I have been interested in monte-carlo approaches to Go since running
my first MC simulations in magnetic phase transitions when I was in
graduate school in the 1980's. What held me back, even when the latest
crop of MC programs started winning against older stronger programs
and my program SlugGo, is that in the physics simulations we know on
theoretical grounds what the shape of the random distributions are, but
in Go we do not. I was amazed at how well UCT helps get around the
problem and still allows the use of nearly flat random distributions
(flat except for a few hand tuned rules).
Recent work on the ELO ratings of patterns comes very close to what I
think is needed to move forward, and we are also working on ways to
determine biased probability distributions that are appropriate for Go
move selection.
I have little doubt that such "weighted playout" (not really heavy or
light)
considerations will lead to the next major step in computer Go progress.
The advantage of such a method is that it intrinsically matches the
basic
premise of MC: the right degree of randomness allows you to search the
problem space appropriately.
Cheers,
David
On 8, Jan 2008, at 11:04 AM, Don Dailey wrote:
I think Dave Hillis coined this term "heavy playouts."
In the first programs the play-outs were uniformly random. Any move
would get played with equal likelihood with the exception of eye-
filling
moves which don't get played at all of course.
But it was found that the program improves if the play-outs are
somewhat
"managed". So Dave starting calling the original formula "light
play-outs" and the slower but better method "heavy play-outs."
And yes, it slows down the play-outs. Still, the play-outs seem to
require a good bit of randomness - certainly they cannot be
deterministic and it seems difficult to find the general principles
that
are important to the play-out policy.
- Don
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
