steve uurtamo wrote: > yes, and the fact that turning a dumpling into a dead > group can take more than a few moves, since you may > have to fill up the eyespace several times, meaning > going fairly deeply down branches with several self-ataris > along the way. > Ok, it's pretty much as I thought. There are relatively simple solutions but they will slow down the play-outs significantly unless someone finds a creative fast solution.
I think the general outline is that you pre-test groups first to see if a self-atari move is "interesting." It's worthy of additional consideration if the stones it is touching have limited liberties and the group you self-atari is relatively small. Then you could go on to other tests which will consume even more time of course. This is probably one of those changes necessary to improve the scaling curve. The slowdown will hurt at low levels and at some point it will break even, then be stronger. - Don > s. > > On Thu, Mar 6, 2008 at 9:23 AM, Don Dailey <[EMAIL PROTECTED]> wrote: > >> > >> > >> > You won't find that in computer vs computer games, because "tricking" the >> > strong programs requires some go skill and it only works if you wait long >> > enough before you "solve" the position. But if you search KGS (LeelaBot, >> > CrazyStone, CzechBot) for even games where the bot lost against a kyu >> > players you will find many. All go more or less like that: >> > >> > A 4-6 kyu human is behind by 10-15 points in the midgame (at that >> > stage the >> > probability of winning is correlated with territory, so the MC bot is >> > building fine.) He creates a 12-16 point worth nakade trick in a corner >> > and does not solve it.The bot is happy, it thinks a bulk five is alive or >> > something like that. Perhaps the human sacrificed another 15 points >> > somewhere to create the trick so he should be dead lost. But, he only >> > has to play on, reduce, etc. As the endgame approaches, the MC bot >> > allows the reduction only until the territorial balance would change the >> > winner. The player is happy, he turned a 25 points loss into a 1.5 point >> > loss (assumed by the program) and has a 12 point surprise. >> > At the end, when the whole board is decided, the player kills >> > the bot's group and the bot turns a sure win into a sure loss and >> > resigns. >> > >> > Because the trick can only be played by similar strength players (much >> > weaker players can't build something like that, much stronger don't >> > need it) >> > it affects the rating of the bots. I guess CrazyStone could be near >> > KGS 1dan >> > with that solved. It is 2k now. But, of course, the solution may not >> > come at >> > the price of making the program weaker. That is the difficult part. >> >> I want to make sure I understand the nakade problem, please correct me >> if I am wrong: >> >> My understanding of this is that many program do not allow self-atari >> moves in the play-outs because in general the overwhelming majority are >> stupid moves. Is that what is causing the nakade problem? And if >> you start including self-atari you weaken the program in general? >> >> And can I assume the tree portion is also inhibited from seeing this due >> to a combination of factors such as heuristics to delay exploring "ugly" >> moves as well as the weakness of the play-outs in this regard (which >> would cause the tree to not be inclined to get close enough to the issue >> to understand it properly?) >> >> - Don >> >> >> >> >> >> > >> > >> > Jacques. >> > _______________________________________________ >> > computer-go mailing list >> > [email protected] >> > http://www.computer-go.org/mailman/listinfo/computer-go/ >> > >> _______________________________________________ >> computer-go mailing list >> [email protected] >> http://www.computer-go.org/mailman/listinfo/computer-go/ >> >> > _______________________________________________ > computer-go mailing list > [email protected] > http://www.computer-go.org/mailman/listinfo/computer-go/ > > _______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
