From newshub.ccs.yorku.ca!ists!helios.physics.utoronto.ca!news-server.csri.toronto.edu!rpi!zaphod.mps.ohio-state.edu!sol.ctr.columbia.edu!bronze!chalmers Tue Jan 21 09:27:09 EST 1992
Article 2887 of comp.ai.philosophy:
Newsgroups: comp.ai.philosophy
Path: newshub.ccs.yorku.ca!ists!helios.physics.utoronto.ca!news-server.csri.toronto.edu!rpi!zaphod.mps.ohio-state.edu!sol.ctr.columbia.edu!bronze!chalmers
>From: chalmers@bronze.ucs.indiana.edu (David Chalmers)
Subject: Re: Combinatorial explosion
Message-ID: <1992Jan18.235207.29414@bronze.ucs.indiana.edu>
Organization: Indiana University
References: <1992Jan17.082347.15868@bronze.ucs.indiana.edu> <1992Jan17.145033.2435@arizona.edu>
Date: Sat, 18 Jan 92 23:52:07 GMT
Lines: 53

In article <1992Jan17.145033.2435@arizona.edu> bill@NSMA.AriZonA.EdU (Bill Skaggs) writes:

>  It's a breeze.  You just recapitulate evolution.  Set up a
>simulation of 10^80 electrons, protons, and neutrons, in
>Einstein's space time, with plenty of free energy, and run
>it for 10^10 years.  This takes an infinitesimal fraction
>of a second, so you can try 10^100 different parameter settings.
>This still takes an infinitesimal fraction of a second.  Wait
>until you get a simulation that figures out that its a simulation
>and rearranges the galaxies into a message for you.

A nice reply, and somewhat plausible, though it has a few problems.

(1) We don't understand the laws of physics well enough to do this.
(2) If we've done the simulation well enough, they won't know they're
    in a simulation.
(3) There will be huge identification problems -- we may have trouble
    seeing past the protons and electrons to the intelligent beings.
(4) It's very likely that the beings won't share any conceptual
    framework with us.

Still, there's obviously something to this.  Maybe I'll restrict the
question to beings that deal with the same sorts of concepts that
we deal with -- e.g. they have to be able to get around in our
environment, or communicate directly with us, or at least perceive
our environment and come to useful conclusions.

If we rule evolution of entire worlds out of court, there's always
still the evolution of individuals -- e.g. we can search the space
of all programs in a jiffy (halting problem isn't a problem, under
the assumption that anything that's still grinding away after
10^10^10 steps without doing anything useful is unlikely to be
much of a candidate).  But what "fitness" criterion will we use?
Of course we can simply look at them and judge ourselves, but that's
putting humans into the loop and we lose most of the speed advantages.
What kinds of criterion could be automated?  I don't see anyway to
automate the recognition of a good perceptual system, for instance.

I'm still very interested to see suggestions about how one might
be able to develop good systems for perception, reasoning, learning,
etc, in the absence of speed constraints (with highly indirect methods
such as world-evolution ruled out).  At least a catalogue of some
things that would be easier: finite game-playing trivially; mathematical
theorem-proving (though not, perhaps, the development of useful
mathematical concepts); perhaps the development of really good
supervised learning algorithms (just give it lots of data sets, and
pick the smallest algorithm that gives good generalization performance);
what else?

-- 
Dave Chalmers                            (dave@cogsci.indiana.edu)      
Center for Research on Concepts and Cognition, Indiana University.
"It is not the least charm of a theory that it is refutable."


