On Thu, 2007-01-25 at 08:23 -0800, terry mcintyre wrote:
> Go, being a matter of efficiency over one's opponent, may be even more
> susceptible to improvement via many small improvements over many moves
> than is chess. As long as you don't leave weak shapes behind, picking
> up a point here, a point there at a slightly faster rate than your
> opponent will give you the game.
And this indeed seems to be the case. When I was testing 19x19 cgos
with Steve, the performance rating between versions that did 2x
simple MC play-outs was enormous, over 300 points per doubling.
That rate of improvement tapered off with higher levels because the
simple MC algorithm is strictly limited in scalability.
I also had a difficult time producing a player that was less than
200 ELO stronger than a random player. Even a single play-out,
which seems hardly enough to discriminate between moves, is
enormously stronger than a random player. It was pretty much
like this:
ASSUME computer is black
1. play 1 random game.
2. If black wins, play one of the first N black moves in the
play-out (all-as-first, for me it's some-as-first.)
3. If white wins, play one of the black move NOT in the play-out.
4. Crush a random player!
Of course 2 play-outs was incredibly effective against against a
1 play-out player, 4 beats 2, 8 beats 4 and the tapering effect
is very gradual up to the level we were able to test - where the
computer was not able to play a game without losing on time.
This was despite the fact that the algorithm is limited. There
is a point in the 9x9 version where it hits the wall. For
Botnoid (AnchorMan) 5000 simulations is almost as good as it
gets and it plays very quickly at this level. This is not
a refutation of the principle, it's just that AnchorMan has
a very un-scalable algorithm. 1 ply monte carlo cannot
discover much of anything even if given infinite number of
play-outs but this changes completely when a tree is added
such as in UCT.
- Don
>
>
>
>
> ______________________________________________________________________
> Get your own web address.
> Have a HUGE year through Yahoo! Small Business.
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
