Path: utzoo!utgpu!jarvis.csri.toronto.edu!cs.utexas.edu!sun-barr!apple!mips!zaphod.mps.ohio-state.edu!think!paperboy!snorkelwacker!bloom-beacon!eru!luth!sunic!tut!santra!mcsun!ukc!edcastle!aipna!cam
From: cam@aipna.ed.ac.uk (Chris Malcolm)
Newsgroups: comp.ai
Subject: Re: Cog Sci Fi (was: STRONG AND WEAK AI)
Message-ID: <1781@aipna.ed.ac.uk>
Date: 30 Dec 89 20:05:07 GMT
Reply-To: cam@aipna.ed.ac.uk (Chris Malcolm)
Organization: Dept of AI, Edinburgh University, UK.
Lines: 77

Stevan Harnad wrote:

> cam@aipna.ed.ac.uk (Chris Malcolm) wrote:

>> my game is assembly robotics... The assembly agent is designed not only
>> to succeed in its tasks, but to present a suitable virtual world to the
>> planner, there is an extra constraint on the task modularisation. That
>> constraint is sometimes referred to as the symbol grounding problem.

> As the mint that coined the term, I think I speak with a little
> semantic authority when I say that that's decidedly *not* the symbol
> grounding problem but rather a symptom of it! Virtual worlds are not
> real worlds, and what goes on inside a simulation is just meaningless
> symbol crunching. The only way to ground symbols is in the real world.

I think Stevan has taken me to mean the opposite of what I intended! When
I said "virtual world" I meant something which bore the same
relationship to the real world as does a virtual machine (such as a
software code intepreter) to a real machine (such as the computer
running the interpreter. In other words, this kind of "virtual world" is
just as real as the "real world", the "virtual" prefix meaning simply
that we are referring to a level within the organisation of a creature
which is hosted by the various physical and information processing
mechanisms which underly it. At the top level the virtual world of a
creature is "what it is like to be" that creature, the umwelten of von
Uexkull. A virtual world in this sense is most emphatically not any kind
of simulation.

If, however, one presumes that the symbol grounding problem can be
solved by the kind of perceptual mechanisms Stevan has outlined in his
various papers, with the occasional qualification that a similar system
can be devised for the motor side, then it is true that the kind of
virtual world I am talking about becomes rather like a simulated world,
such as in the often cited example of the aircraft simulator being
hooked up to the controls and sensors of a real aircraft, and thus
crossing the great divide between simulation of flight and really
flying. The reason for this similarity is the separation of the sensor
and motor hierarchies until they are combined at a symbolic level. This
separation omits the powerful facilities available to a creature by
combining sensing and motor activity to create kinds of sensing and
action otherwise impossible (or much more computationally expensive).
One example of this kind of facility is the servo-mechanism, which uses
feedback to create - as far as the functioning creature is concerned - a
useful and perceptible stability from the otherwise unstable and
ephemeral. Another example is the use of an unstable motor activity
which the environmental conditions will tend to drive into one of two
appropriate limiting conditions. In this case the unstable motor
activity is used as a combination of sensing and response system.  There
are many other kinds of ways in which sensing and action can be combined
to advantage, or subsituted for one another.

Wherever sensing and action have been locally amalgamated in this sort of
way a barrier is created to the extension of separate sensor and motor
processing hierarchies. While these processing hierarchies definitely
exist, and are most important, they can't be extended beyond such a
barrier. An independent set of such barriers can be seen as constituting
a level. There can be many such levels erected atop one another. When I
say "virtual world" in the context of a biological or artificial
creature I refer to such a level. These virtual worlds depend on the
active functioning of the real creature situated in the real world, and
in that sense they are thoroughly real and grounded.

Let me add the prediction that whereas one certainly can ground toy
robots with architectures which involve separate perceptual and motor
hierarchies erected on top of a basic bottom level of feedback control,
that we will not be able to create "intelligent" but dumb (speechless)
robots, let alone a robot which could pass the Total Turing Test,
without multiple levels of such virtual worlds, each one creating new
virtual sensors and effectors, in terms of which the next level can be
constructed. In between these levels, I expect the kind of categorical
perception described by Stevan Harnad, and its motor analogue, will play
an important role. But, IMHO, it cannot, on its own, accomplish all that
is required of symbol grounding in a complex intelligent creature.
-- 
Chris Malcolm    cam@uk.ac.ed.aipna   031 667 1011 x2550
Department of Artificial Intelligence, Edinburgh University
5 Forrest Hill, Edinburgh, EH1 2QL, UK