[computer-go] An attempt to explain UCT for computer Go

[Don Dailey]

On Sun, 2006-09-24 at 16:25 -0400, Don Dailey wrote:
> On Thu, 2006-09-21 at 14:48 -0400, Don Dailey wrote:
> > I'm trying something a bit different than UCT as you describe which
> > gave
> > me better results in self play tests.   I achieved about 59% results
> > with my method over about 200 games.   I'm redoing the test to get a
> > bigger sample and to incorporate some other improvements. 
> 
> 
> My method is still working better than UCT, after 122 samples,  I'll
> keep running to get a lot more samples.   I have seen results like this
> turn around because 100 samples is way too few.
> 
> Of course all of this is subject to some tuning i.e. which constants you
> use in either algorithm, and perhaps other implementation details.
> 
> My current result out of 122 games is 58.197 in favor of the alternative
> formula I'm using.

That's  58.197 PERCENT of the games that are won by my method over UCT
as I understand it.   My tester doesn't report the number of games won
and lost, just the percentage and total - so you can extrapolate if you
want that information.




> - Don
> 
> 
> 
> 
> 
> _______________________________________________
> computer-go mailing list
> computer-go at computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/