Seth Pellegrino wrote: > Hello again, > > I've gotten my speed issues mostly resolved (there's always more > optimization to be done), but I'm still running into a roadblock > reconstituting Dr. Coulom's results. I downloaded the same data set > that he uses (All the games from KGS on January 2006 with no handicap, > available at http://www.u-go.net/gamerecords/), ran it through my > engine code to generate the teams/competitions, fed that result > through Dr. Coulom's majorization-minorization engine (available at > http://remi.coulom.free.fr/Amsterdam2007/ ), and got some gammas > significantly different from the ones he did. I've attached the a file > with my full results*, but some are even qualitatively different. For > example, capture level 2 (re-capturing the previous move) gets a gamma > of 0.53 in the paper, which means that its presence in a team should > reduce the probability that team will be chosen. However, my computed > gamma was a startling 14.299, indicating that level of that feature is > a strong positive presence. On the other hand, my computed gammas for > things like distance seem to match relatively closely to the paper.
That may depend on some details of the features. I don't have time to check in details. Maybe I always multiply "capture of previous move" by "distance to previous move equals to one". Maybe also I make a difference between "capture a ladder" and "capture a non-ladder". I would not worry about such a difference as long as your result matches your intuition of what is a good move. Remi _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
