In Remi's paper on ELO ratings of moves, how is mean log evidence computed? Is that looking at the probability of the training set? e.g. if the selected moves have estimated probabilities of 1/e, 1/e^2, 1/e, and 1/e, then the log evidences would be -1,-2,-1, and -1 for a mean log evidence of -1.25?
I guess I'm kind of surprised by the mean log evidence being about 1/14 with a prediction rate of 34% (forgive me if I misremembered the numbers). I guess I can rationalize that, but it leaves me very curious what the distribution of prediction probabilities are that lead to that result.
_______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
