Newsgroups: comp.ai.neural-nets,comp.ai
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!howland.reston.ans.net!swrinde!gatech!news.sprintlink.net!noc.netcom.net!netcom.com!olea
From: olea@netcom.com (Michael Olea)
Subject: Re: Neural nets: the argument from evolution
Message-ID: <oleaD90My9.9EI@netcom.com>
Organization: NETCOM On-line Communication Services (408 261-4700 guest)
References: <3po0n9$kqp@Mercury.mcs.com> <3po4sn$hfk@canopus.cc.umanitoba.ca> <3prl6k$kdh@info.curtin.edu.au>
Date: Tue, 23 May 1995 05:22:57 GMT
Lines: 69
Sender: olea@netcom8.netcom.com
Xref: glinda.oz.cs.cmu.edu comp.ai.neural-nets:24240 comp.ai:30040

buihh@cs.curtin.edu.au (Hai Hung Bui) writes:

>Christopher McKinstry <cmckin@mbnet.mb.ca> writes:

>>the reason intelligence evolved slowly is because it didn't make a 
>>difference for a very long time... only in the last few hundred thousand 
>>years has having human level intelligence made any difference in the 
>>survival of any life form.

>>genetics find solutions to a problem VERY quickly... the problem is "the 
>>problem"... you must have a fitness test that encourages intelligence, if 
>>the test is valid, the intelligence will emerge quickly.

>>for example in karl simms evolution work... it took only 30 or so 
>>generations to evolve creatures that swam quite well when the test 
>>rewarded movement.

>>if we had a program that did well on a turing test... but did not think, 
>>it could be used to train an evolving neural net. eventually the nueral 
>>net will respond like the turing simulator, except it would be much more 
>>general and could then continue to learn from real people.

>>it's all in the fitness test.

>How one would go about and develop such tests? Does it necessarily involve
>"growing" agents in some sort of environments? If so, will artificial
>enviroments do, or will we enventually have to evolve our agents
>in real natural environment?

>The problem with artificial environments is it wont be as rich as real
>life. Thus, the fitness test might overlook some aspects of intelligence.

>The problem with real natural environment is we cannot speed up the
>evolution process (as we can do in a simmulation). This could be a problem
>too since we dont want to spend thousands of years in evolving our agents.


>-hung
>buihh@cs.curtin.edu.au
>http;//www.cs.curtin.edu.au/~buihh


>>-- 
>>K. Christopher McKinstry               http://www.mbnet.mb.ca/~cmckin/
>>======================================================================
>>-The Truth is of course, is the Universe is playing God with dice...
>>-Plutonium president kill for sale NSA CIA FBI hostage mole infiltrate
>> (above line designed to drive NSA's internet scanning computers crazy)

	The distinction between artificial and natural environments is
itself a bit artificial - evolution is in fact occurring in
micro-environments that might seem artifical, but become natural with
surprising speed, driven by "the market".  Someone mentioned the "ABL"
problem - given the image of a piece of mail locate the address block.
In the abstract it's an artificial, even a toy, problem, but an effective
solution brings with it a measure of economic survival.  I know, I worked
on it, had a contract with the company that came in second in this little
struggle.  The irony is that the company that won the USPS open competition
had, by USPS' own measure, slightly inferior recognition - Recognition
is but a piece of a production mail-processing system; the winner of this
little struggle won by purchasing the manufacturer of key hardware, getting
a sweet deal on the hardware, and bidding the system "at cost".  The
superior "AI" lost to superior positioning of hardware distribution.  No
matter, it is still in broad perspective survival of the most fit.  It is
in every way an evolutionary struggle, though the means of adaptation are
superficialy remote from mutations and crossover and the like.



