Xref: utzoo comp.ai:7706 comp.ai.philosophy:3
Path: utzoo!attcan!uunet!tut.cis.ohio-state.edu!iris.cis.ohio-state.edu!byland
From: byland@iris.cis.ohio-state.edu (Tom Bylander)
Newsgroups: comp.ai,comp.ai.philosophy
Subject: Re: Emergent properties (was: What AI is exactly)
Message-ID: <84118@tut.cis.ohio-state.edu>
Date: 27 Sep 90 15:49:34 GMT
References: <59556@bbn.BBN.COM> <3894@se-sd.SanDiego.NCR.COM> <26FA3460.1C7D@marob.masa.com> <3918@se-sd.SanDiego.NCR.COM> <15132@venera.isi.edu>
Sender: news@tut.cis.ohio-state.edu
Reply-To: Tom Bylander <byland@cis.ohio-state.edu>
Followup-To: comp.ai
Organization: Ohio State University Computer and Information Science
Lines: 46

In article <15132@venera.isi.edu> smoliar@vaxa.isi.edu (Stephen Smoliar) writes:
>... one cannot point to some specific
>system component and say, "Here is where the knowledge to recognize the letter
>A resides."  An outside observer will be able to note that there are parts of
>that automaton which exhibit similar behavior when confronted with various
>presentations of that letter, but one cannot to a Newell-style knowledge level
>analysis of the system.

As I understand it, a knowledge-level analysis of a system specifies
what knowledge can be ascribed to the system based on its external
behavior.  There is no problem in ascribing "the knowledge to
recognize the letter A" to machines that actually recognize the letter
A, no matter how the machines are constructed.

Talking about the knowledge within components of a system is what
Newell called a "mixed model" (which is elaborated upon by Sticklen's
JETAI article "Problem Solving Architecture at the Knowledge Level").
The structure of the components and their interaction are described at
the symbol level, and the components themselves are described at the
knowledge level.  In this style of analysis, one must be careful.
Knowledge can be ascribed to a component only if the component by
itself *behaves* as if it has the knowledge.  Merely having a
representation of the knowledge is insufficient because, for
appropriate behavior to occur, some other component is needed to use
the representation.

So I would reword your first claim quoted above as follows: "One
cannot point to some specific system component and say, `Here is where
the knowledge to recognize the letter A is *represented*'".  I don't
why this kind of situation is so surprising because it is true for
most ordinary programs.  For example, a sorting program "knows" that
less-than is transitive, but, for typical sorting algorithms, it is
not possible to point out a "component" that represents this
knowledge.  (Knowing transitivity would appear to be an "emergent"
property of sorting algorithms.)

Finally, I don't think you want to claim that "one cannot do a
Newell-style knowledge level analysis" of such systems or their
components.  Certainly, it might be difficult to determine how the
component interaction results in recognizing the letter A, but this is
a problem of symbol-level analysis.  With regard to ascribing
knowledge to components, I understand why it is difficult (because it
requires considerable experimentation and analysis), but not why it is
impossible.

						Tom Bylander