On Fri, 2006-12-01 at 08:39 -0800, David Fotland wrote:
What's included in an evaluation? Is each evaluation one random game, or a
set of random games that gives good enough statistics about the value of a
position?
When you say random it conjures up images of aimless wandering - but
the monte carlo programs are not really random in that sense - they are
directed. The statistical terminology is that that are not uniformly
random as with a good random number generator for instance.
But they do start off that way - as they gain information they apply
that information to bias the randomness towards good moves in a
purposeful directed way. And the process continues to feed back on
itself.The method is powerful enough that it produces perfect play
eventually.
A full width global search also produces perfect play but it's not
concerned with the quality of the moves it searches - it searches
everything no matter how awful it is (subject of course to alpha/beta
cutoffs.)
There is no reason whatsoever to avoid other techniques to improve the
process. The judicious application of patterns could focus the search
even more.Any time you can insert knowledge effectively it should be
done because as you yourself have hinted at, a little knowledge can save
a lot of searching and it would be stupid not to take advantage of that
fact. But it's equally stupid to ignore search completely.
- Don
David
On a P4 3.0Ghz mono processor, the number of evaluations per
seconds is in the
order of 4500/s in 9x9, 2500 in 13x13 and 1100 in 19x19.
Sylvain
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
