Copy Link
Add to Bookmark
Report
AIList Digest Volume 8 Issue 111
AIList Digest Thursday, 20 Oct 1988 Volume 8 : Issue 111
Responses:
AI applications to building design and construction
Info on PROTEGE/RIME
CLOS & CommonLOOPS (2 Messages)
PFL
Concept Learning & ID3 (Quinlan) - in prolog (3 Messages)
Robotics; Universities offering
Classifier Systems
Expert systems and weather forecasting
AAAI-88 ordering info
----------------------------------------------------------------------
Date: 11 Oct 88 08:40:01 GMT
From: mcvax!ukc!strath-cs!glasgow!gilbert@uunet.uu.net (Gilbert
Cockton)
Subject: Re: AI applications to building design and construction
Professor Geoffrey Trimble at Loughborough University of Technology has
been associated with a number of construction industry expert systems.
He has a chapter in a forthcoming book on Knowledge Elicitation edited
by Dan Diaper of Liverpool Polytechic and published by Ellis Horwood.
Geoffrey Trimble is Professor of Construction Management in the
Department of Civil Engineering.
Loughborough University of Technology is in Loughborough,
Leicestershire, LE11 3TU, UK.
Domain experts have been heavily involved in the coding of some of
these systems, as well as the knowledge elicitation.
Committment from sponsors to use a system has proved to be major factor
in the succesful completion of a system. Not surprising, but important
to anyone developing something in a research setting.
--
Gilbert Cockton, Department of Computing Science, The University, Glasgow
gilbert@uk.ac.glasgow.cs <europe>!ukc!glasgow!gilbert
------------------------------
Date: Wed, 12 Oct 88 09:01:12 PDT
From: Mark Musen <MUSEN@SUMEX-AIM.Stanford.EDU>
Subject: Re: Info on PROTEGE/RIME
PROTEGE was the subject of my Ph.D. dissertation, and represents a
metalevel knowledge-acquisition tool that generates other
knowledge-acquisition tools that are custom-tailored for particular
application areas. There are not yet any journal articles in print
describing the work, although a paper on PROTEGE appears in the
Proceedings of the 1988 Workshop on Knowledge Acquisition for Knowledge
Based Systems (Banff, Canada). The most thorough description of PROTEGE
is in my dissertation ("Generation of Model-Based Knowledge-Acquisition
Tools for Clinical-Trial Advice Systems," Stanford University, January,
1988). Although I have no more copies available for distribution, the
dissertation can be ordered from University Microfilms (phone
800-521-0600).
A good description of the RIME methodology appears as a chapter by Judy
Bachant in a collection just edited by Sandy Marcus entitled "Automating
Knowledge Acquisition for Expert Systems" (Kluwer, 1988).
Mark Musen
Medical Computer Science Group
Knowledge Systems Laboratory
Stanford University
------------------------------
Date: Wed, 12 Oct 88 16:18 EDT
From: Brad Miller <miller@CS.ROCHESTER.EDU>
Reply-to: miller@CS.ROCHESTER.EDU
Subject: Re: CLOS & CommonLOOPS
Date: Mon, 10 Oct 88 19:03 O
From: Antti Ylikoski tel +358 0 457 2704
<YLIKOSKI%FINFUN.BITNET@MITVMA.MIT.EDU>
I would be very grateful if someone could let me know if an academic
license for the CLOS, the Common Lisp Object System, is available.
I would also like to know whom to contact to obtain it and the price.
Here's what I have in my distribution directory on PCL also known as CLOS:
I hope it helps.
****
Here is the standard information about PCL.
Portable CommonLoops (PCL) started out as an implementation of
CommonLoops written entirely in CommonLisp. It is in the process of
being converted to an implementation of CLOS. Currently it implements a
only a subset of the CLOS specification. Unfortunately, there is no
detailed description of the differences between PCL and the CLOS
specification, the source code is often the best documentation.
Currently, PCL runs in the following implementations of
Common Lisp:
Xerox Common Lisp (Lyric Release)
Symbolics (Release 7.2)
Lucid (2.0)
CMU
VAXLisp (2.0)
ExCL (Franz)
Ibuki Common Lisp (01/01)
HP Common Lisp
TI
Golden Common Lisp
Pyramid Lisp
Coral Common Lisp (Allegro)
There are several ways of obtaining a copy of PCL.
*** Arpanet Access to PCL ***
The primary way of getting PCL is by Arpanet FTP.
The files are stored on arisia.xerox.com. You can copy them using
anonymous FTP (username "anonymous", password "anonymous"). There are
several directories which are of interest:
/pcl
This directory contains the PCL sources as well as some rudimentary
documentation (including this file).
In the directory /pcl the files:
readme.text READ IT
notes.text contains notes about the current state of PCL, and some
instructions for installing PCL at your site. You should
read this file whenever you get a new version of PCL.
get-pcl.text contains the latest draft of this message
/pcl/doc
This directory contains TeX source files for the most recent draft of
the CLOS specification. There are TeX source files for two documents
called concep.tex and functi.tex. These correspond to chapter 1 and 2
of the CLOS specification.
/pcl/archive
This directory contains the joint archives of two important mailings
lists:
CommonLoops@Xerox.com
is the mailing list for all PCL users. It carries announcements
of new releases of PCL, bug reports and fixes, and general advice
about how to use PCL and CLOS.
Common-Lisp-Object-System@Sail.Stanford.edu
is a small mailing list used by the designers of CLOS.
The file cloops.text is always the newest of the archive files.
The file cloops1.text is the oldest of the archive files. Higher
numbered versions are more recent versions of the files.
*** Xerox Internet Access to PCL ***
Xerox XNS users can get PCL from {NB:PARC:XEROX}<PCL>
*** Getting a copy of PCL from ISI ***
ISI distribute PCL with its Common Lisp distribution. For further
information about this send a message to ACTION@ISI.EDU.
Send any comments, bug-reports or suggestions for improvements to:
CommonLoops.pa@Xerox.com
Send mailing list requests or other administrative stuff to:
CommonLoops-Request@Xerox.com
Thanks for your interest in PCL.
----
Brad Miller U. Rochester Comp Sci Dept.
miller@cs.rochester.edu {...allegra!rochester!miller}
------------------------------
Date: 13 Oct 88 09:47 PDT
From: hdavis.pa@Xerox.COM
Subject: CLOS & CommonLOOPS
You request an academic license for CLOS.
1. CLOS is the (now accepted) standard specification for an object-oriented
extension to CommonLisp. In fact, the X3J13 commitee has made it a
standard part of CommonLisp. It is not a particular product to be sold or
licensed.
2. The only currently available implementation of CLOS is called PCL, which
was (and is being) developed at Xerox PARC by Gregor Kiczales. You can ftp
it from arisia@xerox.com via anonymous login. This implementation is in
the public domain; no licensing agreements or payments are needed. Since
PCL was developed by Xerox, you are required to keep the copyright notices
on all the files.
3. The CLOS distribution list is commonloops.pa@xerox.com. Join up!
-- Harley
------------------------------
Date: Thu, 13 Oct 88 07:08 MDT
From: WOODL@BYUVAX.BITNET
Subject: re: PFL
Posting-Version: unknown; site unknown
Subject: re: PFL
In answer to the query about Finnin's frame rep language, I downloaded from
compuserve and have used it on a Mac with ExperCommon Lisp, and I also have
it running on an HP9000-350 in Common Lisp. If you look in the front of the
AI magazine, it will give you the compuserve acct.
Larry Wood, Brigham Young University
------------------------------
Date: 15 Oct 88 01:27:52 GMT
From: cs.utexas.edu!sm.unisys.com!aero!mcguire@ohio-state.arpa (Rod
McGuire)
Subject: Re: Concept Learning & ID3 (Quinlan) - in prolog
In article <395@uiag.UUCP> gerrit@uiag.UUCP (Cap Gerrit) writes:
>Does anybody out there in the world has an implementation of
>the ID3 algorithm of Quinlan
Prolog is one of the worst possible languages in which to write the
ID3 algorithm. I don't know if it can be written to be efficient.
The problem is that one needs to count up statistics in parallel and
then somewhat randomly access these statistics. While in prolog it
is possible to kludge up arrays with mutable elements, I will bet
that the overhead of this technique makes a prolog implementation
slower by a factor of at least 100 to say a fortran implementation.
Since the ID3 algorithm is usually applied to a moderate amount of
data (let's say, to construct a tree discriminating 10 classes from
analysis of 5000 attribute vectors each with 20 attributes that can
take on 1 of 5 values), I think that this performance difference can
make prolog implementation unusable. Also, I think any prolog version
that strives for efficiency is likely to be ugly and far removed
from the functional specification of the algorithm.
However I would love to see an analysis that proves me wrong. Below,
I present in lisp the central part of the ID3 algorithm - the
definition of the metric B(a U) which gives a value for the goodness
of attribute "a" as a discriminant for the set of
class-attribute-vectors "U". Following that is an array-based
implementation for the time consuming parts.
Let U be a set of class-attribute-vectors where each element "u" is
a "cav" data-structure defined as:
(cav-class u) = c, the class determined by vector u.
(range 1 to nc)
(cav-av u a) = v, the value for attribute a in vector u.
(range 1 to nv)
; the metric B for splitting U on attribute "a" is defined as:
(defun (B U a)
(/ (sum (v 1 nv) ; sum for v=1 to nv
(- (sum (c 1 nc)
(* (N c a v U)
(log (N c a v U))))
(* (S a v U)
(log (S a v U)))))
(size-of U)))
where
(S a v U) = number of elements u in U s.t.
(cav-av u a) = v
and
(N c a v U) = number of elements u in U s.t.
(cav-av u a) = v
& (cav-class u) = c
In order to avoid processing the elements in U over and over again,
it is reasonable to pre-compute N (as below) and define S in terms of N.
(defvar N (make-array (list nc na nv)))
(loop for v in U
do (loop for a from 1 to na
do (increment (aref N (cav-class v) a (cav-av v a)))))
------------------------------
Date: 16 Oct 88 01:39:59 GMT
From: quintus!ok@unix.sri.com (Richard A. O'Keefe)
Subject: Re: Concept Learning & ID3 (Quinlan) - in prolog
In article <39500@aero.ARPA> mcguire@aero.aerospace.org (Rod McGuire) writes:
>In article <395@uiag.UUCP> gerrit@uiag.UUCP (Cap Gerrit) writes:
>>Does anybody out there in the world has an implementation of
>>the ID3 algorithm of Quinlan
>Prolog is one of the worst possible languages in which to write the
>ID3 algorithm. I don't know if it can be written to be efficient.
>The problem is that one needs to count up statistics in parallel and
>then somewhat randomly access these statistics. While in prolog it
>is possible to kludge up arrays with mutable elements,
There is not the slightest need for mutable arrays in an implementation
of ID3 or similar algorithms, and Prolog is not at all a poor choice.
The Iterative Dichotomiser involves two kinds of steps:
(1) making a sweep through the training set collecting a random
sample of *incorrectly predicted* examples to add to the
"window" (this is the "Iterative" part)
(2) doing a kind of back-to-front radix sort on the contents of
the "window" to turn it into a decision tree (this is the
"Dichotomiser" part)
>Since the ID3 algorithm is usually applied to a moderate amount of
>data (let's say, to construct a tree discriminating 10 classes from
>analysis of 5000 attribute vectors each with 20 attributes that can
If you are working with tiny training sets like that, you don't need
ID3. Quinlan's innovation was the "windowing" technique -- forming
decision trees is nothing new, you will even find a tiny Fortran
implementation in Algorithm AS 165, JRSS -- which permitted him to
work with training sets having millions of examples. The "windowing"
idea can be applied to many induction schemes:
{Initialise}
set the window to a random sample of N1 examples.
{Induce}
induce a "rule" from the current window.
{Evaluate}
make a pass through the complete training set,
adding a random sample of N2 examples which are incorrectly
classified (if fewer than N2 misclassifications, take all).
If performance was adequate, stop.
{Iterate}
Go back to {Induce}.
Two references:
"Discovering rules by induction from large collections of examples",
Quinlan, J.R.
(in) Expert Systems in the micro-electronic age, (Ed. D. Michie)
Edinburgh University Press, 1979
"Learning efficient classification procedures",
Quinlan, J.R.
(in) Machine learning: an artificial intelligence approach
Eds Michalski, Carbonell, & Mitchell
Tioga press, 1983
The algorithm descriptions in these articles are quite clear.
------------------------------
Date: 18 Oct 88 16:27:45 GMT
From: mist!tgd@cs.orst.edu (Tom Dietterich)
Subject: Re: Concept Learning & ID3 (Quinlan) - in prolog
There is evidence that the windowing feature of ID3 does not provide
much benefit. Consult the following paper for details:
Wirth, J. and Catlett, J. (1988). Experiments on the costs and
benefits of windowing in ID3. In Proceedings of the Fifth
International Conference on Machine Learning, Ann Arbor, MI.
Available from Morgan-Kaufmann, Inc, Los Altos, CA. 87--99.
Here is the abstract:
"Quinlan's machine learning system ID3 uses a method called windowing
to deal economically with large training sets. This paper describes a
series of experiments performed to investigate the merits of this
technique. In nearly every experiment, the use of windowing
considerably increased the CPU requirements of ID3, but produced no
significant benefits. We conclude that in noisy domains (where ID3 is
now commonly used), windowing should be avoided."
The paper reports several studies involving training sets as large as
20,000 examples. The authors state that if you have the physical
memory to store the examples, it is best to avoid windowing.
Windowing seems to work best on noise-free training sets where there
are many redundant features. These turn out to be rather uncommon
although the initial domains in which ID3 was developed had these properties.
--Tom Dietterich
tgd@cs.orst.edu
Department of Computer Science
Oregon State University
Corvallis, OR 97331
------------------------------
Date: 17 Oct 88 15:01:53 GMT
From: att!mtuxo!rsn@bloom-beacon.mit.edu (XMRH2-S.NAGARAJ)
Subject: ROBOTICS; Universities offering
I am still getting e-mail for the above request. Since the number of
requests were quite a few, I decided that I would post a summary
of responses to my request seeking names of US universities offering
graduate degrees in Robotics.
The list (summary) provided below is a partial list since there was
considrable duplication in the info that I received.
Raj Nagaraj
>
Add the University Of Southern California to the list. It's Computer Eng
department offers a combined degree in Artificial Intelligence and Robotics
and the Computer Eng dept is ranked #3 in the nation by IEEE. I know I'm
a biased alumnus :-)
<
>
Raj,
I have been working in Robotics for a few years, and from
what what I know, there is now univ. that offers a degree in
Robotics although following institutions have very strong
programs and research in Robotics.
MIT, Stanford, CMU, Purdue, Univ. Of. Michigan, RPI, USC,
Univ. of Texas at Austin, Univ. Penn.
P.S.: Above list is by no means the final word on standing of
various schools in Robotics and there may even be some
<
Columbia University Computer Science Department has a very active and
prominent Robotics group
>
You might want to check out the University of Texas at Arlington...
(Dallas Fort Worth, Tx area). When I was a grad student there several
years ago, they got a *LARGE* grant to start a robotics facility,
with specific staffing in robotics. Don't know what the status is now,
though.
<
>
try carnegie-mellon.
<
>
One university is,
Carnegie Mellon University through the dept of Civil engineering.
The address is
Dept of Civil engg
Porter Hall
Carnegie Mellon Univ
Pittsburgh
PA 15213
<
>
You'll probably get this suggestion from others: check out Cornell.
The Robotics program here is very strong, the project headed by John
Hopcroft (1986 Turing Award). The emphasis seems to be on solid
modeling, motion planning and machine vision (judged by faculty
representation -- the project also supports some research associates
each year who expand the scope of subjects).
Disclaimer: I'm not even in Robotics, I'm in machine learning and I
know they're doing good robotics work!
<
>
I saw the list posted on the bulletin board in the Electrical Engineering
department a while back, so I don't remember. However some of the schools with
the top comp eng/electrical eng departments were :
Bekerley
Stanford
USC
MIT
UCLA
Caltech
I don't remember the rest.
<
------------------------------
Date: Mon, 17 Oct 88 11:12:13 PDT
From: rik%cs@ucsd.edu (Rik Belew)
Subject: Classifier Systems
Date: 11 Oct 88 14:05:20 GMT
From: steinmetz!boston!powell@itsgw.rpi.edu (Powell)
Subject: Classifier system software packages
Recently, I have read some interesting articles on induction and classifier
systems. To better understand their capabilities and functionalities,
I am looking for a free, classifier software package to experiment with.
Check with Rick Riolo at the Univ. Michigan. He has developed a
comprehensive, portable version of Holland's Classifier System
called CFS-C. I think you can reach him as Rick_Riolo@um.cc.umich.edu .
Richard K. Belew
rik%cs@ucsd.edu
Assistant Professor
CSE Department (C-014)
UCSD
San Diego, CA 92093
619 / 534-2601 or 534-5948 (messages)
------------------------------
Date: 18 Oct 88 17:04:53 GMT
From: hubcap!ncrcae!gollum!rolandi@gatech.edu (Walter Rolandi)
Subject: RE: Expert systems and weather forecasting
Regarding:
>I'm trying to find out if anyone is working on expert systems in weather
>forecasting. Names, addresses, references ... would all be welcome. With
>thanks in advance
>
> Laurence Moseley
You might want to get a copy of the proceedings of 17 JAIIO/PANEL '88
EXPODATA from SADIO, the Argentine Operations Research and Computer
Science Society. At their recent conference, a paper describing an
expert system weather forecaster was presented. The paper was entitled,
"Um sistema especialista para previsao de tempo". Its authors are,
V.H.de Avila Duarte and F.A.de Castro Giorno. I think the researchers
work at the National University of Brazil at Rio but I am not sure.
Walter Rolandi
rolandi@ncrcae.Columbia.NCR.COM
NCR Advanced Systems Development, Columbia, SC
------------------------------
Date: 19 Oct 88 16:48:45 GMT
From: mailrus!uflorida!haven!h.cs.wvu.wvnet.edu!b.cs.wvu.wvnet.edu!sip
ing@ohio-state.arpa (Siping Liu)
Subject: Re: AAAI-88
The following information is from the proceedins of AAAI-88:
To order the proceedings, write to:
Morgan Kaufmann Publishers, Inc.
P.O.Box 50490
Palo Alto, CA 94303
(414) 578-9911 or (415) 965-4081
For AAAI-88, $75/$56.25 AAAI members.
------------------------------
End of AIList Digest
********************
