AIList-REQUEST@SRI-AI.ARPA (AIList Moderator Kenneth Laws) (10/14/85)
AIList Digest Monday, 14 Oct 1985 Volume 3 : Issue 145
Today's Topics:
Opinion - Military Support & AI Hype & CS and AI Definitions,
AI Tools - Workstations & Lisp vs Prolog Implementation Facts,
Call for Papers - IEEE Systems, Man and Cybernetics
----------------------------------------------------------------------
Date: Thu 10 Oct 85 12:07:50-PDT
From: Rich Alderson <ALDERSON@SU-SCORE.ARPA>
Subject: re: military flame
Like it or not, the military CAN take credit for the interstate highway system,
and the "civilian highway system in place before the interstates" as well--at
least, I ASSUME that you are referring to the U. S. highway system. Both were
built under the mandate of the Constitution of the United States (Philadelphia,
1787), which did not view "internal defence [as] a standard rationale for
better internal transportation" but rather saw good internal transportation as
vital to both the internal and external defense of a nation.
Please note that the above is not a matter of opinion. My own opinions are
just that, and I've had enough grief in my life for expressing them publicly.
Obviously, I make NO claim to represent anyone else's opinions if I refuse to
state my own.
Rich Alderson
------------------------------
Date: Thu 10 Oct 85 10:12:48-PDT
From: WYLAND@SRI-KL.ARPA
Subject: AI Hype is unavoidable
Part of the problem of AI hype is unavoidable: it is a result of
the definition of the field, which assumes that, once the problem
is defined, its solution is trivial.
Artificial Intelligence and Computer Science represent two
complementary approaches to the use of computers. AI is problem
oriented, and CS is solution oriented. In CS, one assumes that
the problem is (or is capable of being) well understood, and the
task is to design a good (clean, efficient, fast, etc.) solution
to the problem. In AI, the underlying assumption is that the
solution will be trivial once the problem is understood, and that
the task is to understand the problem.
These two approaches are reflected in the names of the field.
Computer Science is the science of using computers, a clear,
hard, objective definition about the use of a tool to solve
problems. Artificial Intelligence is concerned with the
simulation of intelligence, a fuzzy, open, subjective definition
about the study of the problem of simulating a human behavior
called intelligence for which there is no clear, generally
accepted definition.
Each field encounters problems when its assumption is violated.
In CS, programming disasters can result if you start coding
before you have defined the problem. Therefore, CS has developed
structured programming, programming specifications, etc. tools to
insure that the problem *is* well defined before the solution is
attempted.
In AI, trouble starts when, after the problem is understood, the
solution is *not* trivial in terms of computer performance, such
as execution time and memory space. Then, you start hacking the
algorithms in order to get results in a finite amount of time and
memory, hoping that you are not betraying your understanding of
the problem. Therefore, AI has developed extensive editors,
debuggers, windows, etc. in an attempt to insure that the
implementation of the solution remains trivial.
I believe that much of the AI hype problem stems from the
unstated assumption of trivial solutions. Profound and
impressive insights expressed as toy problems typically do not
"scale up" to the real world. An AI researcher involved in this
situation is embarrassed but not humiliated: the original
research on the problem is still valid; there is just a
"temproary problem in creating a practical solution." This
obviously creates frustration for the customers who thought they
were buying a solution of the problem rather than an
understanding of it.
If the above is true, the problem of AI hype will not go away
until the field develops enough solid understanding of the
problem of intelligence to change its name from AI to a solution
oriented name, like Machine Learning, etc. We are probably in
the same position now that physical science was before Galileo
and Newton when it was called Natural Philosophy - full of
metaphysics, passionate argument, and conflicting data (i.e.,
where the action is).
Dave Wyland
------------------------------
From: CONNOLLY CHRISTOPHER IAN <CONNOLLY@ge-crd.arpa>
Subject: AI Definition and Tools
1) I can't help but think that a cause of the recent arguments on AI
hype rests in the question "What is AI?". Note that I say >>**"A"**<<
cause. Definitions, anyone?
2) AI Machines - My observation is that the startup time on a 3600 without
help is quite long. There are a few people here who have taken a Symbolics
Lisp course and seem to have picked up on the stuff much more quickly (2 or
3 months?). Once you know how to use the machine though, I think it's a
far better programming environment than VAX/VMS. I've seen nothing on a
VAX that parallels the Window Debugger (wherein the entire stack can be
dissected), the Inspector (wherein your data structures can be dissected),
and the Flavor Examiner (wherein your data types can be inspected). The
latter two are also a great help when you have no source. I think it speeds
up my programming by a factor of 5, at least. Anyway, that's yet another
opinion...
------------------------------
Date: Sat, 12 Oct 85 12:28:27 EDT
From: George J. Carrette <GJC@MIT-MC.ARPA>
Subject: Lisp vs prolog, implementation facts
Actually, all three of the Symbolics, LMI, and TI lispmachines give
the lisp-level system programmer access to extremely low level data
type and stack operations, there is no need to go to microcode for
that sort of thing. The LM-PROLOG that I maintain from time to time in
the Symbolics and LMI environment creates its own datatypes by the
usual lisp punning techniques, fooling with locatives, forwarding
pointers and the internals of CDR-CODED lists without needing
microcode support, and amazingly keeping a good deal of
transportability. Such is the ubiquity of certain hacks of lispmachine
implementation. Microcode is used, optionally, only for hand-coding
functions that are also written in lisp. The LM-PROLOG technique is
in the class of CONTINUATION PASSING techniques of prolog
implementations, as described for example in a chapter of Sussman and
Ablesons "Structure and Interpretation ..." book. This kind of
technique has more overhead associated with creation of real function
evaluation frames and such, (at least on a simple stack-machine) and
observably gets about 1/2 or 1/3 the BENCHMARK performance of a lower-level
machine-model implementation such as described by Warren. Overall
performance of a practical prolog "expert-system" may depend more
on virtual memory performance considerations of course rather than
what happens in the register-usage-mostly situation of some benchmarks.
Also consider that commercial systems put into production often have
assembly language coding of important subroutines, so high-level/low-level
language interface issues are important. The continuation passing
technique provides a more natural and efficient interface to
"assembly language" (i.e. LISP on a LISPMACHINE) than other models.
When talking about a commercial lispmachine it is important to think in
terms of LISP as a COMPUTER ENGINEERING technique rather than as having
anything to do with AI programming in particular.
------------------------------
Date: 12 Oct 1985 11:04-EDT
From: milne@wpafb-afita
Subject: call for papers - IEEE SMC
CALL FOR PAPERS
IEEE TRANSACTIONS ON SYSTEMS, MAN AND CYBERNETICS.
SPECIAL ISSUE ON
Causal and Strategic Aspects of Diagnostic Reasoning
Papers are solicited for a special issue of IEEE Transactions on Systems,
Man and Cybernetics that will be devoted to the topic, "Causal and Strategic
Aspects of Diagnostic Reasoning." Dr. Robert Milne, Army Artificial
Intelligence Center will be the guest editor of the special issue.
While it is expected that the research to be reported will be typically
backed up by concrete analyses or system building for real-world diagnostic
problems, the intent is to collect the most sophisticated ideas for
diagnostic reasoning viewed as a generic collection of strategies. Articles
should attempt to describe the strategies in a domain-independent manner as
much as possible. Articles that merely describe a successful diagnostic
expert system in a domain by using well-known languages or strategies will
typically not be appropriate. Papers reporting on psychological studies,
epistemic analyses of the diagostic process, elucidating the strategies of
first-generation expert systems, descriptions of specific diagnostic systems
that incorporate new ideas for diagnostic reasoning, learning systems for
diagnosis are some examples that will be appropriate. It is expected that
most articles will typically concentrate on some version or part of the
diagnostic problem, so it is important that the paper state clearly the
problem that is being solved independent of the implementation approaches
adopted.
Papers will be reviewed carefully by referees selected by the Transactions
Editorial Board. Five copies of the manuscript should be submitted to Dr.
Robert Milne at the following address by January 15. It will be helpful if
people who intend to submit manuscripts for consideration can let the guest
editor know of their intent as soon as possible via arpanet or telephone.
Submit papers by January 15th, 1986 to:
Dr. Robert Milne
US Army AI Center
HQDA DAIM-DO
Washington, D.C. 20310-0700
phone:(202)-694-6913
arpa: milne@wpafb-afita
Author's Timeline:
15 January 1986 Papers Due
15 April 1986 Notificationo of acceptance/rejection
1 June 1986 Final Manuscripts due
November 1986 Publication
------------------------------
End of AIList Digest
********************