> In other words, pachi's dynamic komi should be twice as high as it is ?
I should have been more careful in my choice of words. There can be a big difference between the theoretical value of the first move (which is twice the komi), and the parameter value which should be used to make the bot play best. Bots have many tuning parameters and most of the time the only way to get the optimal values is to try, measure and keep what works best. I will try this experiment and report what works best for Pachi. It's quite possible that the optimal value will be somewhere between 7.5 and 15. Jean-loup 2010/2/11 Le Hir Matthieu <mate...@hotmail.fr> > In other words, pachi's dynamic komi should be twice as high as it is ? > Currently, on first move - at 9 handi - pachi has an advantage of about 66 > points thanks to his komi. It should be about 130? How would that affect his > play ? More second line, slack and "no man's land" moves ? > > My reasoning may be silly but... If the initial dynamic komi is doubled, > the bot has to "work" twice less hard and can play loose, slack moves and > still keep the lead , right ? > As opposed to this, halving the initial komi from 66 to about 35 points, > wouldn't that force to bot to play more efficient, sharp moves ? > > What I'm afraid of, though, is that after a while ( when the bot has caught > up enough) it might start again to play slack again ? > > Matthieu > > > > *-------Message original-------* > > *De :* Jean-loup Gailly <jl...@gailly.net> > *Date :* 11/02/2010 16:00:33 > *A :* computer-go <computer-go@computer-go.org> > *Sujet :* Re: [computer-go] Dynamic Komi's basics > > > Pachi uses 7.5 points per handicap stone > > Pachi is wrong. See the first paragraph of "There is a relationship" > in http://senseis.xmp.net/?Komi/valueOfFirstMove > The rest of the page is quite confusing but but the value of one move > at the beginning of a game is definitely twice the komi. > A move early in the game is worth about 14 points, not 7.5. > I would write the proof as follows. > > Assume x is the value of one move, and komi is the value of the right > to play next. After playing the first move, black has won x points on > the board and lost the right to play next, so blacked gained x-komi in > total. White has gained the right to play next, so komi points. For > this to be fair, the value gained by white must be equal to the value > gained by black: x - komi = komi or: x = 2*komi. > > Jean-loup > > > 2010/2/11 Petr Baudis <pa...@ucw.cz> > Hi! > > On Thu, Feb 11, 2010 at 10:32:49AM +0100, Le Hir Matthieu wrote: > > First, I'm wondering how komi is determined when a dynamic system is used > : > > > > * According to this page : > http://senseis.xmp.net/?Komi%2FvalueOfFirstMove > > the value of komi at first move is half the points the black move is > > supposed to be ( 14 points) > 7,5 komi in even games. > > How does this apply to high handicap games ? > > I *think* that komi value is miai-counted, not deiri-counted, so in miai > counting that corresponds to the 7.5 value also proposed in the article. > I have never been an expert in the counting theory, however, so maybe > I am wrong. > > Pachi uses 7.5 points per handicap stone, in some testing I also used 8 > points per stone which gives similar winrates. It makes no theoretical > sense to use the 0.5 fraction in this, but it seems to work well enough > and I did not have time to tune for better value. > > > From what I think I understood, dynamic komi is supposed to try to keep > the > > game more even. > > If the computer is black, playing at 9 handi, will the " burden komi" ( > > negative) be 9 x - 7,5 ? - 67,5 ? It sounds highly improbable. > > Yes, this is the komi value pachi uses. I have found that higher komi > values work similarly well (perhaps a bit worse) as long as the komi is > not completely ridiculous. > > > On the other hand, 9 handicaps are supposedly giving an advantage of 90 > to > > 120 points, so my natural thought would be that the bot would give itself > at > > least a negative komi of that many points ? > > > > I can't figure out well how komi is determined at first move. > > extra_komi = 7.5 * handicap_stones_count > > Then it is linearly decreased until it hits 0 at move 200. > > This is the most naive implementation. In practice, neither extra_komi > determination nor its application throughout the game should be linear, > probably. There is a lot of experiments to be done. :-) > > > In relation with the previous question, I'm wondering how komi is > determined > > and what its value is for every handicap game ( as black and as white). > Is > > there a specific value for each ( before first move is played) or is it > only > > determined by the way it was programmed and the programmer's preferences > ? > > It is only determined by the programmer's preferences and more > importantly empirical measurements of what kind of values worked best. > > > P.S.: For some measured dynamic komi positive effects, please see the > presentation I posted few months ago. I plan to finish up and submit a > paper with investigation of also other approaches hopefully in time for > the IEEE call for papers. ;-) > > -- > Petr "Pasky" Baudis > A great many people think they are thinking when they are merely > rearranging their prejudices. -- William James > _______________________________________________ > 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/ > > [image: Animations GRATUITES pour votre messagerie - par IncrediMail! > Cliquez ici!] <http://www.incredimail.com/?id=605280&rui=122199674> > > _______________________________________________ > 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/
