Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!csd4.milw.wisc.edu!lll-winken!uunet!portal!cup.portal.com!dan-hankins
From: dan-hankins@cup.portal.com (Daniel B Hankins)
Newsgroups: comp.ai
Subject: Re: Making fires and making minds - the laws of physics prevail
Message-ID: <17468@cup.portal.com>
Date: 22 Apr 89 06:49:15 GMT
References: <10992@bcsaic.UUCP> <16878@cup.portal.com> <2792@crete.cs.glasgow.ac.uk> <17374@cup.portal.com> <2826@crete.cs.glasgow.ac.uk>
Organization: The Portal System (TM)
Lines: 205

In article <2826@crete.cs.glasgow.ac.uk> gilbert@cs.glasgow.ac.uk
(Gilbert Cockton) writes:

>>     Which is to say, no choice at all.  The illusion of choice is one of
>>the most pervasive fictions of Western thought, arising at least in part
>>from our inability to predict with any accuracy our own or other's
>>actions.
>
>I see. And I presume the laws of physics give rise to this illusion. We are
>then, determined to believe in choice?  

     Some of us are.  Others are not.  It's impossible to predict.


>If the laws of physics must prevail over everything, then there must be no
>progress, no history.  All is determined.  

     I don't see where determinism eliminates progress or history.

     History is a record of events that have happened.  Determinism is
irrelevant to that.  Unless, of course, you are talking of _social
theories_ of history.  I don't claim complete knowledge, of course, but
those I am familiar with (Spengler, Marx, Adam Smith, Illuminism) all
ascribe some measure of predictability, and therefore determinism to the
history of societies.
     In theories of history, as in theories of anything else, some will fit
the data and some will not.  Some theories will fit the data but be
ineffective for lack of predictive power.  'Demonic' theories explain
computers quite well, but are ineffective for predicting a computer's
future behavior or for modifying that behavior.  Theories of history based
on the notions of free will, responsibility, and so on may well be
consistent with known history.  But they are poor predictors of future
behavior.

     As for progress, determinism is _not_ the same as fatalism.  Our lack
of knowledge of the future means that we must ever strive to improve it. 
Self-organizing systems such as genetic algorithms and artificial neural
networks are completely deterministic.  Yet they do progress from an
ineffective existence to an effective reality.


>So I presume America's future decisions over abortion, school prayers and
>the budget deficit are all going to be determined by the laws of physics?

     Yes.

     But not knowing what those decisions are going to be, it behooves us
to behave _as if_ we had a choice (free will is a _useful_ illusion, in a
large number of cases - just like Newton's laws of motion).  Fatalism
places the object of Fate outside of the system, in some sense.  In
Fatalism the human is acted on by fate but has no power to act on fate.
Determinism allows the physical system to be _part_ of the much larger
system, and therefore to influence it as it is influenced, just as each of
Lorenz's three equations change the behavior of the other two.

>[discussion of oscillator circuit in part driving itself omitted]
>>     The psychology experiments are an example of feedback of a more
>>complex nature.  By feeding the system under study (a person or group of
>>persons) information about the larger system (the subject-experimenter
>>system) you alter the behavior of the system.
>
>You cannot predict how though.  The subjects range of choice (variation of
>response if you want) is not predictable.  All knowledge here can only be
>post-hoc, and there is no immediate hope of this changing.  I suggest you
>acquaint yourself with research methods in psychology, since the argument
>cannot be carried any further without some common knowledge here.

     I agree.  I cannot predict how informing the human subject of the
purpose of the experiment will affect that human's performance in the
experiment, except within certain _very_ broadly based parameters.

     But the same is true of physical systems.  Something as simple as a
convection roll's behavior cannot be predicted because of the way that each
parameter (speed, top/bottom heat gradient, left/right heat gradient)
alters the other two in some very sensitive ways.

     This is even more true of an artificial neural network program. 
The results of disturbing the system, say with feedback about its own
behavior, cannot be predicted without _complete_ knowledge of the current
state of the net and complete knowledge of the current input.  Even then,
the only way to predict the precise output is to run the program.  If the
complete current state of the net is unavailable, as it might well be on a
highly parallel asynchronous implementation (like a hardware-based neural
network), then its behavior is just as difficult to predict as that of a
biological system (i.e. animal or man) of equal complexity.

     The upshot:  Deterministic systems are in principle just as difficult
to predict as so-called freely willed ones.


>There are already results that show that a digital computer cannot hope to
>apply AI techniques to massive rule-bases in real-time and achieve human
>performance (CACM paper, 1988, example was truth maintainance on a very
>fast computer).  What is possible depends on the theory.  The truth of any
>theory rests ultimately on what can be done with it.

     This clearly sounds like you think the only route AI takes is the
symbolic LISP or Prolog based inferential one, as in expert systems and the
like.  The experiment you describe sounds like it was performed on a
standard Von Neumann architecture, possibly (although not probably, from
your account of the nature of the computations) using vector processors. 
The limitations of the Von Neumann architecture are well known.

     The limitations of massively parallel architectures, such as neural
networks directly implemented in hardware, are not as well known.  It is
known that an electronic neuron is at least as fast as a biological one.
Repeat:  AI techniques are not limited to rule-based inference systems. 
Connectionism (i.e. implementing the important parts of human neural
systems in silicon) is coming on strong.


>>What predicts what the movement of each of these particles will be? 
>>Physical law.  What therefore predicts the movement of the entire arm? 
>>Physical law.
>
>What decides where to put the arm. Me, within my physical limitations. 
>I've just tapped the desk twice. I presume the laws of physics forced me
>to do this here and now.

     Yes.  You could not have done otherwise.  Your only reason for
thinking that you could have done otherwise is that your model of your own
behavior is incomplete.

     You simply _don't know enough_ about the state of your brain and your
body at the time you tapped the desk, even afterwards, to see the causal
connections.

     One might as well say that a cloud has free will.  Any behavior you
can ascribe to free will I can ascribe to action which is not predictable
(incomplete models) and therefore _appears_ to have the possibility of
coming out differently.


>>     Specific knowledge of these areas is not germane to the problem at
>>hand;  meta-knowledge (knowledge about knowledge of these areas) of them
>>is.
>
>This is marvellous example of the self-imposed ignorance of the AI
>type - I don't need to know the details, because I can tell, without
>looking at them, from my meta-knowledge, that I can ignore the
>details.  For this poster at least, the complaint of self-imposed
>ignorance is apt.  Epistemicly, it is as irrational a claim as any
>that comes from the 'religious' corners.

     Would you be happier if I called this meta-knowledge by its more
common name: Epistemology?

     For the purposes of an investigation into the rate of fall of a 1 kg
weight on the moon, I think I can safely ignore the details of polymer
chemistry.


>>     All of these obey the laws of physics.  Perhaps your point is that
>>they obey other laws.  This is true, to some extent.  But these other
>>laws are either approximations of combinations of simpler laws, or are
>>directly reducible to simpler laws.
>
>Not according to systems thinking, if I have understood it correctly.

     Please be more specific.  Holistic theories I have encountered put
holism and reductionism on two sides of the same coin.  Sure, the whole is
greater than the sum of the parts... it's their product :-).  Seriously,
though, it seems that most of the power of holistic reasoning comes out of
the combinatorial explosion and from nonlinearity (chaos, or sensitive
dependence on initial conditions if you like).


>>     AI's sole purpose is not to advance knowledge of human psychology. 
>>It has many purposes.  That's one.  Perhaps it doesn't fulfill that one
>>well.  Another is to come up with potentially useful applications.  That
>>one it fills extraordinarily well.
>
>Examples please.  Folk around here have been asking for them for years. 
>To qualify applications must be in use, have a manual, a user base and be
>applied to real work.

     Genetic algorithms are in use in a variety of commercial environments,
such as flow control on large oil pipelines.

     Expert systems are sold and used in a wide variety of environments
every day.  I remember one in particular as long ago as 1978 assisting in
medical diagnosis at the Mayo clinic.  Expert systems are perhaps AI's
biggest commercial success.

     Robotics.  Robots in a large number of automated and semi-automated
factories are programmed using large symbolic AI systems.

     Databases.  Advances in natural language processing are used in at
least one commercial database system.

     Games.  Computer based games are a multimillion dollar business, and
they are often among the first to employ AI techniques, in order to get
smart opponents.  There are plenty of commercially available chess-playing
programs based on a combination of brute force and AI, combat flight
simulators with smart enemy planes (more on this later), Go playing programs,
Reversi playing programs, role-playing programs employing NL advances, and
so on.

     Military applications.  The military often want to implement smart 
enemies on their simulators, and also they use expert systems for battle
management assistance (not always successfully - note that a poorly written
program does not invalidate the concept, merely the implementers).


Dan Hankins