Copy Link
Add to Bookmark
Report
AIList Digest Volume 4 Issue 093
AIList Digest Friday, 18 Apr 1986 Volume 4 : Issue 93
Today's Topics:
Queries - Machine Translation & Nontrivial Expert Systems,
Representation - Shape,
Project - Real-Time Machine Learning,
News - Max the Robot,
Review - Canadian AI Newsletter, March 1986
----------------------------------------------------------------------
Date: Wed, 16 Apr 86 08:42:55 GMT
From: gcj%qmc-ori.uucp@cs.ucl.ac.uk
Subject: Points Arising.
Re: Computer Dialogue (Vol 4 # 57 & 58).
I wonder how (many) psychological "strokes" were exchanged
in this conversation?
Re: Machine Translation (Vol 4 # 67 & 70).
Impressive claims are made for machine translation systems;
are there any systems that could produce a precis (summary)
of a large document?
Gordon Joly
ARPA: gcj%qmc-ori@ucl-cs.arpa
UUCP: ...!ukc!qmc-cs!qmc-ori!gcj
"They who often look far into the distance have excellent vision."
------------------------------
Date: 16 Apr 86 04:11:07 GMT
From: ihnp4!lzaz!psc@ucbvax.berkeley.edu (Paul S. R. Chisholm)
Subject: Non-trivial expert systems
< The electronic funds transfer is in the electronic mail. . . . >
For those of you who remember my Usenet posting in November, I'm
*still* looking for MS-DOS based expert systems for a review. I'll
shortly post a much longer list of such packages. For the moment, I
want to deal with something more fundamental.
I want to show the difference between an expert system and a simple
decision tree. Yet most of the examples (and the one software package)
I've seen, there *is* no difference . . . or at least, the one can be
transformed into the other.
Nearly all the expert system shells are actually based on
productions: if <foo> and <bar> and ... then <glarch>. The conditions
can be arbitrarily complicated, but usually involve testing "global
variables" against constant values, and possible a bit of trivial
arithmetic. The consequent assert that yet another global variable has
some constant value. There are wrinkles: the most common is having
variables that are "local" to a rule. This isn't strictly necessary,
but saves a lot of tedious, repetitious rule writing.
In theory, it's always possible to treat each production as a node,
and construct a tree of questions without knowing any of the answers
ahead of time. This disturbs me, though I realize generating a
meaningful tree is nontrivial.
In practice, damn near *everyone* draws that tree first, then
writes the rules. This is missing the point! If your "expert"
knowledge is that trivial, you don't need logic, just a branch follower.
I tried drawing a "subway network", and writing rules of how to get
from one station to another. This isn't very instructive: forward
chaining doesn't find anything like an optimal solution, and backwards
chaining takes every damn trip. (*sigh* - Can you tell my first AI
course was taught out of Nilsson's PROBLEM-SOLVING METHODS IN ARTIFICIAL
INTELLIGENCE? Nilsson thought all AI reduced to search.)
I've seen one sample expert system that didn't reduce to a tree,
and I'm not at all sure simple shells can solve it! C. J. Culbert (sp?)
of NASA sent me a wonderful "monkey and banana" system that requires
about eighty inferences. The solution involves things like "move the
ladder under the red box, climb up and get the green key out, climb down
and move the ladder under the green box, climb up, unlock the green box
with the green key, get the blue key . . ." Very nice, but the expert
system shells I've seen can't handle time, e.g., "first move the ladder
to A, then once you've finished this subtask, move the ladder to B".
Once a value (e.g., ladder location) is deduced, it's hard or impossible
to change. "Undoing" isn't always kosher either: if I have a glass of
milk, I can quench my thirst or make butter, but once I've done one . . .
What are my points?
+ First and foremost, I'd like a expert system that can be solved with
simple productions. It shouldn't be an example provided by a vendor I'll
review; that'd potentially give his or her product an edge.
+ Second, I'd like some reassurance that production-based expert
systems go beyond decision tree programs. Please don't flame to the net
on this one. I'm posting to both Usenet and Arpanet groups; if you send
me mail (I'm reachable from both), I'll summarize and repost.
Sorry to ramble, thanks in advance for your help, and I'll post the
MS-DOS expert systems as soon as I can.
--
-Paul S. R. Chisholm, UUCP {ihnp4,cbosgd,pegasus,mtgzz}!lznv!psc
AT&T Mail !psrchisholm, Internet mtgzz!lznv!psc@topaz.rutgers.edu
The above opinions may not be shared by any telecomm company.
------------------------------
Date: Sun, 13 Apr 86 11:45:51 EST
From: Jim Hendler <hendler@brillig.umd.edu>
Subject: shape
Ken-
The work done so far on shape falls loosely into the classes of
``solid modeling'' and ``functionality.'' The questions you ask in the
AI digest were more from the solid modeling camp -- i.e. Can I put this
peg in this hole? I'm not incredibly familiar with that literature, but some
recent work in automated manufacturing (a big buzzword these days) has included
this type of work. Dana Nau (dsn@maryland) has a paper which will be appearing
in a new journal (I forget the name, it's one of the new ``applied AI''
journals) which discusses a frame based approach to solid modeling. He also
has several tech. reports at U of Maryland on such a topic.
The functionality work is more my line (I have a couple of students starting
thesis work on it now). This research has traditionally focused on
recognizing objects from functional information (i.e.
from the dialogue
John: Do you know the time
Mary: 1 PM
we infer that Mary has a watch)
I'm looking into expanding it in a direction that starts to interact more
with the solid modeling stuff. I'd like a description of a ``cleaver''
such that I could INFER that it is a weapon. At the moment most systems
(including my own planners) must have this information stored explicitely.
It is this my students are now looking into.
Hope this is of some help
Jim Hendler
Ass't Professor
University of Maryland
Hendler@maryland
------------------------------
Date: Wed, 16 Apr 86 13:28:00 est
From: Stanley Letovsky <letovsky@YALE.ARPA>
Subject: shape
To: laws@sri-iu
Ernie Davis at NYU was working on this problem; I read a
manuscript of his entitled something like "Buttons, Rakes and Rings"
last year. I don't know if he published it anywhere but you might ask
him what the status of the work is. He was trying to define an ontology
for qualitative and loose quantitative reasoning about shape. -Stan
------------------------------
Date: Thu, 17 Apr 86 13:45:04 PST
From: Scott Turner <srt@LOCUS.UCLA.EDU>
Reply-to: srt@ucla-cs.UUCP (Scott Turner)
Subject: Re: Shape
Jack Hodges of the UCLA Artificial Intelligence Laboratory is working on
EDISON, a program that invents mechanical devices. Jack is in England this
week presenting his work at a conference, so I'm standing in for him in
presenting some references that might be useful for understanding "hooks
and rings".
First of all, the Edison project looks at naive invention: the kind of
tinkering that backyard inventors or (one supposes) children do. There is
no complex of mathematical forces in the project. It focuses instead on
the issues of creativity and problem-solving. How _does_ one get that
great idea? The reference:
EDISON: An Engineering Design Invention System Operating Naively,
Hodges, Dyer, Flowers, Tech Report UCLA-AI-85-20, Dec. 1985
There has been a lot of work done on naive physics. The reference I'm
aware of is:
Hayes, P.J., "The Second Naive Physics Manifesto," pp. 467-486
in _Readings in Knowledge Representation_, ed. Brachman & Levesque,
Morgan Kaufman, 1985
There has been some work done on object representation, primarily:
Lehnert, W.G., _The Process of Question Answering_, LEA 1978.
(see Chapter 9).
Rieger, C. "An Organization of Knowledge for Problem Solving and
Language Comprehension", pp. 487-508, _Readings in Knowledge
Representation_...
Wasserman, K. and Lebowitz, M., "Representing Complex Physical
Objects," _Cognition and Brain Theory_, 6(3), pp. 259-285 (1983)
Finally, for the particular area of children's problem solving:
DeBono, E., _Children Solve Problems_, Penguin, NY 1980.
And not to overlook work by Forbus, DeKleer and Brown, though I won't
bother to type in the cites.
That should get you started.
Scott R. Turner
ARPA: (now) srt@UCLA-LOCUS.ARPA (soon) srt@LOCUS.UCLA.EDU
UUCP: ...!{cepu,ihnp4,trwspp,ucbvax}!ucla-cs!srt
FISHNET: ...!{flounder,crappie,flipper}!srt@fishnet-relay.arpa
------------------------------
Date: Wed, 16 Apr 86 15:46:06 pst
From: malkoff@nprdc.arpa (Don Malkoff)
Subject: real-time machine learning
The "REAL-TIME MACHINE LEARNING LABORATORY" has been established
at the Navy Personnel Research and Development Center, Code 71,
San Diego, CA 92152-6800.
Ongoing work includes:
1. Real-time fault detection and diagnosis in complex control
systems, involving random time variability, and
2. Automated sonar detection and classification.
These and other related project areas make use of machine learning
techniques.
For information contact Don Malkoff, (619) 225-6617.
------------------------------
Date: WED, 10 JAN 84 17:02:23 CDT
From: E1AR0002%SMUVM1.BITNET@WISCVM.WISC.EDU
Subject: Max the Robot
Source: KRLD News, Dallas
A man, for unknown reasons, baricaded himself into an apartment. Since he
had a history of explosive law violations, the police did not want to
enter the apartment. They did not even know if he was still alive as
he was not talking to them and a friend said he was extremely depressed.
They sent in Max the Robot, a tank-like entity complete with camera and
manipulator. It smashed through the window and pushed a drape aside
when the man, in astonishment, left the apartment peacefully without
any shots being fired. [...]
------------------------------
Date: WED, 10 JAN 84 17:02:23 CDT
From: E1AR0002%SMUVM1.BITNET@WISCVM.WISC.EDU
Subject: Review - Canadian AI Newsletter, March 1986
Canadian Artificial Intelligence March 1986 Issue Number 7
Biographies of new officers, letter from someone protesting the research
effort into emulating the brain.
Discussion of Intelligent Computer-Aided Instruction research in
Canada, bemoaned the scarcity of AI researchers in Canada with psychology
training.
Specific Projects:
University of Alberta: automate the teaching of statistics first course
Bell Northern Research: modelled personality interactions between expert
and builder and developed an artificial expert in horticulture
University of Calgary: training materials for medical students including
the use of videodisc
University of Calgary: user modelling project
Concordia University: application of Pask's "Conversation Theory" to
team decision support and a course assembly and tutorial environment
which runs on Apples
ForceTen enterprises: expert system and natural language interface
to be part of their courseware development system. The product runs
on IBM PC's
National Research Council: Project to develop adaptive computer-
based training with natural language interface.
University of Saskatchewan: Project to develop Lisp teacher and
conception corrector, programming environment for first year students
University of Waterloo: attempt to model students in a manner independent
of system used
__________________________________________________________________________
The Canadian Society for Fifth Generation Research has signed an
agreement of understanding with the Japanese ICOT for exchange of
technical informaiton and research meetings. This is the first
collaboration that ICOT has signed with a foreign organization.
__________________________________________________________________________
Machine Vision International has established its head office in Ottawa.
__________________________________________________________________________
"Interest in artifical intelligence and expert systems is relatively
new in Canada"
From a Canadian government report on expert systems by the Office of
Industrial Innovation, Department of Regional Industrial Expansion
October 1985
__________________________________________________________________________
Reviews of
Artificial Intelligence: A Personal, Commonsense Journey by
William R. Arnold and John S. Bowie (an introduction to AI
for lay readers but got a poor review.)
Progress in Artifical Intelligence by Luc Steels and John A. Campbell
(collection of papers from the 1982 European Conference on artificial
intelligence)
(Some short reviews as well)
__________________________________________________________________________
List of AI Tech reports from Canadian Universities
__________________________________________________________________________
Summary of the third and fourth University of Waterloo-University of
Ontario AI workshops , Queen's University Expert System Workshop
__________________________________________________________________________
Obituary for Paul A. Kolers, a researcher in psychology of visual perception
------------------------------
End of AIList Digest
********************
