I do this for all board sizes, with the same bias. Testing shows that the bias makes it play weaker on 9x9, but I havent tried to tune 9x9 performance, so I haven't fixed this.
I didnt test this bias independently of dymanic komi, so I dont know if they interact badly. David > -----Original Message----- > From: [email protected] [mailto:computer-go- > [email protected]] On Behalf Of "Ingo Althöfer" > Sent: Saturday, April 30, 2011 12:02 AM > To: [email protected] > Subject: Re: [Computer-go] MC Beginner question > > David wrote: > > I use Many Faces's score estimate from the old evaluation function > > to give a similar small bias in the UCT tree, with about the same > > result. > > Martin wrote: > > > In Fuego, we give a small bonus for the size of the win/loss. So a > win > > > is worth between 0.98 and 1.00 depending on the size of the win, > and a > > > loss is worth 0.00 - 0.02. This gives a modest but measurable > > > improvement in strength. And it also makes the moves look better, > since > > > otherwise the program has little sense of what good moves are when > it > > > is far ahead or far behind, and most of the stats look the same. > > > Thanks for the information. Some more questions: > > * Do you do this for 19x19 only, or also for 9x9 and 13x13? > * If so, do you have other ranges (instead of 0.98 - 1.00 ...) > for the smaller board sizes? > > * Did you observe, if these modifications in evaluation interfer with > dynamic/linear komi in some way? > > Cheers, Ingo > > > > > > > -- > NEU: FreePhone - kostenlos mobil telefonieren und surfen! > > Jetzt informieren: http://www.gmx.net/de/go/freephone > _______________________________________________ > 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
