terry mcintyre wrote: > I understand that Monte Carlo algorthms are driven by the "winning > probability", and a 0.5 win looks as good - or maybe even better - > than a 100-point win. > It doesn't view 0.5 as "better"
> However, the estimated probability of winning may be way off. It is > well known that Mogo, and perhaps some other programs, fail to > recognize common nakade placements, which leads to fundamental > estimation errors. An algorithm with more of a "fighting spirit" would > defend against nakade, and attack enemy groups; perhaps making up for > the loss of one group by the capture of another. > > Any algorithm which drives the win toward 0.5 is always going to be > brittle; It doesn't "drive" the win towards 0.5. It doesn't view them as any differently. However, they will prefer a bigger win if there is any room for error. Usually a bigger win is a more likely win - it's only in the cases where it isn't that Monte Carlo program do not care. This is a fundamental error in how people think about this. Your intuition is that you should try for a bigger win just in case - or that it improves your overall winning chances. But if 10,000 monte carlo playouts see one line as winning 10,000 times and another line as winning 9,999 times, even if most of those wins are BIG, it will choose the sure thing. Another way to look at it is this: If there are 2 key groups being fought over, and winning either one wins the game, it will choose the group that it is MOST likely to win - even if it is far smaller. There really is no way to improve on this except to trick it into NOT maximizing it's winning probability. You might end up with a program that appears to play more human, but it will sacrifice some playing strength. It might be possible to make it "break ties" when all else is equal and get more natural play without sacrificing playing strength. You say: When monte carlo program lose, they lose big. Yes, that tends to be true. They essentially give up in dead lost positions (even though it may not be obvious to us.) But they NEVER gradually drift into a losing position because they play passively when they should be fighting. If they should be fighting they will fight and play very purposefully. > any problems with the accuracy of the evaluation may place one on the > losing side of that 0.5 divide. Hence, if it is possible to be more > greedy - to seal off groups, to play for every yose point, to fill in > dame first - programs based on monte carlo will improve. No they won't. If you can do this in such as way that the maximizing behavior is not sacrificed, it's possible that they will salvage some lost games. But these will be DEAD LOST games they salvage, not games where there is some hope (by their estimation) - because if there is some hope this very behavior you are criticizing them for will cause them to focus very heavily on ANY move that gives them a fighting chance - which won't be a silly non-move. > It's terrible to lose a won game by drifting over the edge of a > precipice. When monte carlo programs lose, they lose big -- in my > so-far limited observations, incorrect evaluation of life-and-death > status leads to estimation errors which far exceed the 0.5 margin of > error. > > For a large number of playouts, the estimated scores should converge > as the game progresses. This is particularly true if the random > distributions strongly favor moves where each opponent monotonically > increases the score - keeping one's groups alive, keeping the opponent > dead, and growing areas wherever possible. Of course there must be > enough variability to permit sacrifice plays and nakade placements; > throwing a stone into a group will initially look like a bad play, but > if the placement succeeds, it is a very good play, the success of > which must be properly attributed to the earlier placement - or even > further back, to the surrounding and cutting and eye-killing moves > which ultimately led to the placement move. > > Terry McIntyre <[EMAIL PROTECTED]> > They mean to govern well; but they mean to govern. They promise to be > kind masters; but they mean to be masters. -- Daniel Webster > > > ------------------------------------------------------------------------ > Never miss a thing. Make Yahoo your homepage. > <http://us.rd.yahoo.com/evt=51438/*http://www.yahoo.com/r/hs> > ------------------------------------------------------------------------ > > _______________________________________________ > 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/
