Hi,
I'm wondering about how to best make my Monte Carlo playouts within
UCT heavier and which pieces of domain knowledge to better use to bias
the tree and which ones to apply during the playouts, so I would like to
ask about previous experiences.
Currently, I do three basic hints that I check the previous move and
its immediate neighbors against when deciding on next move:
(i) If the last move or any of its neighbors is in atari, play the
last liberty (either captures or escapes; do not try to escape if
it does not add any liberties).
(ii) If the last move or any of its neighbors can be ataried, play
one of the atari-ing moves (either makes atari or prevents it; again
there is extra check so that the prevention isn't itself self-atari,
it turns out that a bot that always fills one eye of surrounded
two-eyed group is not very strong).
(iii) If the last move creates cut shape, cut with 50% probability.
(iv) There was also hint to make a random move in direct vincinity of
the last move, but it seemed to just generate a lot of horrible shapes.
My main question is about the rules (i) and (ii) (I'm not sure about
(ii) actually), other bots seem to do that as well - I wonder, do you
_always_ make the move hinted by these rules during the playout, or only
with some probability? For example, the RR MoGo paper says that
"it looks for the moves capturing stones on the Go board, plays one if
there is any" - does it really _always_ capture a group in atari
anywhere on the board before considering a random move?
Also, do you use the (i), (ii) rules in the tree search in any way?
What about the ladder checking - do you do it in the tree search, or
during the playouts?
Thanks a lot,
--
Petr "Pasky" Baudis
Whatever you can do, or dream you can, begin it.
Boldness has genius, power, and magic in it. -- J. W. von Goethe
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
