these are from groups outside Europe or from very
small groups, and so the direct result is that their names have been
added to the mailing list for the Newsletter. However nine led to
formal applications to join MLnet as Associate nodes. Of the
applications received, three groups have been successful, namely the
groups at Kaiserslautern, Karlsruhe and Catania.
At this point the Convener was rightly asked to explain why the spend
in year 1 had been such a small percentage of the available monies.
His explanation included the following points:
% Projects often have a slow start thus spending more money in later
years.
% In the case of the NoE, the contract was considerably different
from standard ESPRIT contracts and so we were all determining the
"rules". Essentially, as noted elsewhere, NoEs are able to support
infra-structure and not Research or Training per se.
% Monies had been allocated to items in year 1 (such as a Summer
School and a Network Machine) but it had not been possible to
actually carry out these activities.
% The percentage of the budget allocated for the second year was
considerably higher.
It was further suggested that Networking infrastructure could be a
big item if the NoE expands, as it intends, into Central and Eastern
Europe.
Policy Document
Derek Sleeman (DS) explained the Research Technical Committee was
producing a revised State-of-the-Art report (the original was part of
the Proposal/Technical Annex). It was expected this would be
available within 2-3 months; Lorenza Saitta, as Convener of the
Research Technical Committee, was responsible for its production.
Because decisions were being made about the detail for the Fourth
Framework in early 1994, it was suggested to the Academic Coordinator
that it would be timely if MLnet produced a short Policy Document
(see page 4) which essentially included a set of themes which this
community felt should be present in the next program, together with
any comments we might have on how future programs might be better
organised and administered. These then were the principal elements
of the Policy Document produced by DS, with input from the authors of
the several sections of the State-of-the-Art Document. Several
drafts were circulated to members of the Management Board (including
David Cornwell, the Project Officer) for comment. The resulting
document has now been circulated to all nodes, presented formally to
the EC, and, together with the Annual Report, formed a major focus
for discussion which DS had with the Commission in March of this year
(more in the next section).
The policy document was then discussed extensively, when the
following points were raised:
% Data Mining should have been included as one of the "themes".
% Some felt more emphasis should have been given to Applications.
% Some felt that the themes should be structured into techniques and
applications.
This then lead naturally into a review of comments received on the
Policy Document from the EC's officials and an overview of the Fourth
Framework Program.
A Summary of the Fourth Framework Program and a visit to
Brussels
The total allocated budget for the 4th Framework Program is expected
to be around 12 billion ecu of which 2 billion ecu will be for Long
Term Research and R&D in IT. The 2 billion ecu are to fund all
aspects of IT research, i.e. not only the IT industry but also
industries which are IT users. Approximately 10% of this budget is
for Long Term Research (and this figure probably excludes Networks).
Additionally, it is believed that 2% of the total "IT" budget is to
be set aside for training and mobility which are to be administrated
by the "Basic Research" office.
The Commission is aiming for a call in September, but this is thought
to be optimistic by many involved. However, it is generally expected
that new contracts will start during '95.
DS reported that the annual report and the policy document has been
the focus of a seminar he had given in Brussels to Project Officers,
that he had had detailed discussions with policy makers in the
Directorate, and with a number of individual project officers from
both the BRA and Industry divisions. Subsequently he had had a
detailed phone call with Simon Bensasson. The EC's officials had
discussed the current plans for the next program.
The policy document included a number of themes which the Management
Board felt should be included in the next program. DS reported that
he had been reassured that officials believed that to achieve many of
the aspects inherent in the next R&D program, it would be necessary
to use Machine Learning, Knowledge Acquisition and second generation
ES technology. (The phrase Adaptive System had been used). The
program is, in fact, to be organized around 4 focussed clusters:
% Open Microprocessor Systems initiatives;
% High Performance Computing & Networking;
% Integration in manufacturing;
% Technologies for Business Processes;
and so the techniques of Adaptive Systems are seen as being
orthogonal enabling technologies.
Further, DS was told that the call for Long Term Research (formerly
Basic Research) would include an open call, and hence any theme could
be proposed. Additionally, several of the points raised about the
administration of projects were accepted in principle; these
included:
% generally having only small and well focussed consortia;
% that Long Term Research projects would consist largely of 3-year
projects as well as some 1-year projects to try out ideas. (Earlier
we had heard that LT would be composed exclusively of 1-year
projects.)
% regular call for proposals would be helpful to both industry and
academia;
% more formal links, particularly at the level of LT Research
actions, should be enhanced with Japan and America.
Additionally, DS reported on a meeting with the Director of the
Human Capital & Mobility Programme, M. de Nettancourt, (this topic
was reported at the Management Board but not at the general meeting -
it is included here for completeness)where they discussed:
% the overall structure and aims of NoEs like MLnet;
% the fact that HC&M covers all of Science and is oriented towards
the Basic Research end of the spectrum;
% the future plans for the HC&M directorate in the next Framework:
% the earlier plan was to drop the institutional fellowships
programme as these were seen as being unwieldy to administer. (We now
understand that this may be modified as several NoEs have made
representations to their national representatives.)
% there is a proposal to introduce the concept of European
Laboratory Without Walls, i.e. several groups working on joint
research (LT Research) topics. (We agreed that there was a
considerable complementarity between the ESPRIT Basic Research
Networks and the HC&M programme; the former providing the
infrastructure to support Research & Training and the HC&M program
being able to support actual researchers and research teams.)
Derek Sleeman closed the meeting by asking that people send him, or
any member of the Management Board, any other comments which might
subsequently occur to them. Also he stressed again MLnet's openness
to suggestions for events and activities which might be organised.
Luc deRaedt thanked Derek Sleeman for his many efforts on behalf of
the Network.
***********************************
Decisions and News from the Catania MB/TC Meetings
% Electronic Communication: It was agreed that an advanced FTP
service (at GMD) and an experiment in the use of the Andrew File
System (based at Amsterdam) should both be supported.
% Reports or Databases will be produced shortly by the Industrial
Liaison, Research and Training Technical Committees.
% A revised State of the Art Report will be available shortly.
% Joint ELSNET*/MLnet Workshop to be supported (dates to be
announced).
% IFIP-94; A Workshop on ML applications at IFIP is to be supported
by MLnet.
% New nodes: Daimler-Benz (main node); EDF (Electricit de France),
Prague and Oxford (associate nodes).
% ECML-95 and Familiarisation Workshops to be supported by MLnet (see
further information on pages 5-7).
% Strategic Planning: To be featured as a specific item at each of
the next 3 Management Board meetings.
% The EC has agreed to support MLnet for the third year (ie until
September 1995). For support after this date MLnet will need to
apply to the Fourth Framework.
***********************************
POLICY STATEMENT
FROM THE NETWORK OF EXCELLENCE IN MACHINE LEARNING
IMPORTANCE OF THE AREAS OF MACHINE LEARNING (ML) AND
KNOWLEDGE
ACQUISITION (KA).
Building Knowledge Bases has been identified as a major bottleneck
in producing Intelligent Systems. ML and KA have developed a number
of tools and techniques which can help considerably with this
critical phase. Additionally, techniques have been devised to refine
knowledge bases when the results produced by an intelligent system
differ from those anticipated by the domain expert.
We expect in the next decade that learning components will be
embedded in many of the IT systems produced, so that they can
progressively improve their Knowledge Bases. Further, we expect to
see major integration of AI techniques, including ML, with
traditional DP techniques. Additionally, ML is having a considerable
impact on reducing costs of developing software for controlling
dynamic systems, such as robots.
MLnet believes that our subfield is central to building tomorrow's
Intelligent Systems, and thus is central to every forward-looking
European IT company.
MLnet's CURRENT AIM
The aim of the network is to coordinate Machine Learning Research and
Development throughout Europe, to ensure that these technologies
become a pervasive force in European industry, and at the same time
to consolidate and enhance their solid scientific bases.
STRATEGICALLY IMPORTANT ISSUES
MLnet's Research Organization and Coordination Technical Committee
has recently identified the following topics of paramount importance,
to achieve the goals given above:
% The design, implementation and extensive use on demanding tasks of
Workbenches which integrate Machine Learning and Knowledge
Acquisition tools, together with Problem Solving systems (complex
tasks are likely to require a Multi-strategy learning approach).
% Ability of a KBS (Knowledge Based System) to refine its KB
(Knowledge Base) as a result of feedback received from domain
experts, or from the environment; together with the ability to tailor
its explanations to the level/sophistication of its user.
% Refinement of computer programs, given feedback from experts, and
the environment.
% Enhancing Inductive Logic Programming, which studies how to induce
first order logic formulae from observations and background
knowledge, to deal with complex induction problems in scientific
discovery, knowledge acquisition, automatic programming and deductive
databases.
% Reusable Knowledge Bases.
% The ability of a Robot to interact with its world and learn to
interpret sensor signals and actions.
% Genetic Algorithms allow vast sets of hypotheses which occur in
many domains, say such as molecular biology and genetics, to be
searched (as Genetic Algorithms allow parallelism to be exploited).
COMMENTS ON MANAGEMENT OF PROJECTS AND PROGRAMS
ESPRIT, by any standards, has been a successful vehicle for Basic
Research and Research and Development in Europe. Against great odds
it has managed to get Research groups across two major divides,
namely transnational and academic/industrial, working together. Now
that this tradition of cooperation has been established, it seems
important to ask whether this whole process could be more effective,
and whether Europe could glean some of the strong points of program
management from other industrially-advanced nations. MLnet believes
that the following points should be seriously considered for the IT
Specific Program within the Fourth Framework Programme:
In Long Term Research:
% It recommends that the consortium should normally involve only
three or four partners. But that greater interchange of ideas should
be achieved by having a number of annual contractor-only Research
meetings where all projects review their recent results, problems,
and longer term goals. Such meetings would essentially be mini-
versions of the former ESPRIT week meetings and would be very similar
in format to the meetings which the American ONR agency holds
(indeed, it might be argued that the Networks of Excellence would be
appropriate entities to organize these meetings).
% MLnet had major problems with the suggestion that in the 4th
Framework, there might be a number of short projects which would test
the validity of an idea, so that a subset of these projects might
evolve into larger and longer R & D projects.
MLnet's views are that this is not an appropriate way to view the
relationship between Long Term Research and R & D. By its very
nature, Long Term Research addresses fundamental problems in a
discipline. In our view, the role of R & D programs/projects is to
implement advanced (near market) prototypes from the insights gleaned
from Long Term Research actions.
In R & D:
Again we would advocate that consortia would normally be comprised of
four or five partners, where there would be a tight coupling between
technique-developers, industrial partners and "end-user"
organizations. (We would hope that SMEs or groupings of SMEs would
be involved in these consortia.) Again, we suggest that annual
meetings be held, involving projects which clearly have related
interests and goals, ie those that are active in the same sector.
Less frequently, say bi-annually, such meetings should overlap with
the relevant Long Term Research Annual meeting.
Other points which we would like to raise:
% it would be helpful to Research Laboratories, both academic and
industrial, if some form of rolling program, or rolling set of
program calls could be established. This would enable institutions
to plan their workloads more effectively , and more crucially would
allow contract Research staff to plan their careers more effectively.
% it would be helpful, particularly to University-based groups, if
the time between contract announcement and the start date, could be
increased to at least four months. (Currently it is often hard to
recruit appropriately highly skilled contract staff in the very short
lead time).
% MLnet would like to see more formal links, particularly at the
level of Long Term Research actions, enhanced with Japan and America.
% MLnet believes that the Networks of Excellence are an effective
mechanism for coordinating information about a sub-discipline within
Europe, and very much hope that they will continue to have a role
within the 4th Framework. Specifically we believe that MLnet's
Industrial Liaison Workshop, Familiarization Workshops and Summer
School will all be important for strengthening the community of
European Scientists/Technologists working in the areas of ML and KA;
this in turn should enhance the competitiveness of the European IT
industry.
% MLnet believes that NoEs should have an input to the Community's
planning of future frameworks and programs.
Derek Sleeman
Aberdeen
27th January 1994 / revised 15th February
***********************************
ECML-94 Community Meeting
Francesco Bergadano and Luc de Raedt introduced Derek Sleeman
(Academic Coordinator of MLnet) and Lorenza Saitta (Convener of
MLnet's Research Committee).
Derek Sleeman gave a brief overview of MLnet, what it had achieved in
the first year, and what it was planning to do in subsequent years.
One of the roles it had assumed was that of organising the annual
European Conference on ML. This was essentially coordinated by the
Research Committee; he then handed the meeting over to Lorenza Saitta
to discuss the planning of ECML-95.
Lorenza Saitta reported that in response to her request for
nominations for sites for ECML-95 she had received only one response,
namely the FORTH Institute at Heraklion in Crete; and that unless
there were any strong objections that this would be the site for
ECML-95.
Vassilis Moustakis then outlined some of the facilities which they
could offer at FORTH (new CS Dept Building, wide area network etc).
He also explained that hotel accommodation should be both varied and
reasonably priced if one avoids Easter Sunday itself. (In subsequent
discussions it was agreed that the conference would be held on 25-27
April (1995) with the Familiarisation Workshops being on the
afternoon of the 28th and all day on the 29th). Vassilis also
pointed out that it should be possible also to get charter flights at
this time of the year to Crete; these can be with or without
accommodation.
The next phase of the discussion was the selection of the Program
Chairs. After a relatively short discussion it was agreed that these
should be Nada Lavrac (Ljubljana) and Stefan Wrobel (GMD). Given
they are both active in ILP they agreed to ensure that other ML
specialities are strongly represented in the Program Committee.
Derek Sleeman concluded the meeting by proposing a very sincere vote
of thanks to this year's Program Chairs, Francesco Bergadano and Luc
de Raedt; and to all their support staff.
***********************************
ECML-95
8th EUROPEAN CONFERENCE ON MACHINE LEARNING
25Q27 April 1995, Heraklion, Crete, Greece
First Announcement and Call for Papers
General Information :
Continuing the tradition of previous EWSL and ECML
conferences, ECML-95 provides the major European forum for
presenting the latest advances in the area of Machine
Learning.
Program :
The scientific program will include invited talks,
presentations of accepted papers, poster and demo
sessions. ECML-95 will be followed by MLnet Fa-
miliarization Workshops for which a separate call for
proposals is published on page 7 of this Newsletter.
Research areas :
Submissions are invited in all areas of Machine
Learning, including, but not limited to:
abduction analogy applications of
machine learning
automated discovery case-based learning computational
learning theory
explanation-based inductive learning inductive logic
learning programming
genetic algorithms learning and multistrategy
problem solving learning
reinforcement representation revision and
learning change restructuring
Program Chairs :
Nada Lavrac (J. Stefan Institute, Ljubljana) and
Stefan Wrobel (GMD, Sankt Augustin).
Program Committee :
F. Bergadano (Italy) I. Bratko (Slovenia) P. Brazdil (Portugal)
W. Buntine (USA) L. De Raedt (Belgium) W. Emde (Germany)
J.G. Ganascia (France) K. de Jong (USA) Y. Kodratoff (France)
I. Kononenko (Slovenia) W. Maass (Austria) R.L.deMantaras (Spain)
S. Matwin (Canada) K. Morik (Germany) S. Muggleton (UK)
E. Plaza (Spain) L. Saitta (Italy) D. Sleeman (UK)
W. van de Velde (Belgium) G. Widmer (Austria) R. Wirth (Germany)
Local chair :
Vassilis Moustakis, Institute of Computer Science,
Foundation of Research and Technology Hellas (FORTH),
P. O. Box 1385, 71110 Heraklion, Crete, Greece (E-mail
ecml-95@ics.forth.gr). (Phone +30 81 229 346/302).
(Fax +30 81 229 342).
Submission of papers :
Paper submissions are limited to 5000 words. The
title page must contain the title, names and addresses
of authors, abstract of the paper, research area, a list
of keywords and demo request (yes/no). Full address,
including phone, fax and E-mail, must be given for the
first author (or the contact person). Title page must
also be sent by E-mail to ecml-95@gmd.de. If possible,
use the sample LaTeX title page that will be available
from ftp.gmd.de, directory /ml-archive/general/ecml-95.
Six (6) hard copies of the whole paper should be sent by 2
November 1994 to:
Nada Lavrac & Stefan Wrobel (ECML-95)
GMD, FIT.KI, Schloss Birlinghoven, 53754 Sankt Augustin, Germany
(email wrobel@gmdzi.gmd.de) (Phone +49 2241 14 2670/8).
(Fax +49 2241 14 2889)
Papers will be evaluated with respect to technical
soundness, significance, originality and clarity. Papers
will either be accepted as full papers (presented at
plenary sessions, published as full papers in the
proceedings) or posters (presented at poster sessions,
published as extended abstracts).
System and application exhibitions :
ECML-95 offers commercial and academic participants
an opportunity to demonstrate their systems and/or
applications. Please announce your intention to demo to
the local chair by 24 March 1995, specifying precisely
what type of hardware and software you need. We strongly
encourage authors of papers that describe systems or
applications to accompany their presentation with a demo
(please indicate on the title page).
Registration and further information :
For information about paper submission and program, con-
tact the program chairs (E-mail
ecml-95@gmd.de). For information about local ar-
rangements or to request a registration brochure, contact
the local chair (E-mail ecml-95@ics.forth.gr).
Important Dates :
Submission deadline : 2 November 1994
Notification of acceptance : 13 January 1995
Camera ready copy : 9 February 1995
Exhibition requests : 24 March 1995
Conference : 25 Q 27 April 1995
***********************************
MLnet Sponsored Familiarization Workshops, Heraklion, Crete
April 28-29, 1995
MLnet is planning to hold a series of Technical Workshops together
with some other infrastructure discussions, on 28-29 April 1995, ie,
immediately after the ECML95, which will also be held in Heraklion.
Proposals are requested for these technical workshops. Such
proposals should include:
% A detailed discussion of the subarea(s) to be covered.
% Names and addresses of the proposed Program Committee (together
with a note indicating whether the named people have agreed to act).
% Names of possible invited speakers and some indication of their
travel costs.
% Some indication of the form of the workshop (ie, split between
presentations and panels etc).
This information should be sent to Derek Sleeman preferably by
regular mail, to arrive by 8 August 1994. This item will be
discussed at the September Management Board meetings, and proposers
will be informed of the decision by mid September.
Derek Sleeman will be happy to discuss possible topics with potential
organisers.
NB: MLnet expects to have sufficient funds to support the travel to
Crete and the subsistence at the Familiarization Workshop, for
several members of each MLnet
node.
***********************************
MLnet Workshop on Industrial Applications of Machine
Learning
Dourdan, France, September 2 and 3, 1994
Organizer: Yves Kodratoff (LRI & CNRS, University of Paris-South,
Orsay, France).
Program
Friday, September 2: Overview presentations
% Ivan Bratko (JSI, Ljubljana, Slovenia) - On the state-of-the-
art
of industrial applications of ML
% Gregory Piatetski-Shapiro (GTE, USA) - DDB: data mining in data
bases
% Wilfried Achthoven (Bolesian Systems, NL) - The state of the
art
in knowledge acquisition: industrial practice with KADS, machine
learning and case-based reasoning
% Attilio Giordana (Univ. Torino, Italy) - Applications of
machine
learning to robotics
% Franz Schmalhofer (DFKI, Germany) - Unifying KA and ML for
applications
% Vassilis Moustakis (Univ. Crete, Greece) - An overview of
applications of ML to medicine
Saturday, September 3:
Special address: Setsuo Arikawa (Kyushu University, Japan) -
Knowledge acquisition from protein data by machine learning system
BONSAI
Results of ESPRIT projects:
% Nick Puzey (BAE, UK) - Industrial applications of MLT
% Pavel Brazdil (Univ. Porto, Portugal) - Industrial applications
of
STATLOG
% Attilio Giordana (Univ. Torino, Italy) - The results of BLEARN
Reports on results:
% Gend Kamp (Univ. Hamburg, Germany) - Applications of case-based
reasoning
% Pieter Adriaans (Syllogics, NL)- Application of GAs at
Syllogics
% Fabio Malabocchia (CSELT, Italy) - Machine learning at CSELT
% J rgen Herrmann (Univ. Dortmund, Germany) - Learning rules
about
VLSI-design
% Reza Nakhaeizadeh (Daimler-Benz, Germany) - Machine learning at
Daimler-Benz
% Franois Lecouat (Matra-Espace, France) - Case-based reasoning
at
Matra-Space
Demos will take place during the evenings. Participants are welcomed
to also attend the ML Summer School at Dourdan on the following week
(September 5-10). Note that a separate registration will be required
for the meeting. The topics presented in the first two days of the
Summer School will be of greatest relevance to people attending the
workshop.
Registration fee
The registration is 800 FF. Some full and partial grants for travel,
registration and accommodation are available for European students
and researchers. Applicants must send a letter stating their
motivations and a CV before June 26.
Requests for information and registration forms (see central page
of this Newsletter) are to be addressed to Dolores Caamero,
(MLSS'94), LRI, B t. 490, Universit Paris-Sud, F-91405, Orsay Cdex,
France (e-mail: mlss94@lri.fr).
***********************************
MLnet Summer School on Machine Learning and Knowledge
Acquisition
Dourdan, France, September 5-10, 1994
The summer school is organized by Cline Rouveirol (LRI, University
of Paris-South, France). Its aim is to provide training in the latest
developments in Machine Learning and Knowledge Acquisition to AI
researchers, and also to industrials who are investigating possible
applications of those techniques. The school will be held in Dourdan
(some 50 Km south of Paris). It is sponsored by CEC through the MLnet
Network of Excellence (Project 7115), and PRC-IA (Projet de Recherche
Coordonn, groupe Intelligence Artificielle).
Program
Monday Sept 5th
. Morning: Case-Based Reasoning (Agnar Aamodt, Univ. Trondheim,
Norway) [3h]
. Afternoon: Learning and Probabilities (Wray Buntine, RIACS/NASA
Ames, Moffet Field, CA, USA) [3h]
Tuesday Sept 6th
. Morning: Learning and Noise (Ivan Bratko, JSI, Ljubljana,
Slovenia) [3h]
. Afternoon: Knowledge Acquisition (Bob Wielinga, Univ.
Amsterdam,
NL) [3h]
Wednesday Sept 7th
. Morning: Integrated Architectures (Lorenza Saitta, Univ.
Torino,
Italy) [3h]
. Afternoon: Knowledge Revision (Derek Sleeman, Univ. Aberdeen,
UK)
[2h], (Stefan Wrobel, GMD, Bonn, Germany) [2h]
Thursday Sept 8th
. Morning: Knowledge Acquisition and Machine Learning (Maarten
van
Someren, Univ. Amsterdam, NL) [3h]
. Afternoon: Reinforcement Learning (L.P. Kaelbling, Brown Univ.,
USA) [3h]
Friday Sept 9th
. Morning: Inductive Logic Programming (S. Muggleton, Univ.
Oxford,
UK) [3h]
. Afternoon: Inductive Logic Programming (C. Rouveirol, LRI,
Univ.
Paris-Sud, France) [2h], (F. Bergadano, Univ. Catania, Italy) [2h]
Saturday Sept 10th
. Morning: Conceptual Clustering (G. Bisson, LIFIA, Grenoble,
France) [3h]
Invited seminars and demonstrations of software will be organized
during the evenings.
Registration fee
Some full and partial grants for travel, registration and
accommodation can be accorded to European students and researchers
and to members of PRC-IA. Applicants must send a letter stating their
motivations and a CV before June 26.
Requests for information and registration forms (see central page
of this Newsletter) are to be addressed to Dolores Caamero,
(MLSS'94), LRI, B t. 490, Universit Paris-Sud, F-91405, Orsay Cdex,
France. E-mail: mlss94@lri.fr.
***********************************
Technical Reports on the Catania Familiarization Workshops
Report on WS1: Theory Revision and Restructuring in Machine
Learning
by Stefan Wrobel
Organizer: Stefan Wrobel (GMD, Sankt Augustin, Germany)
Program Committee: Hilde Ade, Carl-Gustaf Jansson, Stefan Wrobel.
1. Workshop Topic
With the growing complexity of applications being tackled by Machine
Learning, the field has become increasingly aware that besides
approaches for the initial acquisition of knowledge bases we also
need techniques for theory revision and restructuring, i.e.,
techniques that can use existing learned or human-supplied domain
theories and can modify them to improve their correctness,
completeness, efficiency or understandability.
Traditionally, this topic has been examined in different contexts.
Revision has always been a part of incremental or hill-climbing
learning systems which keep only one current hypothesis and modify it
whenever new examples arise. More recently in ILP, revision has been
identified has an important part of approaches that learn multiple
predicates simultaneously, and incorporated as a central component of
integrated multi-strategy learning systems. Under the name of
refinement, the expert systems community has studied ways of
improving system knowledge bases over time. Finally, revision and
restructuring are also important topics in neighbouring fields of ML,
such as knowledge representation, logic programming or deductive
databases.
The goal of this familiarization workshop was to bring together the
various approaches to revision and restructuring, that are currently
being pursued, both to allow participants to learn about each others
work and to see if a common framework is emerging. The orientation of
almost all the 12 presentations held at the workshop reflects a
strong trend towards a first-order clausal logic framework as a
common basis, paralleling the current focus on ILP in Europe. In
this framework, revision is seen as the task of changing a theory by
generalizing it when positive examples are erroneously not covered,
and by specializing it when negatives examples are erroneously
covered.
2. Overview of the Presentations
The presentations in the morning each took a particular perspective
on this general framework. In the presentation by Hilde Ade, Bart
Malfait, and Luc De Raedt ("RUTH: an ILP Theory Revision System"),
the general framework was extended by allowing not only facts as
positive and negative examples, but also full clauses that are then
interpreted as integrity constraints. Francesco Bergadano and
Daniele Gunetti ("Intensional Theory Revision") showed how
intensional evaluation of clauses for revision can be made tractable
by assuming the initial program is almost correct and using a strong
search bias (clause sets). Stefan Wrobel ("Heuristic Control of
Minimal Base Revision in KRT Using a Two-Tiered Confidence Model")
concentrated on the specialization subtask and showed how heuristic
evaluation functions can be used to select among the possible minimal
revisions of a theory. Whereas Wrobel's KRT system produces minimal
specializations with exception sets, Henrik Bostroem and Peter
Idestam-Almquist ("Specialization of Logic Programs by Pruning SLD-
Trees") showed an approach that needs only standard definite clauses
for (non-minimal) specialization by unfolding.
In a second session, three approaches to revision were presented that
each use different biases to control the revision process. Marco
Botta ("KBL-2: A First Order Theory Refiner") showed how a data
structure called an LT-tree is used to represent all the information
necessary for revision of the theory. While essentially equivalent
to an AND-OR derivation tree, substitution information in an LT-tree
is computed bottom-up and kept in a database system. Floriana
Esposito, Donato Malerba, and Giovanni Semeraro ("INCR/H: A system
for Revising Logical Theories") departed from the standard framework
by considering a different notion of subsumption (theta-subsumption
under object identity) which produces more natural, but no longer
unique, LGGs. They presented a specialization operator for linked
clauses that is shown to be minimal under this definition of
subsumption. Finally, Filippo Neri ("An Approach to Knowledge
Refinement and Theory Revision") gave an overview of theory revision
processes in the WHY system, which is characterized by its separation
of the knowledge base into causal model and phenomenological theory
that are treated differently during revision.
In the afternoon session, the orientation of talks was more varied.
In the only talk on the restructuring side of the workshop, Edgar
Sommer ("Rule Base Stratification: An approach to theory
restructuring") showed how a so-called stratification algorithm can
be used to introduce intermediate concepts into a knowledge base, and
used an example knowledge base to show that this can greatly increase
understandability. Alipio Jorge and Pavel Brazdil ("Incrementality
issues in Sketch Refinement") used a surprising notion of refinement,
not referring to improvements of a theory, but to the instantiation
process of a sketch when turning it into a final hypothesis. Yutaka
Sasaki, Masahiko Haruno, and Shigeo Kaneda ("Grammar Rule Revision by
Rephrasing Unparsable Sentences") then presented a revision approach
embedded into a natural language parsing system for Japanese. They
showed that their revision method improved parsing capabilities on a
simple test set compared to the method previously used. Daniel
Borraja and Manuela Veloso ("Multiple Target Concept Learning and
Revision in Non-linear Problem Solving"), in a continuation of their
main program talk, showed how revision procedures can be used to keep
up to date conditions that recommend operators in particular
situations of a planning problem.
The last talk of the workshop by Susan Craw, Derek Sleeman, Robin
Boswell, and Leonardo Carbonara ("Is Knowledge Refinement Different
from Theory Revision?") turned out to be an ideal starting point for
a discussion about the terminology that is being used in the field to
describe the different tasks. In the end, a hierarchy of tasks ended
up on the blackboard that regarded refinement as the most general
term, subsuming revision (modifying an incorrect or incomplete
theory) and restructuring (modifying a correct and complete theory to
improve other properties like understandability). Theory revision in
turn features generalization and specialization (debugging) as
subtasks, whereas restructuring consists of both performance
enhancement (using EBL, partial evaluation) and understandability
enhancement (using operators for introducing new predicates).
(theory/knowledge) refinement
|
QQQQQQQQQQQQQQQQQQQQQQQQQQQQQQQQQQQQ
| |
revision restructuring
| |
QQQQQQQQQQQQQQQQQ QQQQQQQQQQQQQQQQQQQQQ
| | | |
specialization generalization performance understandability
(debugging) (EBL, PE)
In summary, especially because of the final discussion, I believe the
workshop has served to clarify a little bit the terminology used in
the field, and presented a lot of work on one part of the spectrum,
namely revision. For the future, it seems that restructuring should
be more of a focus, particularly given the importance of
understandability noted in the main conference by Yves Kodratoff and
Lorenza Saitta.
P.S. As announced at ECML, the proceedings (extended abstracts) of
this workshop are available via FTP and will appear as a GMD
technical report:
"Proc. MLnet Familiarization Workshop on Theory Revision and
Restructuring in Machine Learning (at ECML-94, Catania, Italy), ed.
Stefan Wrobel, Arbeitspapiere der GMD, GMD, Pf. 1316, 53754 Sankt
Augustin, Germany, 1994. Available via FTP from ftp.gmd.de as /ml-
archive/MLnet/Catania94/theory-revision.ps.gz."
If you cannot access FTP, or would like a printed copy, send mail to
stefan.wrobel@gmd.de.
***********************************
Technical Reports on the Catania Familiarization Workshops
Report on WS2: Knowledge Level Models of Machine Learning
by Walter Van de Velde
Organizer: Walter Van de Velde (AI Lab, Vrije Universiteit
Brussels)
Program Committee: Agnar Aamodt, Dieter Fensel, Enric Plaza,
Walter Van de Velde and Maarten Van Someren.
1. Workshop Topic
The aim of this workshop was to discuss knowledge level modeling
applied to machine learning systems and algorithms.
An important distinction in current expert systems research is the
one between knowledge level and symbol level [Newell, 1982]. Systems
can be described at either of these levels. Briefly stated, a
knowledge level description emphasizes the knowledge contents of a
system (e.g. goals, actions and knowledge used in a rational way)
whereas the symbol level describes its computational realization (in
terms of representations and inference mechanisms). There is a
consensus that modeling at the knowledge level is a useful
intermediate step in the development of an expert system [Steels and
McDermott, 1993]. So called second generation expert systems
explicitly incorporate aspects of their knowledge level structure,
resulting in potential advantages for knowledge acquisition, design,
implementation, explanation and maintenance (see [David et al., 1993]
for an overview on the state of the art). The technical goal is to
construct generic components which can be reused and refined as
needed, guided by features of the domain and the task instead of by
software engineering considerations.
This workshop investigated the results of describing learning systems
at the knowledge level, hoping to gain some of the same advantages
that were experienced with second generation expert systems. Although
the earliest attempts to do this [Dietterich, 1986] failed to lead to
useful results, later efforts provided interesting insights [Flann
and Dietterich, 1989]. Maybe a more important reason for the
exploration of the knowledge level of learning systems is that the
notion of knowledge level itself, as it is currently used in expert
systems research, is no longer equivalent to Newell's [Van de Velde,
1993]. Currently the knowledge models used are considerably more
manageable, structured and, in a sense, more engineering oriented.
Knowledge level analysis of learning systems can directly benefit
from the developments in knowledge modeling that are currently taking
place (see e.g. [Klinker, 1993] for recent work). Moreover the
knowledge level analysis of machine learning systems can be done
directly in available environments allowing for the easy integration
with problem solving or knowledge acquisition systems.
Note that the relevance of the knowledge level ideas to machine
learning is broader than what is described here (e.g., learning of
knowledge level models). To keep the present workshop relatively
focussed it was suggested to stick closely to the main topic:
knowledge level modeling of machine learning. This topic is relevant
for several reasons. It provides insights into essential features,
differences and similarities of machine learning algorithms. It
contributes to the flexible and problem specific configuration of
learning systems, and their integration with performing systems.
Knowledge level analysis of learning systems also enables the
exchange and reuse of results in machine learning and knowledge
acquisition, which is one of the main bottlenecks in current research
practice. The topic of knowledge level modeling of machine learning
is also well suited for an MLnet familiarization workshop. Europe has
a strong tradition in knowledge level modeling, with the developments
of such methodologies as KADS and Components of Expertise, and of
languages and environments for constructing knowledge level models
(KARL, MoMo, KresT, FML, and so on) and large scale projects such as
MLT and aspects of KADS-II. The workshop attempted to be a bridge
between knowledge acquisition and machine learning, using concepts of
the KA community to understand results in the ML community.
2. Overview of the Presentations
The invited talk ("Is a knowledge-level characterization of robotic
learning always possible?") was presented by Luc Steels (VUB AI-Lab).
In his well known entertaining fashion he challenged the scope of the
task that the workshop had set itself. That knowledge level
descriptions of problem solving are possible, Luc Steels would be the
last to deny. That such descriptions of learning are feasible and
interesting he would not dispute. But that knowledge level modeling
of robotic behaviour is useful or possible, is less clear.
Luc's presentation generated some confusion, as might be expected.
How is his notion of 'motivation' different from goals? Isn't he just
doing control theory? Maybe Franz Schmalhofer and Stuart Aitken
(DFKI, Kaiserslautern, Germany) could have clarified some of these
issues..., if only they had been there. Their paper ("Beyond the
knowledge level: Behaviour descriptions of machine learning systems")
attempts to combine knowledge level modeling with three other types
of models to capture aspects of skill and performance, rather than
just competence.
The presentation by Celine Rouveirol (LRI, Paris) and Patrick Albert
(ILOG) addressed a central issue of the workshop. "Knowledge Level
Modelling of Generate and Test Learning Systems" covers more than
this title suggests. What knowledge is involved in the life-cycle of
a learning task? Tasks for selecting examples, representations,
algorithms and their parameters, evaluating test results and so on.
This project aims at building a configurable learning environment
that assists in these decisions.
The next presentation approached the same problem in a complementary
way. Aurelien Slodzian (VUB AI-Lab, Brussels), presented a concrete
experiment on applying knowledge level modeling to learning. He
described how, within the knowledge engineering workbench KresT, a
wide class of decision tree learning algorithms can be modeled,
configured, operationalized and integrated. The experiment leaves
little doubt about the feasibility. KresT allows one to put into
place models of problem solving, learning and the meta-reasoning that
may be involved in selecting and controlling these. Some of the
details of the actual methods and knowledge for configuration and
selection still have to be filled in, but the project provides all
the hooks to store and operationalize them. Wouldn't it be great to
use the machine learning models from Celine and Patrick within the
LearnKit of KresT?
The paper by James Cupit, Nigel Shadbolt and Sean Wallis from the
University of Nottingham ("The application of a knowledge acquisition
methodology to the analysis of large databases") describes a specific
methodology for analysing large data-sets. What makes this
particularly relevant is that they use a model of analysis to direct
the process of constructing and revising conceptual expertise models.
The model is also used to access various tools that can be
instrumental in the process.
Enric Plaza and Luis Arcos' (IIIA, Blanes, Spain) presentation on the
reification of learning methods in NOOS was cancelled. This was
unfortunate, because NOOS is one of those systems that takes a
knowledge level analysis of case-based reasoning methods as the basis
for a computational architecture. Progress along those lines provides
primary data points to assess the practical value of the enterprise.
Dieter Fensel (AIFB, University of Karlsruhe) asked the question
'What makes it difficult to apply Machine Learning in Model-Based
Knowledge Acquisition?' It seems that machine learning has not
followed the shift in knowledge acquisition from direct extraction of
knowledge in operational form (e.g. rule or frame) to knowledge level
modeling. So, it seems that there is a growing gap between what
machine learning can produce and what knowledge acquisition needs. To
bridge this gap, Dieter argued, several research topics must be
addressed (and occasionally they are). How can knowledge level
description be used to support selection, modification, combination,
and creation of machine learning techniques related to given learning
tasks? How can a complex learning task be decomposed? How can one
type be used to guide the acquisition of others? How can the bias of
a machine learning technique be represented at the knowledge level?
Can formal languages, claimed to be highly successful for models of
problem solving be used for this purpose?
The workshop ended where it had started. Agreeing that knowledge
level descriptions are not all there is, and asking what can they be
used for and what is needed in addition to account for a full theory
of learning behaviour? No doubt this will be an issue for a follow up
workshop that is already in the course of being organized.
Acknowledgements
This workshop was organized with the help of Agnar Aamodt (University
of Trondheim, Norway), Dieter Fensel (University of Karlsruhe,
Germany), Enric Plaza (IIIA, Blanes, Catalunya, Spain), Walter Van de
Velde (VUB AI-Lab, Brussels, Belgium) and Maarten Van Someren (SWI,
University of Amsterdam, The Netherlands). Thanks to MLnet for the
overall organisation and support to the familiarization workshops.
Some of the papers are available on the World-Wide Web at URL
http://arti.vub.ac.be/~walter/ECML/kl-ml.html
References
[David et al., 1993] David, J.-M., Krivine, J.-P., and Simmons, R.
(Eds.). (1993). Second Generation Expert Systems. Springer Verlag,
Berlin.
[Dietterich, 1986] Dietterich, T. G. (1986). Learning at the
knowledge level. Machine Learning, 1, 287-316.
[Flann and Dietterich, 1989] Flann, N. and Dietterich, T. (1989). A
study of explanation-based methods for inductive learning. Machine
Learning, 4(2), 187-226.
[Klinker, 1993] Klinker, G. (Ed.). (1993). Special Issue: Current
issues in knowledge modeling, volume 5 of Knowledge Acquisition.
Academic Press.
***********************************
Technical Reports on the Catania Familiarization Workshops
Report on WS3: Declarative Bias
by Celine Rouveirol
Organizer: Celine Rouveirol (LRI, Orsay, France)
Program Committee: F. Bergadano, F. Esposito, N. Lavrac, I.
Mozetic, C. Nedellec, E. Plaza, L. Popelinsky, D. Sleeman, T. Van de
Merckt, M. Van Someren.
Seven papers were presented at the workshop; these were all quite
diverse. Three papers came from the Inductive Logic Programming
community, which has had for a long time a strong interest in making
the bias explicit in empirical concept learning. H. Ade, L. De Raedt,
M. Bruynooghe's paper "Declarative Bias for bottom up systems",
presented a comparative study of language bias effects in bottom up
ILP generalisation systems. The paper of M. Grobelnik : "Declarative
Bias in the MARKUS system" also presented some language biases for a
top-down MIS like system. D. Mdalenic "Declarative Bias in the ATRIS
rule induction shell" gave an overview and comparison of a range of
search strategies within a hypothesis space. An ILP related paper by
Kukolva and Popelinsky investigates the use of the INDEX system of P.
Flach in extracting "relevant" constraints that are verified in a
relational database. T. Van de Merckt gave an overview of commonly
used language and search biases in some Similarity Based Learning
systems, more precisely in Instance Based, Neural Networks and
Decision Tree learning. J. Herrmann elaborated on biases in his multi
strategy system COSIMA, that was also presented with a different
focus at ECML-94. F. Neri developed his view of bias in an
incremental learning system, where the order in which the examples
are provided to the system can be seen as a major factor that
influences learning.
The presented papers, together with the discussions during the
workshop and at the closing session, demonstrated that there is
currently a substantial effort to find a unifying framework to
compare the different learning biases used in learning systems. Some
open questions that were raised are: Is there some classification of
biases that can be discussed independently from specific learning
systems? What are the biases that should be made delarative in a
learning system (that is, shiftable, automatically or partially with
the help of a user)? How can the effects of a given bias be
characterized with respect to a ML system performance? Some answers
to the latter questions were provided in the workshop , especially in
the comparison papers of Ade et al, Mdalenic, Decastaeker and Van de
Merckt. Some recent results about the complexity of learning under
different language biases produced by the ML and PAC communities have
also contributed to an understanding of these issues.
To sum up, the issue of Declarative Bias seems central to the
successful use of learning. The area of research is quite vast, and
there is still no widely agreed common classification of biases, nor
a clear understanding of how particular biases effect the learning
outcome. But the workshop showed that more and more effort is
currently focusing on finding unifying frameworks to study, represent
and experiment with biases for families of comparable learning
systems.
The working notes of the workshop are available upon email request to
celine@lri.fr.
***********************************
Technical Reports on the Catania Familiarization Workshops
Report on WS4: Machine Learning and Statistics
by Gholamreza Nakhaeizadeh
Organizer(s): Gholamreza Nakhaeizadeh (Daimler-Benz, Ulm) and
Charles Taylor (University of Leeds)
Program Committee: Hans-Hermann Bock, Georg Bol, John Hand, Bob
Henery, Igor Kononenko, J rgen Kreuziger and Rafael Molina
1. Workshop Topic
Statistics and Machine Learning (ML) can help each other's
development. On one hand, the statistical and probabilistic
approaches can be applied by ML-researchers in developing different
ML algorithms. In this connection, one can mention the concepts
developed originally by statisticians and which have been applied by
the ML-community to optimize, prune, and especially to evaluate
different ML-algorithms. Concepts like probabilistic decision trees
and causal networks can be seen in this connection as well. On the
other hand, some ML algorithms, which can perform classification and
forecasting tasks, were developed originally by theML-Community.
They are, however, of interest to statisticians and can demonstrate
new possibilities in adapting classical statistical approaches to
enable the handling of real-world applications. In this connection,
one can mention some of the decision tree and rule based algorithms,
as well as the Case-Based Reasoning approach. Furthermore, perhaps
the ML community can adapt some statistical algorithms to become ML
algorithms. The existing barrier which prevents the statistical
algorithms from becoming ML algorithms is that the partition regions
can not (yet) be described in a way that is meaningful to, and
evaluable by humans. For example, the k-NN method might lend itself
here, if the decision regions could be approximated by a small number
of Voronoi polygons, and the centre of these cells then termed as
prototypes for use in classification.
The above facts show that the two communities, Statistics and ML, can
learn a lot from each other. The main aim of this workshop has been
to bring ML-researchers and statisticians together to discuss the
different aspects of the interface between ML and Statistics. These
***********************************
Focus on Machine Learning Research at Daimler-Benz
by Jutta Stehr
Daimler-Benz is often still seen as a car manufacturing company;
probably not so well known are its various research activities in the
area of computer science and information technology. Most of the IT
research departments are located at the Daimler-Benz Research Centre
in Ulm and in Berlin. One of the Ulm groups is concerned with Machine
Learning and it will be briefly introduced here.
Research in an industrial company tends to be more constrained than
academic research in terms of time and goals. Research projects often
focus on specific real-world problems and the results should be
either easily incorporated into the company's products (in the case
of Daimler-Benz IT research this is software for cars, aeroplanes,
trains, and industrial machinery) or should contribute to various
industrial processes like engineering, manufacturing, or marketing.
Because of this background the research activities within the ML
group cannot be confined to basic research alone. Rather, the group
is concerned with bridging the gap between research and industrial
real-world problems. Therefore a principle thrust of the work is to
develop software solutions by using advanced data analysis and
Machine Learning algorithms to meet the needs of industry without
being directed to mere application development or commercial
products.
The ML group is part of the department for Software Engineering
Research (Manager: Wolfgang Hanika) and its basic roots lie in the
fields of Statistics and Inductive Learning. The leader of the group,
G. Nakhaeizadeh, is Professor of Econometrics at the University of
Karlsruhe. He has been involved in statistical ML research for
several years; this spans work at the University of Karlsruhe and
within Daimler-Benz.
The activities of the group cover a broad range of ML topics.
However, research divides into three major areas: (i) application of
ML algorithms to fields like image processing, text and speech
understanding, and quality insurance, (ii) development of ML systems
in various domains of financial engineering, logistic and production
planning. (iii) evaluation and tuning of ML algorithms.
The ML group currently includes 3 permanent members at the Daimler-
Benz research centre, with 3 PhD students at Ulm and another 2 at
Karlsruhe, and a couple of students involved in master level
research.
There were several internal cooperation with other departments in the
last years. For example, together with the Marketing Department of
Mercedes-Benz the possibilities of the application of Machine
Learning to the prediction of trucks' market development was
investigated. Another project was in the area of quality insurance:
here the test of automatic transmissions was investigated by
automatically generating rules from statistical data. Beside the
internal projects the ML group has been involved in the Esprit
project StatLog. Daimler-Benz has directed the project which had the
overall aim to give an objective assessment of the potential of
different classification algorithms in solving significant commercial
and industrial problems. 23 algorithms including ML- and statistical
algorithms, and Neural Nets were tested on about 22 large-scale
problems. The participants developed a set of objective criteria for
comparing classification algorithms and established an interactive
test environment (the public domain software packages "Evaluation
Assistant" and "Application Assistant").
Current work include:
% development of an intelligent Credit Scoring System that will be
the basis of a risk management concept. Emphasis is laid upon
incremental learning, the use of cost functions, and the comparison
of Machine Learning Technology, Neural Networks, and Statistics (Karl
D bon, Gholamreza Nakhaeizadeh).
% development of a hybrid forecasting and classification method by
combining several Machine Learning Strategies and different types of
inference including Neural Networks, Case based Reasoning and
Statistical Methods (Stefan Ohl, Gholamreza Nakhaeizadeh).
% application of ML algorithms to short-term and medium-term exchange
rate prediction. A system will be implemented which uses symbolic ML
algorithms and Neural Networks as well as ARIMA-Models and
econometric techniques for forecasting non stationary, economic time
series (Elmar Steurer).
% evaluation of Case Based Reasoning techniques to support real-world
planning and design tasks. The work tends to focus on the
practicability of CBR in engineering and manufacturing environments
(Jutta Stehr).
% evaluation of the applicability of ML technology, especially
incremental learning techniques to office environments (Udo Grimmer).
This project includes a proposed cooperation with Washington State
University on the application of ML algorithms to form filling.
Other projects are on genetic algorithms for industrial-size
classification problems, on hybrid forecasting methods for economic
time series including k-nearest neighbour method, Neural Networks and
Regression Trees, and on the comparison of Symbolic and Statistical
ML- Algorithms.
Selected Publications:
% Graf, J. and Nakhaeizadeh, G. (1991): Application of Statistical
and Connectionist Systems for predicting the Development of Financial
Markets. In Heilmann, W.R. et al. (eds.): Geld, Banken und
Versicherungen, VWW Karlsruhe, pp. 1705 - 1720.
% Graf, J. (1992): Stock Market Prediction with neural networks. In
Gritzmann, R. et al. (eds.): Operations Research '91, Tagungsband
1992, pp. 496 - 499.
% Graf, J. (1992): Long-Term Stock Market Forecasting using
Artificial Neural Networks. In Novak, M. (eds.): Neural Network
World, Volume 2, No. 6, pp. 615-620.
% Graf, J. and Nakhaeizadeh, G. (1993): Neural Nets and Symbolic
Machine Learning Algorithms to Prediction of Stock prices. In
Plantamura, V. et al. (eds.): Logistic and Learning for Quality
Software Management and Manufacturing.
% Graf, J. and Nakhaeizadeh, G. (1993): Recent development in Solving
the Credit Scoring Problem. In Plantamura, V. et al. (eds.): Logistic
and Learning for Quality Software Management and Manufacturing.
% Knoll, U. Nakhaeizadeh, G. and Tausend, B. (1994): Cost-sensitive
Pruning Methods for Decision Trees. In Bergadano, F. and De Raedt, L.
(eds.) Proceedings of the Eight European Conference on Machine
Learning (ECML-94). Springer Verlag.
% Merkel A. and Nakhaeizadeh, G. (1992): Application of Artificial
Intelligence Methods to Prediction of Financial time Series. In
Gritzmann, P. et al (eds.). Operations Research 91, pp. 557-559.
% Nakhaeizadeh, G. (1992): Inductive Expert Systems and their
application in Statistics. In Faulbaum F. (ed.) SoftStat 91. Advances
in Statistical Software. Gustav-Fischer, pp. 31-38.
% Nakhaeizadeh, G. (1992): Application of Machine Learning to solving
industrial problems. In Gritzmann, p. et al (eds.). Operations
Research 91, pp. 560-563.
% Nakhaeizadeh, G. (1993): Application of Machine Learning in
Finance. In: Kirn, S. and Weinhardt, C. (eds.). KI-Methoden in der
Finanzwirtschaft, Fachtagung fr KI, Berlin. pp. 137- 142.
% Nakhaeizadeh, G. (1994): Learning Prediction of Time Series. A
Theoretical and Empirical Comparison of CBR with some other
Approaches. In Richter, M. et al (eds.)Proceedings of EWCBR-93,
Universitaet Kaiserslautern.
% Nakhaeizadeh, G. and Reuter A. (1994): Application of Machine
Learning to Predicting Activities in the Automobile Market. To appear
in Langley, P. (ed.): Fielded Applications of Machine Learning,
Morgan Kaufmann.
% Steurer, E. (1993): Nonparametric Exchange Rate Prediciton by using
a modified Nearest Neighbour Method. In Refenes, A. N. et al. (eds.):
Neural Networks in the Capital Markets, Proceedings, London Business
School
% Steurer, E. (to appear): Application of Chaos Theory to Predicting
the Development of Exchange Rates. In: Hipp, Christian et al. (eds.):
Tagungsband zur 6. Tagung Geld, Finanzwirtschaft, Banken und
Versicherungen, Karlsruhe
For further information contact:
Gholamreza Nakhaizadeh
Daimler-Benz AG Department F3W
P.O. Box 2360
89013 Ulm Germany
Tel: +49 731 505-2860
Fax: +49 731 505-4210
email: reza@fuzi.uucp
***********************************
Funding for Workshops/Meetings
MLnet is happy to receive proposals for European based workshops in
the areas covered by the Network re Machine Learning; Knowledge
Acquisition; Case Base Reasoning etc.
To be eligible the meeting must be international in character.
Generally support is only available from MLnet for helping to
organise the meeting. (Organisers can apply separately to the EC for
other funds to support the academic/technical parts of the program.)
For further details please contact:
Derek Sleeman (Academic Coordinator)
Fax: + 44 224 273422
email: (sleeman or mlnet) @csd.abdn.ac.uk
Lorenza Saitta (Convener of Research Technical Committee)
Fax: + 39 11 751 603
email: saitta@di.unito.it
***********************************
Guidelines for Proposals to MLnet
% The proposal should specify clearly the tasks to be done.
% It should clearly specify the outcomes to be delivered with
appropriate time scales
% All resources requested (equipment, staff support, phones etc.)
should be justified.
Also note
% Given the terms of the EC contract, no overheads can be charged.
% Aberdeen are only able to pay cheques in either UK pounds or ECUs
(any other currency will cause us major problems as the EC contract
does not allow the Univ. of Aberdeen to charge bank charges).
***********************************
Putting Together a Machine Discoverer: Basic Building
Blocks*
Jan M. Zytkow
Whichita State University
1845 North Fairmont
Wichita KS 67208-1595
USA
zytkow@wise.cs.twsu.edu
1. Main Direction of Machine Discovery
Machine discoverers can be briefly defined as computer systems that
autonomously pursue knowledge. Research in machine discovery has been
growing in two main directions: (1) automated scientific discovery,
and (2) knowledge discovery in databases. Both directions differ in
the search techniques used and the expected results of discovery.
Database applications are focused on data collected for purposes
different than discovery, and typically sparse as a source of
information about real world phenomena. In contrast, scientific
applications of machine discovery include fine data generation as a
part of the discovery process.
Knowledge discovery in databases has been described in several
collections of papers, edited by Piatetsky-Shapiro (1991, 1993),
Piatetsky-Shapiro & Frawley (1991), Zytkow (1992), Ras (1993), Ziarko
(1993), and many other papers. Automated scientific discovery is
primarily concerned with reconstruction of discovery mechanisms in
sciences such as physics, chemistry and biology. Many of the recent
results can be found in collections edited by Shrager & Langley
(1990), Edwards (1993) and Zytkow (1992, 1993). This research can be
further split into automated discovery of empirical laws and
discovery of hidden structure.
We will focus on automated discovery of empirical laws and on a long
term goal to build automated robotic discoverers, who develop
theories of the real world through empirical investigation.
The other branch, discovery of hidden structure, is not considered in
this paper. It would require an independent substantial treatment.
Contributors include Langley, Simon, Bradshaw & Zytkow (1987); Rose &
Langley (1986); Rose (1989); Karp (1990); Valdes-Perez (1990);
Kocabas (1991); Fischer & Zytkow (1992); Valdes-Perez, Zytkow, &
Simon (1993). The qualitative models approach also belongs here.
Contributors include Sleeman, Stacey, Edwards & Gray (1989); Roverso,
Edwards & Sleeman (1992); Gordon (1992); Metaxas (1993).
2. Cognitive Autonomy of a Discoverer
Throughout history, human discoverers have had to rely on their own
judgement, because the knowledge they proposed was new, often
contradicting the accepted beliefs. A discoverer can be characterized
by autonomous pursuit of new knowledge, accomplished by own choices
in the repertoire of discovery techniques and results. We want to put
the same qualities into machine discoverers.
Difficulties with directly programming knowledge into AI systems led
to a widespread belief that intelligent systems should learn by
imitation of human learning. After many years of research on machine
learning, however, we are still far from an integrated learning
mechanism which would do the job. The areas of greatest strength of
machine learning, such as concept learning from examples and
clustering, are of little use in learning scientific knowledge.
It becomes clear that efficient knowledge acquisition requires an
agent far more active and autonomous than current learning systems.
For instance, a good learner must have a broad understanding of
various forms of knowledge. It must also know how to link new pieces
of knowledge to its previous knowledge. Such links are typically
missing in instruction. Human learners need surprisingly little
instruction. Through the shortfalls of machine learning we come to
appreciate the autonomous absorption of knowledge by a good learner.
Good learners are discoverers; hence understanding the discovery
process is fundamental for understanding learning.
Research in machine discovery, by its focus on cognitive autonomy and
automation of many cognitive steps, is critical for understanding
automated knowledge acquisition. Symptomatically, the motivation for
the workshop on Learning in Autonomous Agents at the MLnet meeting in
Blanes reads like another case for machine discovery (Van de Velde,
1993).
There must be profound reasons why we are discoverers. Certainly we
were discoverers before we became learners. Otherwise we would not
discover the purpose of the learning situations, such as the pointing
gesture by which, as infants, we are taught the meaning of our first
words. Discovery must also dominate learning in animals, as given the
limitations in their language and culture, they must acquire much of
their knowledge by discovery.
To be useful in our research on machine discovery, the notion of
autonomy requires clarification. Suppose that agent A discovers some
piece of knowledge K, which is already known to others, as is often
the case with our machine discoverers. Agent A can be considered a
discoverer of K, if A did not know K earlier and did not learn about
K from external sources. It is relatively easy to trace the external
guidance received by a machine discoverer because all details of the
software are available for inspection. It is true that existing
machine discoverers lack autonomy in many ways. They would not make
discoveries if humans did not provide help by setting system
parameters, selecting search strategies, preparing input data, and
providing them with empirical systems with which to experiment.
These breaches in autonomy do not disqualify machine discoverers,
however, because they are also characteristic of even the greatest
human discoverers, who make relatively small steps beyond their
inherited background of knowledge and method (Zytkow, 1993).
Existing systems are autonomous only to some degree, but future
research in machine discovery will increase their cognitive autonomy.
An agent becomes more autonomous as it is able to make more choices,
satisfy more values and investigate a broader range of goals.
Overcoming the individual limitations of autonomy is a big challenge.
The mere accumulation of new methods, however, does not suffice. The
methods must be strongly integrated, so that more discovery steps can
be performed in succession without external intervention. When
external intervention is replaced by automated search, which must
stay within tractable limits, the accumulation of discovery steps
becomes an even bigger challenge. But it provides motivation and
opportunity for asking the right research questions and gives the
perspective necessary for the answers. A single cognitive step rarely
permits a sound judgement about the results. A combination of steps
provides more informed reasons for acceptance.
3. Anatomy of a Discoverer
The goal of empirical discovery is to develop the theories of
elementary interactions and processes in the world, which can be
combined to create models of physical systems. In modern science,
the path to serious discovery leads through many steps. We will now
analyze the emerging theoretical framework for machine discovery,
reconstructing the discovery method at the level of the main goals.
These goals and plans that carry them out can be called recursively,
until plans are reached which can be directly carried out.
Long experience of machine discovery leads to a vision of automated
discoverers that consist of several basic building blocks: (1)
empirical semantics, (2) experimentation strategies, (3) theory
formation from data, (4) recognition of the unknown, (5)
identification of similar patterns, that may lead to successful
generalizations, (6) theory decomposition to capture elementary
physical interactions.
Induction, which has been often considered the key element of
discovery, is a part of (3), as one of many skills needed in the
process.
Empirical inquiry requires physical systems to experiment with. In
machine discovery little has been done to understand the design of
such systems. Rajamoney (1993) considered situations in which two
competing theories T1 and T2 cannot be distinguished by experiments
on a particular physical set-up S. His system uses S to design other
set-ups that permit crucial experiments to distinguish between T1 and
T2. Other empirical discovery systems, however, take a set-up
experiment S as a given, and only manipulate parameter values within
S.
3.1 Empirical Semantics
Empirical discoverers use manipulators and sensors to undo the actual
experiments. Examples of manipulators are hand, gripper, burrette, or
heater. Examples of sensors are eye, camera, balance, or
thermometer. Manipulators and sensors are applied to the set-up
experiment S, creating states of S desired by the scientist and
recording the actual outcomes.
Software necessary for real-world experiments includes programs which
control sensors and manipulators, so-called device drivers. But
meaningful to the discovery process are not individual operations of
device drivers, but their combinations, prescribed by operational
definitions (Bridgman, 1927; Carnap, 1936; Zytkow, 1982).
Operational definitions are algorithms expressed in terms of
elementary actions of sensors and manipulators, by which states of S
are set or measured. To be scientifically useful, each operational
definition must be adjusted to the details of a particular empirical
set-up (Zytkow, Zhu & Zembowicz, 1992).
Operational definitions, device drivers, concrete devices, and the
experiment set-up, form jointly an empirical interpretation of the
discovery mechanism. Such an interpretation is needed for a concrete
real-world application of an automated discoverer. Discovery by
experiments with a simulation requires a similar, albeit much
simpler, interface (Shen, 1993).
3.2 Empirical Theory Formation
After devices are linked to an empirical system S, and operational
procedures are fine-tuned to fit S, we can abstract from empirical
semantics, and represent S by a multi-dimensional space E, defined as
a Cartesian product of possible values of all parameters that can be
controlled or measured in S. Experiments are the only way for
obtaining information about E, through data which they generate. Data
are generalized into knowledge.
The discovery task is to generate as complete and adequate a theory
of E as possible, including regularities between control variables
and dependent variables, and boundary conditions for regularities.
The theory should be adequate within empirical error. The task can
also include detection of patterns, such as maxima and
discontinuities of dependent variables, and regularities for
parameters of those patterns.
3.3 Experimentation Strategies
An autonomous explorer controls the values of all independent
variables in E and can measure the physical response in terms of
values of dependent variables. Each experiment consists of selecting
a value for each independent variable, and in measuring the values of
all dependent variables.
Experiments are typically organized in sequences, to enable better
calibration, verification, detection of outliers, and error analysis.
The sequences are generated according to different schemas for
different goals, such as induction of empirical equations,
verification, or detection of the scope of applications of a given
theory (Langley et.al. 1987; Kulkarni & Simon, 1987; Koehn & Zytkow,
1986). Shen (1993) considers another experimentation strategy, driven
by the need for pieces of knowledge required in problem solving.
3.4 Theory Formation Mechanism
Several subgoals may be needed on the way to a complete empirical
theory of a multi-dimensional space E. They include discovery of
regularities and other patterns in two variables, recursive
generalization to further dimensions, discovery of regularity
boundaries, data partitioning, identification of similar patterns,
and recognition of areas in which theory is still missing. We will
now discuss these tasks.
Finding the regularities between one control variable and one
dependent variable is an important scientific goal, and a subgoal to
many others. Such regularities, typically empirical equations, have
been the target of discovery systems developed by Gerwin (1974),
Langley et al. (1987), Falkenhainer & Michalski (1986), Nordhausen &
Langley (1990a), Kokar (1986), Wu & Wang (1989), Wong (1991),
Zembowicz & Zytkow (1992), Moulet (1992,1992a), Schaffer (1993),
Dzeroski & Todorovski (1993), Cheng & Simon (1992), and others.
Quantitative discovery systems traditionally focused on regularities
in the form of equations, whereas scientists are often interested in
other patterns, such as maxima, minima, and discontinuities. The
maxima can, for instance, indicate various chemical species, whereby
maximum location indicates the type of ion, while the maximum height
indicates the concentration (Zytkow, Zhu, & Hussam, 1990).
Discontinuities may indicate phase changes.
When an equation Q has been found for a sequence of data, new goals
are to find the limits of Q's application and to generalize Q to
other control variables. When the former goal is successful, that
is, when the boundaries for application of Q have been found, this
leads to the goals of finding regularities beyond the boundaries.
Generalization, in turn, can be done by recursively invoking the
goals of data collection and equation fitting (Langley et al, 1987;
Nordhausen & Langley, 1990, 1993; Koehn & Zytkow, 1986).
If an equation which would fit the data cannot be found, the data can
be decomposed into fragments and the equation finding goal can be
invoked for each fragment separately. Data partitioning can use
maxima, minima, discontinuities, and other special points detected in
the data (Falkenhainer & Michalski, 1986; Rao & Lu, 1992; Zytkow et
al. 1990, 1992). If no regularity can be found, a data set can be
treated as a regularity in the form of a lookup table used for
interpolation.
3.5 Identification of Similar Patterns
When many patterns have been detected, it is important to group them
together by similarity in meaning. Grouping the corresponding
patterns, technically based on their similarity, is a precondition
for successful generalization within each group (Zytkow, Zhu &
Hussam, 1990).
3.6 Recognition of the Unknown
Discoverers explore the unknown. They must be able to examine the
existing state of knowledge to find the boundaries that separate the
known from the unknown. Then they cross the boundaries to explore the
unknown world beyond them. Machine discoverers can use the same
strategy (Scott & Markovitch, 1993; Shen, 1993; Zytkow & Zhu, 1993).
Each discovery goal corresponds to a limitation of knowledge, for
instance, to an area in E which has not been explored, a boundary
which has not been determined, and a generalization which have not
been made. Not every knowledge representation mechanism makes it easy
to determine the unknown. Increasingly, discovery systems use graphs
to represent relationships between the incrementally discovered
pieces of knowledge, and use frame-like structures to represent
knowledge contained in individual nodes in the graphs (Scott &
Markowitch, 1993; Nordhausen & Langley, 1990, 1993). A knowledge
graph can model the topology of laws and their boundaries in the
space E (Zytkow & Zhu, 1991, 1993).
The graph which represents the current state of knowledge can be
examined at any time to find its limitations, which become new goals
for future discovery. We can call this approach knowledge-driven goal
generation. Each knowledge state can be transcended in different
directions, leading to alternative goals, from which one must be
selected according to a selection mechanism. A big advantage of this
approach lies in separating knowledge, goals, and discovery methods
from each other. The mechanisms for goal generation, selection of the
next goal, and selection of the method to approach the goal, can be
independent. Other discoverers, using the same knowledge graph, can
select different goals and apply different methods. This creates a
situation similar to real science, making machine discoverers more
flexible and efficient.
3.7 Discovery of Elementary Interactions
Thus far we have concentrated on finding a network of empirical
equations to describe the space E of many empirical variables over a
fixed physical system S. The next important goal leads from the
equations to their components that describe elementary interactions
in S. Scientists interpret the equations to assign their component
terms physical meaning, for instance the momentum or kinetic energy
of each individual object in S. Equation transformations leading to
such interpretations form a search space explored by Zytkow (1990).
4. Summary
Cognitive autonomy of a discoverer is a matter of degree and it grows
by acquiring more means, goals, and values. We argued that discovery
plays a central role in learning, and that many years of research on
discovery systems have identified a small number of generic goals
needed to discover empirical theories. The system of goals, methods
for goal satisfaction, measuring and manipulating devices, and a
network of the discovered knowledge elements, are the basic building
blocks with which to construct automated discoverers.
Acknowledgement: comments from Peter Edwards helped to clarify many
issues.
References
% Bridgman, P.W. 1927. The Logic of Modern Physics.
% Carnap, R. 1936. Testability and Meaning, Philosophy of Science,
Vol.3.
% Cheng, P.C. & Simon, H.A. 1992. The Right Representation for
Discovery: Finding the Conservation of Momentum. in: Sleeman &
Edwards eds. Proc. of Ninth Intern. Conference on Machine Learning,
62-71.
% Dzeroski, S. & Todorovski, L. 1993. Discovering Dynamics, Proc. of
10th International Conference on Machine Learning, 97-103
% Edwards, P. ed. 1993. Working Notes MLnet Workshop on Machine
Discovery. Blanes, Spain, Sep.23.
% Falkenhainer, B.C. & Michalski, R.S. 1986. Integrating quantitative
and qualitative discovery: The ABACUS system. Machine Learning,
Vol.1, 367-401.
% Fischer, P., & Zytkow, J.M. 1992. Incremental Generation and
Exploration of Hidden Structure, in: Zytkow J. ed. Proc. of ML-92
Workshop on Machine Discovery, Aberdeen, UK, July 4, 103-110.
% Gerwin, D.G. 1974. Information processing, data inferences, and
scientific generalization, Behav.Sci. 19, 314-325.
% Gordon, A. 1992. Informal Qualitative Models in Scientific
Discovery. in: Zytkow J. ed. Proc. of ML-92 Workshop on Machine
Discovery, Aberdeen, UK, July 4, 98-102.
% Karp, P. 1990. Hypothesis Formation as Design. in: J.Shrager & P.
Langley eds. Computational Models of Scientific Discovery and Theory
Formation, Morgan Kaufmann Publishers, San Mateo, CA, 275-317.
% Kocabas, S. 1991. Conflict Resolution as Discovery in Particle
Physics. Machine Learning 6, 277-309.
% Koehn, B. & Zytkow, J.M. 1986. Experimenting and Theorizing in
Theory Formation. in: Ras Z. & Zemankova M. eds. Proc. of the
International Symposium on Methodologies for Intelligent Systems.
ACM SIGART Press, 296-307.
% Kokar, M.M. 1986. Determining Arguments of Invariant Functional
Descriptions, Machine Learning, 1, 403-422.
% Kulkarni, D., & Simon, H.A. 1987. The Processes of Scientific
Discovery: The Strategy of Experimentation, Cognitive Science, 12,
139-175.
% Langley, P., Simon, H.A., Bradshaw, G.L. & Zytkow, J.M. 1987.
Scientific Discovery: Computational Explorations of the Creative
Processes. Cambridge, MA: The MIT Press.
% Metaxas, S. 1993. The Prediction of Physical Properties with
CRITON. In: Edwards P. ed. Working Notes, MLnet Workshop on Machine
Discovery, Blanes, Spain Sep.23, 61-65.
% Moulet, M. 1992. A symbolic algorithm for computing coefficients'
accuracy in regression, in: Sleeman D. & Edwards P. eds. Proc. of
Ninth Intern. Conference on Machine Learning.
% Moulet, M. 1992a. ARC.2: Linear Regression In ABACUS, in: Zytkow J.
ed. Proc. of ML-92 Workshop on Machine Discovery, Aberdeen, UK, July
4, 137-146.
% Nordhausen, B., & Langley, P. 1990. An Integrated Approach to
Empirical Discovery. in: J.Shrager & P. Langley eds. Computational
Models of Scientific Discovery and Theory Formation. Morgan Kaufmann
Publishers, San Mateo, CA. 97-128.
% Nordhausen, B., & Langley, P. 1990a. A Robust Approach to Numeric
Discovery, Proc. of Seventh International Conference on Machine
Learning, Palo Alto, CA: Morgan Kaufmann. 411-418.
% Nordhausen, B. & Langley, P. 1993. An Integrated Framework for
Empirical Discovery, Machine Learning, 12, 17-47.
% Piatetsky-Shapiro, G. ed. 1991. Proc. of AAAI-93 Workshop on
Knowledge Discovery in Databases.
% Piatetsky-Shapiro, G. ed. 1993. Proc. of AAAI-93 Workshop on
Knowledge Discovery in Databases.
% Piatetsky-Shapiro, G. & Frawley, W. eds. 1991. Knowledge Discovery
in Database}, The AAAI Press, Menlo Park, CA.
% Rajamoney, S.A. 1993. The Design of Discrimination Experiments.
Machine Learning, 12, 185-203.
% Rao, R.B. & Lu S.C. 1992. Learning Engineering Models with the
Minimum Description Length Principle, Proc. of Tenth National
Conference on Artificial Intelligence, 717-722.
% Ras, Z. ed. 1993. Journal for Intelligent Information Systems,
Vol.2.
% Rose, D. 1989. Using Domain Knowledge to Aid Scientific Theory
Revision. Proc. of Sixth Intern. Workshop on Machine Learning, Morgan
Kaufmann Publishers, San Mateo, CA.
% Rose, D. & Langley, P. 1986. Chemical Discovery as Belief Revision,
Machine Learning, 1, 423-451.
% Roverso, D., Edwards, P. & Sleeman, D. 1992. Machine Discovery by
Model Driven Analogy. in: Zytkow J. ed. Proc. of ML-92 Workshop on
Machine Discovery, Aberdeen, UK, July 4, 87-97.
***********************************
Mobal 3.0 Released
The ML group at GMD have now released Mobal 3.0, an enhanced version
of their knowledge acquisition and machine learning system for first-
order KBS development on Sparc workstations. Mobal is a multistrategy
learning system that integrates a manual knowledge acquisition and
inspection environment, a powerful first-order inference engine, and
various machine learning methods for automated knowledge acquisition,
structuring, and theory revision.
As the most visible change, the new release 3.0 no longer requires
Open Windows, but features an X11 graphical user interface built
using Tcl/Tk. This should make installation trouble-free for most
users, and through its networked client-server structure, allows easy
integration with other programs.
As a second change resulting from work in the ILP ESPRIT Basic
Research project, Mobal 3.0 now offers an "external tool" facility
that allows other (ILP) learning algorithms to be interfaced to the
system and used from within the same knowledge acquisition
environment. The current release of Mobal includes interfaces to
GOLEM by S. Muggleton and C. Feng (Oxford University), GRDT by V.
Klingspor (Univ. Dortmund) and FOIL 6.1 by R. Quinlan and M. Cameron-
Jones (Sydney Univ.).
GMD grants a cost-free license to use Mobal for academic purposes.
The system can be obtained from ftp.gmd.de, directory /ml-
archive/GMD/software/Mobal (login anonymous, password your E-Mail
address). For details about the scientific background of Mobal, see
the book "Knowledge Acquisition and Machine Learning", by K. Morik,
S. Wrobel, J.-U. Kietz and W. Emde (Academic Press, 1993). A user
guide is available via FTP.
***********************************
News from the University of Dortumnd
Katharina Morik's move from GMD to the University of Dortmund as a
full professor in AI is now complete. Following her appointment to
Dortmund in 1991, she had continued as external representative of
GMD's MLT project, with Stefan Wrobel taking care of internal
business. With the end of the MLT project in mid-1993, the ML group
at GMD is now lead by Stefan Wrobel. The group consist of four
scientists (Werner Emde, Jrg-Uwe Kietz, Edgar Sommer and Stefan
Wrobel) and two students (Roman Englert and Marcus L bbe), and will
further develop Mobal and other ML systems. As in the past, the
groups at University of Dortmund and at GMD will continue to
collaborate closely.
***********************************
1996 INTERNATIONAL MACHINE LEARNING CONFERENCE
CALL FOR PROPOSALS
In 1993 the ML Journal Board agreed that a cycle of meetings would be
set up so that this conference will be held on the Eastern Coast of
America, West Coast of America and Europe on a three yearly cycle.
In June (or July) 1996, the Thirteenth International Machine Learning
Conference will be held at a European site. The purpose of this call
is to invite groups interested in organizing and hosting the
conference to submit proposals. The group selected to run the
conference will be given full authority and responsibility for
organising the conference.
Proposals should address the following issues:
1. Organization and Format.
In previous years, the format of the conference has alternated
annually between a single plenary session (1988, 1990) and a set of
parallel workshop sessions (1989, 1991). However, since 1992, the
format has involved 3 days of plenary sessions preceded or followed
by one day of specialized workshops. In 1992 a poster session was
held, while in 1993 and 1994, parallel sessions were held in addition
to the plenary session. Please indicate which format you propose and
how you would arrange the schedule to suit the format. You may
present more than one possible format, if you wish.
In the past, the conference has been organized by an Organizing
Committee whose membership included organizers of past conferences
and other senior researchers. Review of papers has been conducted by
a Program Committee selected by the organizers with the advice of the
Organizing Committee. Most recently, all reviewing has been
conducted via email. Please indicate what organization you would
employ.
2. Local Parameters.
% Accessibility. Is it easy and inexpensive for people (especially
graduate students) to travel to the conference site? (Compute mean
airfares from Europe and North America.)
% Meeting Rooms, AV Equipment, etc. What are the physical facilities
like?
% Meals and Lodging. Is there low-cost, medium quality housing
available for attendees (especially graduate students)? How far is
this housing from the meeting rooms? How will attendees get between
the two sites? Where will attendees eat?
% Demo facilities. Will there be computing equipment and space
available to support demos?
3. Local Machine Learning Community.
Is there a local ML group/community that can help with organization
and funding?
4. Organizational and Financial Support.
Can the host institution(s) provide support for registration and
financial management (e.g., credit card payments, accounting, etc.).
How will the conference be funded? Provide a draft budget covering
expenses, expected registration fee schedule, and sources of
financial support (this is very important). The host institution
must agree to forward any unused funds to the host of the 1997
conference. In previous years, funding has been obtained from
federal granting agencies, corporations, and universities.
Proposals should be sent before July 5, 1994 to
Tom Diettrich
Department of Computer Science
303 Dearborn Hall
Oregon State University Corvallis,
OR 97331-3202
tgd@cs.orst.edu
fax: 503-737-3014
Email is preferred.
The choice of organizers will be made at the July meeting of the
Editorial Board of the Machine Learning Journal, which is tentatively
scheduled for the evening of July 11 during the 1994 International
Machine Learning Conference at Rutgers University.
***********************************
Technical Meetings Between MLnet and Other
Networks
A joint meeting with ELSNET* is planned (details will be announced
shortly).
Proposals are welcomed from Network members for joint meetings
between MLnet and other NoEs. Generally these meetings will receive
some financial support from MLnet.
Please send proposals to:
Derek Sleeman
Fax: + 44 224 273422
email: sleeman@csd.abdn.ac.uk
Alternatively, he would also be happy to discuss outline ideas for
such meetings.
You are encouraged to start discussing details well before the
planned dates.
***********************************
News from our "data base" of industrial
applications
We gathered information from almost 60 European companies involved in
Machine Learning or Knowledge Acquisition. This data has been sent
back to the concerned companies in order to get their approval for
publication, (especially for those that provided us with information
that has been tailored to our format).
We have since started investigating as well applications in domains
near to ours: applications of Neural Networks, Genetic Algorithms,
and Case-Based Reasoning. If readers know of any company involved in
these topics, please let us know.
Information to:
Dr Yves Kodratoff
CNRS & Universite Paris-Sud
LRI, Batiment 490
91405 Orsay
France
email: yk@lri.lri.fr
Fax: +33 1 6941 6586
***********************************
Procedures for joining MLnet
Initial enquiries will receive a standard information pack (including
a copy of
the Technical Annex)
All centres interested in joining MLnet are asked to send the
following to MLnet's
Academic Coordinator:
% A signed statement on Institutional notepaper saying that
you have
read and agreed with the general aims of MLnet given in the Technical
Annex;
% One hard-copy document listing the Machine Learning (and
related
activities) at the proposed node and three copies of any enclosures;
the document
should include a list of scientists involved in these field(s), half
page
curriculum vitae for each of these senior scientists, current
research students,
lists of recent grants and relevant publications over the last 5 year
period;
% A statement of the Technical Committees which the Centre
would be
interested in joining, and a succinct statement of the potential
contributions of
the Centre to the Network and its Technical Committees.
Two members of the Management Board will be asked to look at the
material in
detail and will present the proposal at the next Management Board
meeting.
Through the Network's Coordinator, the members may ask for additional
information.
The Academic Coordinator will be in touch with the Centre as soon as
possible
after the Management Board meeting.
The Management Board is not planning to set a fixed timetable for
applications,
but advises potential nodes that it currently holds Management Board
meetings in
November, April and September, and that papers would have to be
received at least
six weeks before a Management Board meeting to be considered.
(Contact Derek
Sleeman for details).
***********************************
Main Nodes
Professor D Sleeman, Aberdeen University (GB)
Tel No: +44 224 27 2288/2304
Fax No: +44 224 27 3422
Professor B J Wielinga, University of Amsterdam (NL)
Tel No: +31 20 525 6789/6796
Fax No: +31 20 525 6896
Professor R Lopez de Mantaras, IIIA AIRI, Blanes (ES)
Tel No: +34 72 336 101
Fax No: +34 72 337 806
Professor K Morik, Dortmund University (DE)
Tel No: +49 231 755 5101
Fax No: +49 231 755 5105/2047
Dr L DeRaedt, Leuven Katholieke Universiteit (BE)
Tel No: +32 16 20 10 15
Fax No: +32 16 20 53 08
Dr Y Kodratoff, Paris Sud University, Orsay (FR)
Tel No: +33 1 69 41 69 04
Fax No: +33 1 69 41 65 86
Professor L Saitta, Torino University (IT)
Tel No: +39 11 742 9214/5
Fax no: +39 11 751 603
Dr Gholamreza Nakhaeizadeh, Daimler-Benz, Ulm (DE)
Tel No: +49 731 505 2860
Fax No: +49 731 505 4210
Mr T Parsons, British Aerospace plc, Bristol (GB)
Tel No: +44 272 363 458
Fax no: +44 272 363 733
Mr F Malabocchia, CSELT S.p.A., Torino, (IT)
Tel No: +39 11 228 6778
Fax No: +39 11 228 5520
Dr D Cornwell, CEC Project Officer
Tel No: +32 2 296 8664/8071
Fax No: +32 2 296 8390/8397
Associate Nodes
% Alcatel Alsthom Recherche, Marcoussis (FR) % ARIAI, Vienna (AT) %
Bari University (IT) % Bradford University (GB) % Catania University
(IT) % Coimbra University (PT) % CRIM-ERA, Montpellier (FR) %
Electricite de France, Clamart , Paris (FR) % FORTH, Crete (GR) %
Frankfurt University (DE) % GMD, Bonn (DE) % Kaiserslautern
University (DE) % Karlsruhe University (DE) % Ljubljana AI Labs (SL)
%JNottingham University (GB) % Oporto University (PT) % Oxford
University (UK) % Paris VI University (FR) % Pavia University (IT) %
Prague University (CZ) % Reading University (GB) % Savoie University,
Chambery (FR) % Stockholm University (SE) % Tilburg University (NL) %
Trinity College, Dublin (IE) % Ugo Bordoni Foundation, Roma (IT) %
VUB, Brussels (BE) % ISoft, Gif sur Yvette (FR) %JMatra Marconi
Space, Toulouse (FR) % Siemens AG, Munich (DE)
Academic Coordinator:
Derek Sleeman
Department of Computing Science
University of Aberdeen
King's College
Aberdeen AB9 2UE
Scotland, UK
Tel: +44 224 27 2288/2304
Fax: +44 224 27 3422
email: {mlnet, sleeman}@csd.abdn.ac.uk
Documents available from Aberdeen:
State of the Art Overview of ML and KA
Recently Announced projects (ESPRIT III)
MLnet Flyer
First Year Report
Policy Statement
**********************************
MLnet NEWS 2.3 END
**********************************
------------------------------
End of ML-LIST (Digest format)
****************************************
