On Wed, Apr 02, 2008 at 02:13:45PM +0100, Jacques Basaldúa wrote: > Jonas Kahn wrote: > > > I guess you have checked that with your rules for getting probability > > distributions out of gammas, the mean of the probability of your move 1 > > was that that you observed (about 40 %) ? > > If I understand your post, there may be a misunderstanding by my fault. > Here gamma is not a gamma function nor a gamma distribution but a constant. > It is just the same Greek letter. I don't remember if it was in > Crazystone's > description or in some other paper I read to understand what Bradley Terry > models are, I just got used to that notation.
Sorry, you were clear, it's only me who did not remember BT correctly. I thought of a transform P = Lambda(gamma2 - gamma1) in the end, like for Elo estimation. The heavy tails could still very well come from inadequacy of the formula P(i) = gamma(i) / sum_j gamma(j) for the smaller probabilities. The most natural explanation for your underestimating the proability that the first move is the right one would come from correlations between patterns. A short toy model where only the first move can have one or two patterns, with same individual value, but with varying combined value, suggests that correlations between patterns should be positive to explain your result. I am a bit surprised, since I would have thought that correlations between patterns would rather be negative... Maybe you can test that in the following way: if you have say 500 patterns that have a much greater value than others, you could add to your learning an entry 'there is both pattern 1 and pattern 2' for all pairs of patterns among those 500, and see if you get better predictions in the end. Of course, if too many patterns are really relevant, this is hard to do. Anyhow, as you said, what's important is having good suggestions for urgent moves, not this kind of discrepancies. Jonas _______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
