It is possible that Your uniform playout part is a lot more efficient than UCT part, because of costly move choosing procedure (loop).
On last Computer and Games conference I and Jakub Pawlewicz published an article describing 1+epsilon trick that increases efficiency of proof-number search. It also applicable to UCT. I believe that UCT with epsilon trick may be as efficient as the rest of the playout. I will put the article on my homepage in few days. Best Regards Lukasz On 1/10/07, Don Dailey <[EMAIL PROTECTED]> wrote:
> Just add a new child only when parent is visited more times than X. I'm testing this very idea with UCT. I'm testing 3 versions right now where X is different in each version. I'm testing 5, 10 and 100. I've always used a higher value than 1 but never thoroughly checked this out. In my tests, X=5 is winning but there are not enough games to be statistically convincing. I'm running at long time controls so that the answer I get is representative. The interesting thing is that it can do a lot more play-outs when when X is high, although it is less strong. I need to understand why. Based on the paltry data I have now it's a mistake to use X that is very high. Rating Win perc Tot Gms Ave Time Player ------- -------- ------- -------- ------ 2105.4 58.065 31 1035.9 laz005 2055.1 50.000 44 881.9 laz010 2000.0 41.379 29 633.4 laz100 Black wins: 29 55.8 % White wins: 23 44.2 % On Fri, 2007-01-05 at 12:05 +0100, Łukasz Lew wrote: > Just add a new child only when parent is visited more times than X. > > You will loose only a minimal amount of information. > > Hope this helps :) > Lukasz > _______________________________________________ > 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/
_______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
