Hi, oakfoam was using large (circular) patterns (up to 9) in the last slow bot tournament. They were trained statistically. I read, that this is suboptimal and came up with an "idea" of another training:
During play agains gnugo I compared for every move the order of move quality, which results from the features, with the one which results from mc calculation. My guess was: the features should try to produce the same order as the result from mc calculation. If I unprune in the order of the features this is good, as I unprune more or less in the order of the mc quality. After every move I change the values of all features in the "better-fitting to mc order" direction by a very little amount. As this seems to be stable for all features including 3x3 patterns it tries to remove most of my large patterns. I double checked for bugs, but of cause there might be some left. Did somebody try a similar approach? My understanding is: The large patterns do not improve the order. As they clearly improve the playing strength (>100ELO for sure) there must be another reason, why they do it. My guess is: they introduce better tenukis into the tree to calculate different positions on the board. Therefore getting live and death better?! Or am I totally wrong and I should search very deep for a bug again:) Thanks Detlef _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
