[extropy-chat] Bluff and the Darwin award

[Russell Wallace]

On 5/18/06, Eliezer S. Yudkowsky <sentience at pobox.com> wrote:
>
> Russell Wallace wrote:
> >
> > At the end of the day though, your claim boils down to the idea that you
> > can become a grand master at chess _without ever having played a game of
> > chess in your life_;
>
> This sounds correct.  Certainly, without ever having played a game of
> chess against an external environmental player.  Possibly, without ever
> internally observing a specific chess game played by two algorithms
> against each other.  The latter option strikes me as silly in practice,
> that is, a suboptimal use of computing power, but doable if some
> superintelligence wanted to do it.


The latter option strikes me as likely to be impossible in principle: it's
been proven that complex nonlinear systems in general don't have analytical
solutions. I don't remember off the top of my head whether that's been
proven for chess in particular, though I vaguely recall someone proving Go
is EXPTIME-hard. Certainly the lack of an analytical solution has been
proven for real physics in even as simple a case as the Newtonian 3-body
problem.

The former, you're on firmer ground; one could avoid the need for an
external opponent by effectively simulating a population of players. "Silly
in practice" would be a good term here - I don't think you could actually
beat a good human player first try doing that - but it's not provably
impossible.

The reason for that, however, is...

All the *useful* information about how to play chess against a generic
> opponent, derivable from any chess game ever witnessed and any
> experience of playing chess, is implicit in the rules of chess.


...the rules, the initial conditions and the fact that fully simulating a
chessboard is computationally tractable. (Well, also switch "useful" to
"necessary"; and "generic opponent" is a nontrivial caveat.)

But in real life, the complete rules and initial conditions aren't
available, and full simulation is not computationally tractable, not even
with a galaxy-sized nanocomputer.


> By the time a superintelligence is told the rules, before the game
> starts, in principle it knows everything it needs to know to beat any
> human player.


The word "superintelligence" just fogs the issue here - a screenful of code
suffices to play optimal chess given infinite computing power. The problem
is that infinite computing power is not in fact available.

More importantly, in chess there is a tractable algorithm that will tell you
how a given move will change the state of the board. In the game of real
life, there is no such algorithm.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20060518/eb594117/attachment.html>