Re: [gnugo-devel] Re: [computer-go] SlugGo v.s. Many Faces

["Jay Scott" <jay@xxxxxxxxxxxx>]

David G Doshay <ddoshay@xxxxxxxxxxxxxxxxx>:
>It is interesting that 16 moves of once-branched play is worth something
>like 3 stones. We will continue to gather statistics to find out.

This form of search, deep analyses of single lines of play, I have called
"playout analysis" <http://satirist.org/learn-game/inspire/playout.html>,
by analogy to the "rollout analysis" done by backgammon programs.

I think many people find it intuitive that playout analysis should work
in games of imperfect information (where it is performed by repeated
playouts with different sequences of dice rolls, or distributions of
cards in a card game, etc.). And in fact, as far as I know, those are the
games where it's most often used. I think it's also fairly clear that
playout analysis is worth trying when a player can make good decisions on
the basis of short-term information, but cannot see long-term
consequences; that's why my web page above mentions using it in chess
endgames, where programs can play "fortress draw" positions near-
perfectly but fail to understand that they are drawn. ("I can win this! I
know I can! Keep playing, I'll find the way!")

But I'm fascinated by your finding, because it is not obvious. By playing
out strings of relatively poor-quality decisions, somehow enough
information is being transmitted back from the 16-move future, through 16
noisy links, to greatly improve the initial decision. Under what
circumstances is that possible? Or, more generally, what is the best
shape for a search tree, and when, and why?

Maybe you should collaborate with somebody who's interested in the math
of decision-theoretic search trees.

  Jay

_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/