Re: No restraint on creativity

Matthew L. Ginsberg <ginsberg@t.stanford.edu>
Date: Wed, 27 Nov 91 06:29:25 PST
From: Matthew L. Ginsberg <ginsberg@t.stanford.edu>
Message-id: <9111271429.AA01641@t.Stanford.EDU>
To: sowa@watson.ibm.com
Subject: Re:  No restraint on creativity
Cc: INTERLINGUA@isi.edu, KR-ADVISORY@isi.edu, SRKB@isi.edu

Hi John:

You write:

  When you quoted those two lines from my previous note,
  
  > You can argue all you want about how standards might inhibit certain
  > lines of research, but the alternative is even more stultifying.
  
  you omitted the immediately preceding line, which said that unless
  we propose something better, "the winner by forfeit will be SQL2." ...
  which would inhibit EVERY line of research ... I don't believe that
  a logic-based standard for information interchange would inhibit
  ANY line of research ...

But the spirit of your remark is that it is better for us to allow
some lines of research to be inhibited by an FOL standard than to
allow all lines of research to be inhibited by an SQL2 standard.  This
still strikes me as chilling -- and let me point out that I am
squarely in the group of people likely *not* to be affected by an
FOL-based standard.  But I have a fundamental responsibility to stand
up for people outside of that camp, misguided though they may be.  : )

  In fact, I liked your recent article in AI Magazine very much;
  I didn't interpret it as being against standards, but as quite
  a reasonable framework for standards.

Thank you; I liked the article as well.  : )  But I would continue to
be happier if we could drop the word "standards" from what we are doing
and discussing.  The use of this word is going to make our funders happier
in the short run but *very* unhappy in the slightly longer run, and we are
all going to suffer.
  
   1. At a minimum, the standard should include first-order logic.
      Every version of nonstandard reasoning includes classical FOL
      as a limiting case.

Not so.  Many people work with proper subsets of FOL.  They may never
be able to "translate out" of an FOL-based standard.
  
   2. Since almost every AI system has a type hierarchy, I would like
      to see a typed or sorted version of logic.  If you want an untyped
      system, you could use the undefined type at the top of the hierarchy
      for every variable.

This is innocuous enough as put, and resembles one of the examples in my
AI Mag paper.
  
   3. One of my complaints about some of the proposed standards is with
      what they include more than with what they leave out.

I agree completely -- this is also in keeping with the views presented
in my paper.
  
   4. Since we are talking about standards for describing systems,
      we need some metalanguage for talking about systems.

Here, I think we actually differ -- I can't see that we *do* need such
a metalanguage at this point (and it is clearly missing from my proposal
in the AI Mag).  The reason is that I do not believe that machines will
be able to make decisions at the metalevel any time soon -- if I axiomatize
my language and you axiomatize yours, there is simply no way that a machine
is going to understand the connection.

As an example, if you use circumscription, eventually saying somehow
"minimize ab" and I use defaults, saying somewhere "by default, not
ab", I can't see a machine making the connection -- although I can see
a person doing so.  My complaint with a second-level axiomatization
is that it *might* be insufficient to describe some particular
representation, that it *will* cause researchers to devote substantial
effort to axiomatizing their representations using this language, and
that there really isn't any obvious payoff at this point.
  
   5. I very much like McCarthy's idea of treating contexts as first
      class objects.

I like it, too.  I think we agree that we wish we *could* make
effective use of the description I argued against in the pervious
paragraph, but we simply aren't there yet.  Your arguments for FOL at
the base level don't apply at the metalevel, since there aren't really
any effective metalevel systems out there.  Let's wait till we have
some answers to this sort of thing before looking for a common
language in which such answers can be expressed!

						Matt Ginsberg