Quoting Darren Cook <[EMAIL PROTECTED]>:
Valkyria uses to methods to bias playouts towards better moves.
Thanks for the reply Magnus. You said it will always try to react to the
last move, and only if no reaction needed will it choose a random move.
It sounds like that is something you only want late in the game, so is
that what you are doing? Do you just use move number as an estimate of
game stage, or do you perform some analysis to see when the board is
looking settled?
No it does this uniformely. Assume white has captured some stones. Any attempt
to save the stones by black (well unitentional because most moves are random)
should be stopped. This means that in all playouts the stones are indeed
captured.
An important design issue is also that if you look at my description
you should
realize that my program computes as little as possible. It only looks
myopically on the last moves and then possibly examines the local surrounding
of a potential random moves. In my playout algorithms that scans the board is
forbidden for efficency reasons.
The non-random moves in heavy playouts are not always correct of course, but
then it is the task of UCT to sort this out near the root.
The ideal I am striving for is that the heavy playout play "non-randomly" as
much as possible as long as those moves are guaranteed to be better than a
random move.
I think Mogo does this even more than Valkyria, if I understood their
paper and
discussions here on the list the Mogo playouts often have long sequences of
forced moves.
-Magnus
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
