On 17-nov-08, at 15:33, Don Dailey wrote:
On Mon, 2008-11-17 at 13:17 -0200, Mark Boon wrote:
1- Capture a stone in atari with a certain probability (like David
Fotland says he's doing).
2- Forbid playing on the 1st or 2nd line unless there's a stone
within manhatten-distance 2.
3- Forbid putting yourself into atari with a large number (>6) of
stones.
Hi Mark,
I'm still working on improving the refbot and I'm using the Michael
Williams modification too.
Many of the things I've tried don't directly affect the playouts. One
of them you can try is this:
1. Do not play to any edge point unless a stone (of either
color) is
diagonally or orthogonally adjacent.
I think that is a good practical rule. It's not a perfect rule,
in the
sense that such a move could still be best - but I think for a playout
strategy it is good.
Yes, I have something like this in my UCT search as well. But there
are a few cases where it overlooks the only move to make a group
alive so I didn't mention it until I knew a bit more of the
ramifications.
For a practical player you should veto that move as a root move choice
too, because the rule will severely reduce the number of samples
for the
edge points and bias them towards wanting to be played, the
opposite of
what you want.
In fact, that has to be considered for anything you do. You must
distinguish between using a rule to adjust the playout strategy and
actually using the scores of the moves for this or that. Some UCT
implementation try to use AMAF data to shape the tree and you could
get
into trouble if you are not careful.
On another note, as an experiment I have a bot running on CGOS that
is the ref-bot but instead of using a fixed number of simulations I
use a fixed amount of time that slowly diminishes towards the end of
the game. The result is it does about 200K simulations per move for
most of the game on a single processor. Its rating is currently stuck
at 1367 or so. With 2K simulations the rating tends to be 1280
without using the weighted formula. This one uses 100 times as many
simulations and the weighted formula, so I had expected it to rate
higher than that. Is this normal? Does the MC-AMAF combination just
not scale at all? Or could it be because there currently doesn't seem
to be a very large population of bots playing?
Mark
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
