> I would like to know what exact experiments with "virtual komi" have > been made and why thay failed. .. > So why exactly shouldn't it work?
(warning: I have not tried this yet, just been thinking about the issue) While virtual komi sounds like an attractively simple way of solving the issue, I don't think it does so without fail - it is slightly too simple. The issue, for me, is that MC's evaluation function is a mixture of statistics and cut-off, so if all the statistics remain below or above the cut-off, moves can no longer be distinguished and play becomes random (the cut-off is important: just going for best average score does not appear to perform as well as going for best potential to win). Just adding virtual komi shifts the cut-off, which should help to make evaluation more selective in the problematic cases. Unfortunately, it does so at the price of making evaluation less selective in formerly unproblematic cases (what previously was the clear best move now has to "compete" with lesser moves that also make the lower cut-off). What seems to be needed is a stratified approach: 1 try to make a clear choice of best move with unmodified MC 2 only if 1 fails, move the cut-off and try again 3 if nothing helps, resign (not just polite, but speeds up testing;-) As usual, one could trade time vs space: instead of repeating the statistics with different cut-offs, record all moves wrt different cut-offs (a small range around the komi) in the first run, then use the statistics for the cut-off that is closest to normal komi but still able to distinguish better moves. Claus _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
