Rémi Coulom wrote:
In Crazy Stone (maybe that is the case of MoGo, too), nakade is such a
big problem because the program avoids playing self-atari in playouts.
Crazy Stone will play the self-ataris anyway, but with a low
probability, so they are played at the end of the playout only. In case
there is a semeai around the nakade, this behaviour generates awful
evaluation errors.
Don't you miscount all sekis as well? At the end of the game the
self-atari is the last move left, you play it, and you lose.
The self-atari/seki/nakade isssue is for me one of the biggest pains
with playouts.
If you do not allow self-atari, sekis are OK, but you do not understand
life & death in many simple situations.
If you allow self-atari, you misjudge sekis.
Misjudiging sekis seems to be preferable in terms of playing strength. I
think one is alternately miscounting the seki for both sides in the
playouts, so the effect is not so bad.
The biggest problem is with scoring games. Humans have problems if you
cannot understand simple L&D situations or claim their stones in seki
are dead.
The "solutions" I could come up with so far are:
a) Don't allow self-ataris until the game is passed out and you are
lost. This can handle seki and nakade. The problem is that you are still
miscounting, even though it shouldn't change the winner. Unfortunately,
it seems to make the program weaker.
b) Do not play self-ataris for "big" groups. Maybe "big" should be
determined experimentally. I would guess at 4. I think I remember Remi
saying that Aya does this and it caused a loss at UEC Cup? Haven't tried
it myself.
I would appreciate feedback from anyone who has struggled with the same
problem. Particularly if you solved it :)
--
GCP
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
