>> I would like to know what exact experiments with "virtual komi" >> have been made and why thay failed. ...
I'm only aware of Don's experiment [1], which he admits he doesn't have any details for and only remembers: "I did a bunch of experiments and ALWAYS got a reduced wins when I faked the komi". On the other side we have some experiments by Kato-san [2] (where he reports a 100 ELO improvement over GnuGo, but only from a few tens of game) and a subjective experiment by Okasaki-san where he "reported Mogo played clearly stronger" on KGS [3]. My own experiments are even more subjective and small-scale, and in the context of 9x9 endgames, not 19x19 handicap openings. However they were enough to make me think the technique is viable, but that if you don't adjust the komi down so the winning rate is near 50% it is wasted effort (*), and so you need to replay the same move over and over with different komi until you zero in on that point. *: I.e. the program still plays weak moves if you've only adjusted komi to go from 80% to 65%, or from 25% to 35%. >> kill all - instead you just overplay a little in order to catch up >> slowly but steadily. > > You just hit the nail on the head. Dynamic komi does not encourage > a program to overplay the position. Since you are starting from a > losing position you HAVE to overplay a bit. You have to attack when > it is futile. If the handicap is correct then you don't really need to overplay. As the stronger player you might guide the game towards more complex positions to encourage more mistakes, but mainly you are just sitting around waiting for those inevitable mistakes. But, the real point of adjusting komi is it is an easy to understand way to overcome MCTS's problem when seeing all moves as winning/losing, and choosing effectively randomly instead of falling back on an opponent model as a human would do. Ingo's suggestion (of two buttons to increment/decrement komi by one point) was to make it easy for strong humans to test out the idea for us. Darren [1]: http://computer-go.org/pipermail/computer-go/2008-August/015870.html [2]: http://computer-go.org/pipermail/computer-go/2008-February/014283.html [3]: http://computer-go.org/pipermail/computer-go/2008-August/015877.html _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
