On 1-aug-08, at 14:15, Peter Drake wrote:
On Aug 1, 2008, at 8:08 AM, Mark Boon wrote:
The neighbours of the last move come in the picture because
usually it's only the last stone played that can be escaping a
ladder and it's the neighbours of the last move that could have
been put into atari. Nothing to do with the additional
complexities Don mentioned.
Let me give a specific example. Suppose that, during a playout, the
tree leads us to this position, with O to play:
.........
...OO....
..O##a...
...Ob....
....c....
.........
.........
.........
.........
Having reached the frontier of the tree, we now finish the game
using Monte Carlo with a ladder reader. The last stone played, to
the left of a, is trapped in a ladder, but can escape if not
chased. Our ladder reader therefore suggests O play at a.
For the next move in the playout, if # only reads ladders from the
last move played, it will see that the O stone at a is not in a
ladder, so move is suggested. The playout now turns completely
random, and it's a coin toss as to whether the group will escape.
That's it. That's all there's to it. (Assuming you meant to write "so
no move is suggested".)
If we also search stones next to the last stone played, things only
get slightly better. # sees that its stones are in a ladder from
which they cannot escape, so it doesn't suggest b.
Exactly. This is the neighbour part. In case 'a' was played by
accident and # can escape it will suggest 'b'.
If we tell it to play a ladder breaker in such situations, it might
play c, which is fine. However, on O's next turn, c is not in a
ladder, nor is any stone next to c, so no move is suggested.
Specifically, O does not make the vital capture at b.
I don't bother with ladder-breakers.
It seems too expensive to search every point on the board for
ladders. What to do?
What you could do is keep track of stones with one or two liberties
and keep them in a list. But I found it's still too expensive for
relatively little gain in strength, although I'd have to admit not
yet having exhausted all possibilities here.
What was the question again? ;-)
Mark
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
