Ingo is correct. The territory chart's fluctuations are because it uses the old territory evaluation without any lookahead. So for example if a move threatens a big group, and the reply saves it, there will be fluctuation, since the group's territory went from solid to unsettled, back to solid. If I turned on the quiescence search it would be much smoother, but also much slower to draw. If you have the program you can click on a move (which goes to that position), and ask for a score display, and see what is going on.
The early fluctuations in the win rate chart show that MCTS is not very good at fuseki. > -----Original Message----- > From: computer-go-boun...@dvandva.org [mailto:computer-go- > boun...@dvandva.org] On Behalf Of "Ingo Althöfer" > Sent: Sunday, January 15, 2012 6:33 AM > To: computer-go@dvandva.org > Subject: Re: [Computer-go] Game 2 goes to Zen: 1-1 > > Hi Darren, > > I think, I an also answer your questions on the charts. > > > > Winning percents for White, as seen by "Many Faces of Go" > > > http://www.althofer.de/zen-tromp-02-percents.jpg > > > > > > Territory estimates for White, as seen by ManyFaces > > > http://www.althofer.de/zen-tromp-02-territory.jpg > > ... > > David, what is the territory chart saying? Is it based on playouts (i.e. > > based on terminal positions), or is it counting the current board > > position (i.e. as if both players were to pass)? > > No playouts at all involved. "Territory" > is just counting the current board position in the traditional > pre Monte-Carlo way. It is one of the nice old MF-10 features within > ManyFaces, and thanks to David for leaving them in the code. > It is quite typical that the fluctuations in Territory are much larger than > in the relevant part of the Monte-Carlo percent chart. > > > (The territory charts says 15pts to black, with 52-55% to white at > > around move 45; then says 40pts to white, with about 52% to white at > > move 80.) > > The two charts are generated completely independently of each other. > (This independence gives additional information especially to experienced > users.) In the Monte-Carlo percent diagram each data point stems from 1,000 > random games starting at the corresponding position. > For reasons of computational speed only every third position is evaluated: > after moves 3, 6, 9, 12, ... The positions after moves of Black are related > to black dots, the others after moves of White. > > It is a bit strange that ManyFaces generates rather large fluctuations for > the very first data points. Perhaps it would make sense to start the data > points only at move 20 or so (David, do you hear me?). For me it would be > also nice to have an option computing the % not only for every third but for > every position in the game. > > Regards, Ingo. > -- > Empfehlen Sie GMX DSL Ihren Freunden und Bekannten und wir > belohnen Sie mit bis zu 50,- Euro! https://freundschaftswerbung.gmx.de > _______________________________________________ > Computer-go mailing list > Computer-go@dvandva.org > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go _______________________________________________ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
