It's been a while for me too. What I remember is that one can equivalently use scores that are the log of the gammas, and then the whole thing looks just like a natural generalization of logistic regression, where you basically have a bunch of features that you combine linearly, and then the probability of each result is proportional to the exponential of that linear combination.
On Tue, Apr 2, 2013 at 5:59 PM, Jason House <[email protected]>wrote: > On Apr 2, 2013, at 4:08 PM, Don Dailey <[email protected]> wrote: > > > I like that paper a lot, but I really had a difficult time swallowing > the assumption that a team could be rating by summing together each > individual member. The paper of course makes a disclaimer that it may > not be a good assumption "all the time" but in my view it's rarely a good > assumption - unless you are playing tug of war. > > > > It may be a reasonable however if the number of features in the pattern > you are comparing is the same. Then summing the ELO's of the features is > exactly the same as taking the average. I don't remember if that was the > case here. > > It's been a while since I read the paper, but I do believe the number of > features per point is constant for the data set Remi used. He had > independent grouping of features, and each group would have one member for > each point on the board. I logically treat "absent" as an extra level with > a gamma of 1.0. With such exhaustive partitioning, one member of a group > must have a gamma of 1.0 or there won't be a unique result. For example, > all gammas in the "MC Owner" category could be multiplied by 1000 and it'd > still yield the same probability distribution. > _______________________________________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go >
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
