I have c=0.7, and I expand after the 1st visit. My bots occasionally crash on the server I'm running things on, I've no idea why. I will try lowering c to 0.5 and raising the required visits to 50 and compare. - James
On Fri, Oct 22, 2010 at 11:57 AM, Christoph Birk < [email protected]> wrote: > Hi James, > > > On Wed, 20 Oct 2010, James Pettit wrote: > >> I just modified my bot to do just playouts, no tree, move selection is >> most >> wins. It's on CGOS now as GoGo-10k. I'll leave it there for a while. It >> definitely exhibited some of the same problems as the UCT version, but I >> cannot for the life of me find a bug in my playout implementation. >> > > After about 100 games (your bots stopped playing a few hours ago) > the "random-playout" results indicate that our basic implementation > is similar. I don't think you have any (more) bugs in that part > of the code than I do. > > But the UCT part is significantly (bayes-elo) worse than mine. > What parameters are you using? In the UCT formula > uct_score = c * sqrt(log(nvisits/pvisits)) > I use c=0.5 and I expand nodes after 50 visits. > > Please note that my UCT-bot does _not_ use RAVE. > > > Christoph > _______________________________________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go >
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
