Hi Erik,
Perhaps the interesting part is in fading out more slowly than with ordinary priors (i.e., by 1/sqrt(n) instead of 1/n)?
I agree with that. My experiment are as follows. In 1400 playouts, result is not good. In 10000 and 20000 playouts, it seems bigger G is better. I guess fading out too much is not so good. 10000po 19x19, K=600 constant, games 400 and 1000, selfplay G winrate 0.0025 0.557 0.005 0.622 0.01 0.653 0.02 0.574 0.04 0.446 0.08 0.074 0.16 0.044 1400po 19x19 0.01 0.481 0.02 0.390 0.03 0.277 10000po 13x13 0.005 0.589 0.01 0.671 0.015 0.698 0.02 0.696 0.04 0.603 0.08 0.297 20000po 9x9 0.01 0.611 0.02 0.643 0.03 0.638 Regards, Hiroshi Yamashita _______________________________________________ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go