From newshub.ccs.yorku.ca!torn!cs.utexas.edu!sdd.hp.com!zaphod.mps.ohio-state.edu!sol.ctr.columbia.edu!spool.mu.edu!darwin.sura.net!cs.ucf.edu!news Tue Nov 24 10:52:36 EST 1992
Article 7693 of comp.ai.philosophy:
Newsgroups: comp.ai.philosophy
Path: newshub.ccs.yorku.ca!torn!cs.utexas.edu!sdd.hp.com!zaphod.mps.ohio-state.edu!sol.ctr.columbia.edu!spool.mu.edu!darwin.sura.net!cs.ucf.edu!news
>From: clarke@acme.ucf.edu (Thomas Clarke)
Subject: Re: It is AI when...
Message-ID: <1992Nov19.141459.4313@cs.ucf.edu>
Sender: news@cs.ucf.edu (News system)
Organization: University of Central Florida
References: <BxpIsD.H1E@cs.bham.ac.uk>
Date: Thu, 19 Nov 1992 14:14:59 GMT
Lines: 84

In article <BxpIsD.H1E@cs.bham.ac.uk> axs@cs.bham.ac.uk (Aaron Sloman) writes:
> clarke@acme.ucf.edu (Thomas Clarke) writes:
> 
> > Date: 12 Nov 92 20:22:05 GMT
> > Organization: University of Central Florida
>     ....
> > If you can discern how it works it's not truly intelligent behavior.
> > You'll know its truly intelligent when you can't figure out how it
> > works.
> 
> 
> This implies that
> EITHER
> 
>     (a) human beings, chimpanzees, squirrels (and other things) are
>     not truly intelligent
> 
> OR
>     (b) we'll never understand how they work
> 
> I see no reason to believe either, though understanding how they
> work is very difficult and is likely to take many more years. I see
> current AI/Cognitive science as being at a stage that could be
> compared with Galileo's understanding of physics: i.e. some
> important new ideas have emerged, but there`s still a very long way
> to go.
 
Let me clarify my definition of what I mean by understanding.  
We already understand in a general way how intelligence works in
humans et al:  it involves a massively parallel network of neurons.

I have in mind a more precise definition of understanding:
(I call this strong understanding, but then someone has
probably already used this term in some other sense.  Not being
a professional philosopher, I'm not sure.)
 
  A process is strongly understood when it is in principle possible 
  to manipulate the process so as to control the process over a finite 
  time span using control inputs of smaller Kolomogorov-Chaitin 
  complexity than the process behavior.

I insert "in principle" since while something may be understood, its
control may be beyond engineering practice.  Control of the brain
may require 10^N (a larger number) of fine platinum electrodes to be
embedded in the appropriate neurons - possible in principle. but
certainly not in current practice.

The stuff about "Kolomogorov-Chaitin" complexity is there to eliminate
control some sort of real-time statistical control wherein the 
inputs are adjusted continuosly in some sort of tight feedback loop.
No fair to make the person go to the refrigerator and eat fudge by
controlling all the individual afferent neurons involved in the
activity.  You have to know where the "higher level centers" are and
the "code" that they use so you can input "hunger, sweet, cold, get"
or some equivalent and have this brief message generate the
complex behavior.

"Smaller" means of smaller exponential order, or one of various
similar concepts appropriate for the process.

The "finite time span" is there to take care of chaotic dynamics.
I think one can say that weather is can be strongly understood.
However, chaos prevents control of the weather beyond a horizon of 
several days; no matter how complete the thermodynamic transducers 
available to the meteorologor, without real-time feedback the weather 
will diverge from the desired outcome in a few days.  

"Finite" should be taken in the mathematicians since of having a lower 
bound.  I should be more precise about what "finite" means, but I'm 
trying to keep my definition "finite":-)

I don't think my earlier statement, which now becomes "intelligence
will never be strongly understood", applies only to AIs that are
continuous or neural.  An algorithmic (Turing machine) AI could fail to
be strongly understood. Its code could be generated by some sort
of genetic process that fails to supply documentation so that the
problem of strongly understanding the code in order to control the 
AI's behavior is NP-complete or, worse yet, undecidable.

--
Thomas Clarke
Institute for Simulation and Training, University of Central FL
12424 Research Parkway, Suite 300, Orlando, FL 32826
(407)658-5030, FAX: (407)658-5059, clarke@acme.ucf.edu


