On Wed, 2007-04-04 at 10:40 -0700, terry mcintyre wrote: > If both Monte Carlo players strive for 0.5 point wins, then almost any > ballpark komi would lead to a 50/50 split, a sort of self-fulfilling > prophecy?
If the komi is wrong, you would see it in the score of white vs black over a large number of games. Someone once suggested that I take the scores on CGOS, and "see what would happen if I pretended the komi was various different values." The idea was that I could find the "correct" komi. But that would fail badly with the monte carlo programs. It's very common for Lazarus to be content with a half point win when it could easily get a much larger win. And of course it WOULD go for the larger win if it were necessary because of komi. Many other programs are like this too. However, it would be accurate to play a few thousand games at 7.5 komi, then a few thousand and 6.5 and compare the white/black win percentage. The one closest to 50% would be the one to use. > What happens when a player sets a more difficult komi than the one > used for scoring? Sometimes Go players use larger komi as a sort of > handicap. Would it be possible to encourage more substantial wins by > tweaking the "internal komi" used to drive move selection in this > fashion? I think a lot of us have tried this. It's an interesting try - but it doesn't work! I tried this in Lazarus by constantly changing the komi so that the program though it was losing somewhat, but not hopelessly. Of course this causes it to take more chances - even when it has a sure win. You are essentially forcing it to try desparate moves by throwing away safe moves which it now views as just not cutting it. There is a pretty safe way to fix this behavior to an extent. You have to have an external procedure which orders the moves in a more natural way (prefering big wins) and then you can give UCT at the root just a very slight incentive for prefering that move ordering. This works best when the game is almost a forgone conclusion (in either direction) but there are still points to be won and lost, however irelevant. Personally, I could care less - I guess because I am a chess player and I think it's weaker players who are impressed most by big wins. Very strong chess players tend to take the long way around - taking the sure win to the dramatic flashy quick but risky win. Weaker chess players tend to judge the skill of the player based on how many moves it takes to win a game but strong players know this is more a matter of playing style - not skill. It actually surpises me that go players care about this. I thought GO was more about the beauty of ommision and the unstated understanding of event that don't actually have to happen to be appreciated. - Don _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
