Re: [pygame] PyGameDB coming along well

[kschnee <kschnee@xxxxxxxxxx>]

On Tue, 19 Aug 2008 07:49:16 -0700, James Paige <Bob@xxxxxxxxxxxxxxxxxxx>
wrote:
> That reminds me of something I read recently. Apparently the first ever
> computer program that can play Go with near-human skill was developed
> this year:
> http://www.usgo.org/index.php?%23_id=4602
> 
> Of course, it only beat the human in 1 out of 3 matches :)


From what I understand, the reason that AI for these games is so hard is
that the general method being used is a brute-force lookup table. A chess
computer has a list of each possible board position that could result from
its next move, and the possible moves from there, eventually dead-ending at
some kind of evaluation of whether a board position is "good." So the AI
picks the move that its lookup table says is most likely to give a good
result. Because there are 20 possible moves for White at the beginning and
400 possible board configurations after Black goes once, that's a
ridiculously huge database for looking even a few moves ahead. Go has even
more possible moves, making it harder to take on _with that strategy_. In
other words the trouble in building a Go computer isn't so much because Go
is innately more complicated than Chess, but because of the specific way
that it's complicated.

What's been proposed as an alternative to this strategy is to try to get a
machine to recognize patterns of pieces, the way that a human might
recognize a "knight fork" situation, so that it can develop general
strategies instead of relying on exact known board layouts. If anyone can
build a program like that, it'd be useful for other purposes too.