Copy Link
Add to Bookmark
Report
AIList Digest Volume 2 Issue 095
AIList Digest Wednesday, 25 Jul 1984 Volume 2 : Issue 95
Today's Topics:
Expert System - Ben-Zoma Mailing List & Plant/Induce reference,
AI Tools - XLISP Sources,
Parapsychology - Justification & Mailing List,
Adminstrivia - Lab Reports and Project Descriptions,
Seminar - Learning State Variables,
Project - Engineer's Assistant for Fault Diagnosis
----------------------------------------------------------------------
Date: Mon, 23 Jul 84 15:00:46 pdt
From: Angela Shiflet <shiflet@lll-crg.ARPA>
Subject: Ben-Zoma Mailing List
RESEARCH ANNOUNCEMENT
Ben-Zoma is a knowledge- and experience-based scientific/engineering advisor
which converses in technical English. It learns by abstracting
solution methods and language understanding derived from assisting users.
Requests, once converted to a frame-like form, are dispatched to a
distributed consortium of experts. These sources of expertise (e.g. MACSYMA
on LISPM, LM on a Cray, SMP on a VAX, as well as new experts) are callable
through drivers written in LISP. Ben-Zoma will create and dispatch code for
appropriate special-purpose processors (e.g. Crays, Cosmic Cubes, data flow
machines, and VAX arrays). Graphical displays of numerical data will be
incorporated where appropriate.
Work on this project has begun under the direction of Dr. Ted Einwohner
at the Univeristy of California, Lawrence Livermore National Laboratory,
Computing Research Group, under contract to the Department of Energy.
Comments and suggestions are welcomed:
ben-zoma-discussion@lll-crg.ARPA
To be added to this list send mail to:
ben-zoma-discussion-request@lll-crg.ARPA
------------------------------
Date: Mon 23 Jul 84 14:06:09-PDT
From: Michael Walker <WALKER@SUMEX-AIM.ARPA>
Subject: Plant/Induce reference
Roy,
The Plant/Induce work was done by Ryzard Michalski at the
University of Illinois, Urbana, Illinois. The paper I have is:
Michalski, R.S., and Chilausky, R.L. Learning by being told and learning
from examples: an experimental comparison of the two methods of knowledge
acquisition in the context of developing an expert system for soybean
disease diagnosis. International Journal of Policy Anaysis and Information
Systems, Vol 4, No. 2, 1980.
I believe he also published a version in 1981 in the International
Journal of Man-Machine Studies.
If you are interested in systems that learn rules automatically,
you might also want to look at Peter Politakis' SEEK system described in
Artificial Intelligence, January 1984.
Mike Walker
walker@sumex-aim.arpa
------------------------------
Date: 20 Jul 84 8:28:06-PDT (Fri)
From: hplabs!hao!seismo!rlgvax!cvl!jcw @ Ucb-Vax.arpa
Subject: XLISP sources posted to net.sources
Article-I.D.: cvl.1193
After dozens of requests reached me, I decided to post the sources to
David Betz' XLISP interpreter. It is written entirely in C, commented
uncommonly well, fairly portable, and has some rather neat features
including primitives for object-oriented programming.
The program is in the public domain, Copyright by David Betz.
Jay Weber
..!seismo!rlgvax!cvl!jcw
..!seismo!rochester!jay
jay@rochester.arpa
------------------------------
Date: 20 Jul 84 13:10:50-PDT (Fri)
From: hplabs!hpda!fortune!amd!decwrl!dec-rhea!dec-pbsvax!cooper@Ucb-Vax.arpa
Subject: Why discuss super- and para-normal phenomena
Article-I.D.: decwrl.2741
"Alex.Rudnicky" asks:
"It may be fun to speculate about the super-normal and the para-normal,
but what does it have to do with AI?"
Answer: A number of things. First let's discuss the "super-normal", a phrase
which I will take to refer to "exceptional human performance."
1) Like it or not, we're stuck (at least for now) with describing the "I" in
"AI" in terms relative to human performance. In most instances average human
performance is all that is required of our programs, but sometimes, the
exceptional is called for. Human performance serves as a guide to what CAN
be done, because it HAS been done. According to the standard assumptions of
AI, if humans can do it, so can sufficiently powerful, well-programmed
machines. If humans cannot, then there may well be a NP-hard or worse problem
involved. For example: can the type of associative memory retrieval associated
with human intelligence be merged effectively with total recall? Or must the
information available by free-form association always be strictly limited?
If total well-indexed recall is done by at least one human being than it can
presumably also be done by a machine. Otherwise, it remains an open question.
2) An understanding of human information processes is seen as either absolutely
necessary or (depending on what "school" of AI philosophy you subscribe to) at
least very useful to AI programming. If your model of human information
processing cannot account for exceptional human performance then it is either
incorrect or incomplete. Knowledge that some adults have "eidetic" memory
(near perfect image memory) may well be critical to understanding all memory.
Knowledge that a large percentage of children (perhaps all if we could test
them young enough) have eidetic memory and then lose it as they grow up,
should be taken into account in theories of information acquisition from a
near "tabla rasa" state.
In other words, knowing the limits of human information processing is, in the
long term, very important to the field. In the short term, given our
distance from our ultimate goals, the need is less critical. Some relatively
brief exchanges in an informal forum seems appropriate to keep people thinking
about it.
Which brings us to the paranormal. First of all, claims of paranormal
abilities would seem to be included as exceptional human information processing
capabilities. My previous comments about the "super-normal" applies.
Some effort (probably not much from the viewpoint of current AI) should be
applied to determine whether or not, in general, the phenomena exist and if so
whether they should be considered as an exceptional cognitive skill or only
an exceptional perceptual skill. In the latter case its relevance is much
reduced. (My own opinion is: the experimental evidence makes it much more
likely than not that psi exists, and I would tend to see it as sensory/motor
rather than cognitive, though ESP seems to share many characteristics with
memory and dreaming).
Second, paranormal phenomena brings considerable doubt to the basic assumption
of AI: that human cognitive function can be explained as information processing
and therefore can be simulated or approximated by a sufficiently powerful and
well programmed artificial symbolic processor. This is of minor pragmatic
concern if psi is simply a rarely used IO channel. Several parapsychologists
have theorized, however, that psi functioning as perception is simply "leakage"
from its fundamental purpose in the organism; to wit, an essential part of
one or another cognitive function. Candidate functions I have seen mentioned
are intuition, creativity and memory. If so, (and I personally doubt it) then
human cognitive processing may not be simulatable on a Turing machine but only
on a Turing machine plus (you'll pardon the expression) oracle.
IN SUMMARY: while it seems premature to spend too much time now worrying about
exceptional (particularly paranormal) human performance, the AI community
should remain aware of this area. It might become very important to us and
we should not be caught unaware.
Topher Cooper
USENET: ...decvax!decwrl!dec-rhea!dec-pbsvax!cooper
ARPA: cooper%pbsvax.DEC@decwrl.ARPA
------------------------------
Date: 20 Jul 84 13:12:05-PDT (Fri)
From: hplabs!hpda!fortune!amd!decwrl!dec-rhea!dec-pbsvax!cooper@Ucb-Vax.arpa
Subject: Continuation of ESP discussion available.
Article-I.D.: decwrl.2742
On May 21 Ken Laws posted a reply summarizing an article from Dr. Dobb's
containing a theory of ESP. I replied with a detailed criticism of the theory
(at least as summarized) and suggested that further contributions be mailed
to me rather than posted. I have put together a single file containing:
1) A repeat of the original pair of articles.
2) Some corrections/updates to my article.
3) The five responses I got from my article.
4) My replies to those five responses.
The compilation is 745 lines long. Anyone who is interested in getting a
copy should send me mail requesting it. Unless you request otherwise you
will also be added to a mailing list to receive the next round, if there is
one. Your name and location will be kept confidential. Any submissions for
the next round will be public unless you request that I remove your name and
location from the posting.
Topher Cooper
USENET: ...decvax!decwrl!dec-rhea!dec-pbsvax!cooper
ARPA: cooper%pbsvax.DEC@decwrl.ARPA
------------------------------
Date: Tue 24 Jul 84 23:17:30-PDT
From: Ken Laws <Laws@SRI-AI.ARPA>
Reply-to: AIList-Request@SRI-AI
Subject: Lab Reports and Project Descriptions
Moderating the digest has become sufficiently routine that I
can devote increased time and creative energy to shaping the
contents. I can thus accept submissions to new digest
"departments" for lab reports, project summaries, and abstracts
of recent or current work. My intention is to better inform
readers by publishing "promotional" material originally written
for other audiences. This is similar to AIList's circulation
of seminar abstracts, a feature that I consider highly successful.
I therefore encourage list members to send abstracts of their
technical reports, conference papers, and journal articles to
AIList. Usenet members should preferably send such items directly
to AILIST@SRI-AI rather than through net.ai, although the usual
mechanisms will operate to prevent double distribution of net.ai
submissions. I shall screen the items and publish them in
coherent groups as the digest load permits. The digesting
delay for such material may be several weeks, but I shall try
to keep the backlog to a reasonable size by publishing special
issues of abstracts as necessary.
I shall also pass along a limited number of carefully edited
messages derived from Arpanet-distributed position postings and
similar material. I shall take considerable liberties with the
arrangement and format of the original texts without inserting
[...] annotations, and shall suppress explicit solicitations
(although the unofficial custom on the Arpanet has been to permit
such commercialism by academic institutions). I shall also try
to avoid repeating boilerplate lab descriptions that AIList has
already published. Nonacademic institutions may [occasionally]
submit similar promotional material so long as Arpanet standards
are respected. My decision to distribute such material will be
based solely on interest to the general AIList reader, not on the
potential benefit of filling AI-related positions.
Please don't dump all of your archived blurbs on me today
or tomorrow; we have plenty of time. I should like to see
the submissions dribble in over a period of >>years<<, so
wait until an appropriate opportunity (e.g., when a related
discussion comes up in the digest or when your dissertation
goes to press). Eventually we shall reach a steady state
with material being submitted as it is produced for other
purposes.
I anticipate that these news items will require more editing than
normal submissions, particularly the lab reports derived from
promotional material. You can simplify my job if you provide a
meaningful "Subject:" line such as the "Seminar - ..." headers I
have been distributing. Keywords such as "Abstract" and
"Project" should be followed by a very short title that readers
can use to screen the messages. The submissions themselves
should be concise and closely related to the interests of the
AIList readership. (The enthusiasm of your colleagues, bosses,
and sponsors for your 200 papers on educational parapsychology
may not be shared by a general audience.) Please include
sufficient "Contact:" information (e.g., address and phone number)
that I shall not have to help readers wanting further information.
I shall be fairly strict about screening material I consider
marginal, and should appreciate your consideration in minimizing
this unpleasant part of my responsibilities. Rejections will be
handled by "form letter", and generally will not include detailed
justifications. I hope that few will interpret such a notice
as an invitation to debate or the opening round in a series of
negotiations.
Comments to AIList-Request@SRI-AI on this new policy will be
helpful in determining whether this experiment should be modified
or discontinued. (Your silence will be interpreted as lack of
disapproval.) I shall keep list readers informed of any
significant trends in the expressed opinions.
-- Dr. Kenneth I. Laws
AIList Moderator
------------------------------
Date: Tue 24 Jul 84 12:36:44-PDT
From: Juanita Mullen <MULLEN@SUMEX-AIM.ARPA>
Subject: Seminar - Learning State Variables
[Forwarded from the Stanford SIGLUNCH distribution by Laws@SRI-AI.]
DATE: Friday, July 27, 1984
LOCATION: Chemistry Gazebo, between Physical & Organic Chemistry
TIME: 12:05
SPEAKER: Tom Dietterich
Heuristic Programming Project
Stanford University
TOPIC: Learning About Systems That Contain State Variables
It is difficult to learn about systems that contain state variables
when those variables are not directly observable. This talk
formalizes this learning problem and presents a method called the
iterative extension method for solving it. In the iterative extension
method, the learner gradually constructs a partial theory of the
state-containing system. At each stage, the learner applies this
partial theory to interpret the I/O behavior of the system and obtain
additional constraints on the structure and values of its state
variables. These constraints can be applied to extend the partial
theory by hypothesizing additional internal state variables. The
improved theory can then be applied to interpret more complex I/O
behavior. This process continues until a theory of the entire system
is obtained. Several sufficient conditions for the success of this
method will be presented including (a) the observability and
decomposability of the state information in the system, (b) the
learnability of individual state transitions in the system, (c) the
ability of the learner to perform synthesis of straight-line programs
and conjunctive predicates from examples and (d) the ability of the
learner to perform theory-driven data interpretation. The method is
being implemented and applied to the problem of learning UNIX file
system commands by observing a tutorial interaction with UNIX.
------------------------------
Date: Thu 12 Jul 84 04:33:14-PDT
From: MEISENSTADT@USC-ECLB.ARPA
Subject: Project - Engineer's Assistant for Fault Diagnosis
Human Cognition Research Laboratory, Open University, Milton Keynes, England
LOCATION: 50 miles north of London (between 38 and 60 minutes
by train, depending upon the service).
COMPUTING FACILITIES:
Symbolics 3600 Lisp Machine (for the dedicated use of this project), VAX
11/750 running NIL, Prolog, and POP-11, and dedicated lines to the Open
University's three DECsystem-20's running Interlisp, Maclisp, Edinburgh
Prolog, etc. ALL TERMINALS IN OUR LAB ALSO HAVE DIRECT ARPANET ACCESS.
ACTIVE AI PERSONNEL: Two tenured staff members, (Marc Eisenstadt and Jon
Slack), three research fellows, three Ph.D. students, and one consultant
programmer, all of whom constitute the Human Cognition Research
Laboratory's mainstream AI people. The OU also has other active
AI researchers on site, working under Max Bramer in the Maths Faculty
and Tim O'Shea in the Institute of Educational Technology. We are a
vigorous and growing group of researchers, and our current manageable
size enables us to offer the best AI computing facilities of any
academic institution in Europe.
PROJECT: "A Knowledge Engineer's Assistant for Constructing
Knowledge Based Fault Diagnosis Systems"
PROJECT SYNOPSIS: We are building a repertoire of rapid prototyping tools
intended to speed up both the analysis of verbal protocols (such as
those obtained during interviews with domain experts), and also the
encoding of elicited knowledge into implementable form. The applied
aspect of this work is the design of intelligent cross-referencing
and browsing facilities linked directly to a coding window. The
theoretical aspect of this work is an investigation of the process
of theory-formation as typified by modern day Knowledge Engineers.
CONTACT: MEISENSTADT@USC-ECLB, or telephone (international) 011-44-908-653149
or 011-44-908-661566.
------------------------------
End of AIList Digest
********************
