Copy Link
Add to Bookmark
Report
Machine Learning List Vol. 3 No. 13
Machine Learning List: Vol. 3 No. 13
Tuesday, August 6, 1991
Contents:
Machine Learning: Issues, Answers, and Quandaries
TR on Bayesian Mixture Modeling
Call for participation for AAAI 1992 Spring Symposium Series
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 /usr2/spool/ftp/pub/ml-list/V<X>/<N> or N.Z
where X and N are the volume and number of the issue; ID & password: anonymous
------------------------------
Date: Mon, 29 Jul 91 15:26:54 PDT
From: Tom Dietterich <tgd@turing.cs.orst.EDU>
Subject: Machine Learning: Issues, Answers, and Quandaries
Reply-To: tgd@turing.cs.orst.EDU
Several people requested copies of the slides from my AAAI invited
talk entitled "Machine Learning: Issues, Answers, and Quandaries".
Hence, I've decided to make them available as a postscript file. The
file is located on host "cs.orst.edu" (128.193.32.1). You can connect
via anonymous ftp:
local> ftp cs.orst.edu
Connected to cs.orst.edu.
220 lynx FTP server (Version $Revision: 15.15 $ $Date: 89/08/31 10:33:40 $) read
y.
Name (cs.orst.edu:tgd): anonymous
331 Guest login ok, send ident as password.
Password: foo@a.b.c.edu
230 Guest login ok, access restrictions apply.
ftp> cd pub/tgd
200 CWD command okay.
ftp> binary
200 Type set to I.
get aaai-slides.ps.Z
200 PORT command okay.
150 Opening data connection for aaai-slides.ps.Z (128.193.36.1,2631) (107809 byt
es).
226 Transfer complete.
local: aaai-slides.ps.Z remote: aaai-slides.ps.Z
107809 bytes received in 1.1 seconds (96 Kbytes/s)
ftp> bye
local> uncompress aaai-slides.ps
local> lpr -P(your postscript printer) aaai-slides.ps
This postscript file requires a fairly large postscript printer. I
have printed it successfully on a personal laser writer. It does not
print successfully on an HP laserjet III si because of some kind of
incompatibility. It should print 55 pages.
It is not clear whether AAAI will be producing a written form of the
talks.
--Tom
------------------------------
From: Radford Neal <radford@ai.toronto.EDU>
Subject: TR on Bayesian Mixture Modeling
Date: Mon, 29 Jul 1991 16:47:04 -0400
[The following message is the first example of a change in editorial
policy. Previously, I had hesitated to send out announcements of the
availability of single technical reports, fearing that the list would be
inundated with them and I would spend all my time fowarding such requests.
I still firmly believe that the readers of the list would be better
served by announcing ALL machine learning TRs from various sites
together with ordering information. However, this information does
not appears to be easily available to people who post here.
If I get too many requests, I may create special abstract issues
or look for a different moderator. I suggest keeping abstracts short
(i.e. to a about 15 lines of text so the whole mesage fits on
one screen of a terminal)-
MP]
The following technical report is available for ftp from the neuroprose
archive. A hardcopy may also be requested. (See below for details.)
Though written for a statistics audience, this report should be of
interest to those interested in machine learning, as it reports a
Bayesian solution for one type of "unsupervised concept learning".
The problem is similar to that addressed by Cheeseman, et al's
AutoClass system, while the technique employed is related to that used
in Boltzmann Machines.
Bayesian Mixture Modeling by Monte Carlo Simulation
Radford M. Neal
Technical Report CRG-TR-91-2
Department of Computer Science
University of Toronto
It is shown that Bayesian inference from data modeled by a mixture
distribution can feasibly be performed via Monte Carlo simulation.
This method exhibits the true Bayesian predictive distribution,
implicitly integrating over the entire underlying parameter space.
An infinite number of mixture components can be accommodated without
difficulty, using a prior distribution for mixing proportions that
selects a reasonable subset of components to explain any finite
training set. The need to decide on a ``correct'' number of components
is thereby avoided. The feasibility of the method is shown empirically
for a simple classification task.
To obtain a compressed PostScript version of this report from neuroprose,
do the following:
unix> ftp cheops.cis.ohio-state.edu
(if this doesn't work, try the numerical address 128.146.8.62)
(log in as "anonymous", with password "neuron")
ftp> binary
ftp> cd pub/neuroprose
ftp> get neal.bayes.ps.Z
ftp> quit
unix> uncompress neal.bayes.ps.Z
(the PostScript version is now present in neal.bayes.ps)
To obtain a hardcopy version of the paper by physical mail, send mail
to:
Maureen Smith
Department of Computer Science
University of Toronto
6 King's College Road
Toronto, Ontario
M5A 1A4
or use the e-mail address:
maureen@ai.toronto.edu
------------------------------
Subject: Call for participation for AAAI 1992 Spring Symposium Series
Date: Tue, 06 Aug 91 12:10:10 -0700
[From comp.ai. I have edited this message to highlight topics
of interest to ML-LIST readers. The full Call for participation
can be obtained from sss@aaai.org- MP]
Call for Participation
AAAI Spring Symposium Series
March 25,26, & 27, 1992
Stanford University
Sponsored by
The American Association for Artificial Intelligence
445 Burgess Drive, Menlo Park, CA 94025
(415) 328-3123
sss@aaai.org
Introduction
The AAAI (in cooperation with Stanford University's Department of Computer
Science) presents the 1992 Spring Symposium Series to be held Wednesday
through Friday, March 25--27, 1992, at Stanford University.
The topics of the nine symposia in the 1992 Spring Symposium Series are:
Artificial Intelligence in Medicine
Cognitive Aspects of Knowledge Acquisition
Computational Considerations in Supporting Incremental Modification and Reuse
Knowledge Assimilation
Practical Approaches to Scheduling and Planning
Producing Cooperative Explanation
Propositional Knowledge Representation
Reasoning with Diagrammatic Representations
Selective Perception
Most symposia will be limited to approximately 60 participants. Each
participant will be expected to attend a single symposium. Working notes
will be prepared and distributed to participants in each symposium.
A general plenary session will be scheduled in which the highlights of each
symposium will be presented and an informal reception will be held on
Wednesday evening, March 25.
In addition to invited participants, a limited number of other interested
parties will be allowed to register in each symposium. Registration
information will be available in December 1991. To obtain registration
information write to the address listed on the front of this brochure.
Submission Requirements
Submission requirements vary with each symposium, and are listed in the
descriptions of the symposia. Please send your submissions directly to the
address given in the description. DO NOT SEND submissions to AAAI. All
submissions must arrive by November 15, 1991. Acceptances will be mailed by
December 13, 1991. Material for inclusion in the working notes of the
symposia will be required by January 31, 1992.
ARTIFICIAL INTELLIGENCE IN MEDICINE
Send all materials by November 15, 1991, to:
Michael G. Kahn
Medical Informatics Laboratory
Department of Internal Medicine
Washington University School of Medicine
Campus Box 8005
660 South Euclid Avenue
St. Louis MO 63110
EMAIL: kahn@informatics.WUSTL.EDU
FAX: 314-362-8015
Phone: 314-362-4320
Program Committee: Michael Kahn (co-chair), Jack Smith (co-chair),
Bruce Buchanan, Mark Musen, Peter Szolovits.
COGNITIVE ASPECTS OF KNOWLEDGE ACQUISITION
The objective of this Symposium is to bring together a multi-disciplinary
group of researchers to focus on issues associated with cognitive aspects
of knowledge acquisition. It will have as its theme the knowledge
processes of society, what these are, how the knowledge construct arises in
model, what purpose it serves as a theoretical and practical construct, how
individual cognitive processes mesh with socio-cultural processes, and the
roles of information technology and artificial intelligence in the
development of these processes.
The issues center around the relation between the individual as a skilled
autonomous agent and the knowledge processes within the cultures and
societies within which that individual is embedded. What artefacts arise
when we model knowledge as located within the individual and neglect the
dependence of that individuals knowledge processes on supporting cultures?
In particular, the individuals knowledge acquisition processes seem
critically dependent on socio-cultural phenomena which have so far not been
overtly taken into account in the study of knowledge acquisition in
artificial intelligence.
It would be unrealistic to expect the Symposium to provide answers to these
questions. Rather it is intended to nucleate an ongoing discussion between
various communities of interest that have related ideas, problems, and
complementary contributions to a very significant and wide-ranging research
domain. The sessions will be structured around provocative position papers
raising and clarifying major issues, with a focus on discussing the
significance of these issues for different sub-disciplines of artificial
intelligence, and the contributions to these issues both from within
artificial intelligence and from many other disciplines including
philosophy, psychology, anthropology, cognitive science, sociology and
economics.
Those wishing to participate should submit five copies of a position paper
of five to twenty pages. Theoretical papers addressing the basic issues and
practical papers on relevant experience and the impact on knowledge
acquisition methodologies are equally welcome. Papers should be concise
and crisp in style, raising issues and stating them clearly for general
discussion.
Prospective participants are encouraged to contact committee members for more
information on the symposium.
Papers must be sent so that they arrive by 15 November, 1991, to:
Brian Gaines
AAAI Spring Symposium
Knowledge Science Institute
University of Calgary
Calgary, Alberta T2N 1N4
CANADA
Program Committee: John H. Boose, Boeing Advanced Technology Center
(john@atc.boeing.com); Bill Clancey, Institute for Research on Learning
(William_Clancey.PARC@xerox.com); Brian Gaines, University of Calgary
(gaines@cpsc.ucalgary.ca); Alain Rappaport, Neuron Data
(Alain.Rappaport@ML.RI.CMU.EDU)
COMPUTATIONAL CONSIDERATIONS IN SUPPORTING INCREMENTAL MODIFICATION AND REUSE
Submissions should be sent to arrive by November 15th to:
Subbarao Kambhampati
Department of Computer Science and Engineering
Arizona State University,
Tempe, AZ 85287
U. S. A.
Program Committee: Ashok Goel (goel@cc.gatech.edu), Subbarao
Kambhampati (Chair, rao@cs.stanford.edu), John Mylopolous
(jm@ai.toronto.edu), Bill Swartout (swartout@isi.edu).
KNOWLEDGE ASSIMILATION
In recent years, much of machine learning research has concentrated
on algorithms for the two relatively separate tasks of accelerating
problem solvers, and inducing concepts from examples. Important new
techniques have emerged in both areas, notably explanation-based
learning and probably-approximately-correct algorithms. However,
relatively little attention has been paid to learning techniques
that can enable an agent to improve its performance along multiple
dimensions over time. The symposium will focus on this task as a
potential new unifying theme for research.
A system whose performance improves in an independently changing
environment must be capable both of acquiring fresh information and
of increasing the effectiveness with which it uses the information it
already possesses. The title ``Knowledge Assimilation'' emphasizes
the need to mesh together old and new knowledge, in sharp distinction
to both pure speedup learning and pure concept induction. Overall
performance improvement is also more than speedup learning plus
concept induction, because new and old knowledge must be restructured
and interlinked to allow all information to be used effectively.
Moreover, knowledge assimilation cannot have a single objective:
ecologically useful learning algorithms must make rational tradeoffs
between increasing an agent's stock of information, its speed, and
other aspects of performance.
Some specific research questions relevant to the theme of knowledge
assimilation include:
- What knowledge must an agent possess before it can start learning?
- How is knowledge assimilation different for robots
and for disembodied knowledge base?
- How does the availability of a teacher or the ability to
perform experiments influence learning?
- How should an agent allocate resources to different learning subtasks?
- What are the implications of PAC results for practical learning agents?
- When is it useful to introduce new ontological distinctions?
- How can theories be restructured to accommodate new concepts?
- How can the usefulness of alternative knowledge (re)organizations
be measured?
- How should analytical and empirical support for
elements of knowledge be combined?
- How should knowledge be updated to account for contradictory evidence?
Although it is hoped that participants will reflect on the the theme
of knowledge assimilation, and make efforts to discuss their work from
a shared perspective, there will be room at the symposium for
conflicting points of view. The fundamental objective is to bring
together researchers in machine learning and related areas to discuss
the opportunities for consolidating past work and moving forward in a
common direction.
Concretely, the symposium will consist of individual presentations and
panel debates, with ample opportunity for all participants to relate
experiences and express opinions. Those who wish to make
presentations should submit a draft paper, of length at most ten
pages. In addition to reports on current research, which may be
preliminary, critical reviews and rational reconstructions of previous
work are also welcome. All prospective submitters are encouraged to
contact the program committee (preferably by email) to discuss how
what they wish to present is coordinated with the objectives of the
symposium. Paper submissions are encouraged from students as well as
from experienced researchers.
Those who wish to attend without presenting a paper should send a
description of their research interests and a list of their related
publications.
Four copies of all submissions should be sent to arrive by November 15,
1991, to
Charles Elkan
Department of Computer Science and Engineering
University of California, San Diego
La Jolla, California 92093-0114
The program committee consists of Tom Dietterich (tgd@cs.orst.edu),
Charles Elkan (elkan@cs.ucsd.edu), Oren Etzioni (etzioni@cs.washington.edu),
and Bart Selman (selman@research.att.com).
PRACTICAL APPROACHES TO SCHEDULING AND PLANNING
Government and industry require practical approaches to a diverse set
of complex scheduling and planning problems. While scheduling has
been studied in isolation for many years, recent advances in
artificial intelligence, control theory, and operations research
indicate a renewed interest in this area. In addition, the scheduling
problem is being defined more generally, and work is beginning to
consider the closed-loop use of scheduling systems in operational
contexts. This symposium will serve to bring together theorists and
practitioners from diverse backgrounds, with the aim of disseminating
recent results and fostering the development of a cross-discipline
understanding.
The symposium will focus on issues involved in the construction and
deployment of practical scheduling systems that can deal with resource
and time limitations. To qualify as ``practical'', a system must be
implemented and tested to some degree on non-trivial problems
(ideally, on real-world problems). However, a system need not be
fully deployed to qualify. Systems that schedule actions in terms of
metric time constraints typically represent and reason about an
external numeric clock or calendar, and can be contrasted with those
systems that represent time purely symbolically.
Issues to be discussed at the symposium include, but are not strictly
limited to, the following.
- Integrating planning and scheduling.
- Integrating symbolic goals and numerical utilities.
- Managing uncertainty.
- Incremental rescheduling.
- Managing limited computation time.
- Anytime scheduling and planning algorithms, systems.
- Dependency analysis and schedule reuse.
- Management of schedule and plan execution.
- Incorporation of techniques from discrete event control.
- Incorporation of techniques from operations research.
- Learning.
- Measures of schedule and plan quality.
- Search techniques.
- Methodology.
- Applications.
Prospective participants are encouraged to submit papers that deal with any
of these issues. Research results, position declarations, and system
descriptions are all appropriate. A paper need not describe final results,
so reports on work in progress are welcome. Papers must be no longer than
10 pages and should be as short as possible. All papers are to be sent via
electronic mail to zweben@ptolemy.arc.nasa.gov (standard LaTeX or pure
ASCII only), to arrive by November 15, 1991. Prospective participants
may contact committee members for more information on the symposium.
If electronic mail is impossible, send five paper copies (same deadline)
clearly marked ``AAAI Spring Symposium'' to:
Monte Zweben
NASA Ames Research Center
MS: 244-17
Moffett Field, CA 94035
U.S.A.
Program Committee: Mark Drummond, NASA Ames Research Center
(drummond@ptolemy.arc.nasa.gov); Mark Fox, Carnegie-Mellon University
(mark.fox@golem.cimds.ri.cmu.edu); Austin Tate, AI Applications
Institute (A.Tate%ed@nsfnet-relay.ac.uk); Monte Zweben, NASA Ames
Research Center (zweben@ptolemy.arc.nasa.gov)
PRODUCING COOPERATIVE EXPLANATIONS
Electronic submissions are preferred, and should be sent to
alex@wiliki.eng.hawaii.edu. If electronic mail is unavailable, four paper
copies should be submitted to:
Dr. Alex Quilici
Department of Electrical Engineering
University of Hawaii at Manoa
2540 Dole St, Holmes Hall 483
Honolulu, HI, 96822
All submissions must be received by November 15, 1991.
Program Committee: David Chin, Johanna Moore, Cecile Paris, Alex
Quilici (chair).
PROPOSITIONAL KNOWLEDGE REPRESENTATION
Submissions should be sent, to arrive by November 15, to:
Deepak Kumar
Department of Computer Science
226 Bell Hall
State University of New York at Buffalo
Buffalo, NY 14260
kumard@cs.buffalo.edu
Phone: (716) 636 2193
Fax : (716) 636 3464
Program Committee: Stuart C. Shapiro (Chair), State University of New York
at Buffalo (shapiro@cs.buffalo.edu); John Barnden, New Mexico State
University (jbarnden@nmsu.edu); Joao P. Martins, University of Lisbon,
Portugal (ist_1416@ptifm.bitnet); John F. Sowa, IBM (sowa@watson.ibm.com)
REASONING WITH DIAGRAMMATIC REPRESENTATIONS
Hardcopy submissions (4 copies) are strongly encouraged though we will also
accept electronic submissions. Submissions (please include e-mail address
if available) may be sent to arrive by November 15, 1991 to:
Hari Narayanan
Laboratory for AI Research
Department of Computer & Information Science
The Ohio State University
Columbus, OH 43210, USA
e-mail: narayan@cis.ohio-state.edu
tel: (614) 292-1413
Enquiries and requests for more information may also be sent to this address.
Program Committee: B. Chandrasekaran (Co-chair), Yumi Iwasaki, Hari Narayanan,
Herbert Simon (Co-Chair).
SELECTIVE PERCEPTION
Those interested in attending should submit a 3-5 page extended abstract
describing their current work in this field and/or prospects for future
research. The submission should make clear how the use of selective
perception facilitates the system's tasks.
Four (4) copies of each submission should be sent to arrive by November 15,
1991 to:
Reid Simmons
School of Computer Science
Carnegie Mellon University
5000 Forbes Avenue
Pittsburgh, PA 15213
Submissions may also be sent by electronic mail to reid.simmons@cs.cmu.edu.
This address may also be used to obtain more information about the goals of
the symposium and for other inquiries.
Program Committee: Dana Ballard, Tom Dean, James Firby, Reid Simmons (chair)
------------------------------
END of ML-LIST 3.13