Many Faces still uses UCT to control exploration. David
> -----Original Message----- > From: [email protected] [mailto:computer-go- > [email protected]] On Behalf Of Petr Baudis > Sent: Monday, March 08, 2010 12:30 PM > To: Nick Wedd > Cc: computer-go > Subject: Re: [Computer-go] Congratulations to Many Faces of Go! > > On Mon, Mar 08, 2010 at 04:43:29PM +0000, Nick Wedd wrote: > > Congratulations to Many Faces of Go, the undefeated winner of > > yesterday's KGS bot tournament. > > > > My report is at http://www.weddslist.com/kgs/past/57/index.html > > I hope you will report any errors, as usual. > > Thank you for the report! > > On the hardware - Puego was r1098, still run on the i7 920, but this > time with 8 threads instead of 4. CzechBot was still run on the same > hardware. pachi2 was AFAIK running on 21 cores, but probably not of > three Mac Pro machines - I hope Jean-Loup will be able to fill on the > details. > > I beg to disagree with the analysis of the pachi2-ManyFaces1 game. > I think that the move at 177 was neccessary, there does not seem to be > alternative for white to live with its roup. Also, pachi2 kept being > very optimistic for the rest of the game and I also had the feeling > that the game looks favorable for white (though I did not count). > It seems to me that the blunder was only at the D19 move (267); > leaving the ko seems to be too dangerous and the bottom corner is not > enough compensation, especially given black's nasty followups. IMHO > the bot did not read the E19 ko correctly; I do not have detailed > winrate logs (yet?), though. > > Thank you for the praise of Pachi. :-) I'm sorry I confused you > on KGS previously, the bug was not in the codebase that was used > in the tournament. It does seem our flaw indeed is time management, > though in normal byoyomi games its clever implementation by jlg > boosts the strength - the problem seems to be poor handling of S.D. > and S.D.-like settings, probably due to faulty changes of mine. > > The surprise of the tournament for me was the excellent stv's > performance, looks like Steenvreter is to be reckoned with - I wonder > if its author plans to divulge more details on its internal operation. > > Finally, a general note on the UCT idea conclusion - I would say that > while the UCT is evolutionary at the beginning of current bots, the > current RAVE-based MCTS algorithm actually has nothing to do with UCT > anymore since the upper confidence bounds are not used anymore - the > RAVE information fully substitutes them as online exploration bias. > Perhaps "RAVET" instead of "UCT". ;-) > > P.S.: I wish I would name Pachi's key component 'mcts' instead of > 'uct' now. :-) > > -- > Petr "Pasky" Baudis > http://pasky.or.cz/ | "Ars longa, vita brevis." -- Hippocrates > _______________________________________________ > 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
