[computer-go] ELO Ratings of move pattern

[Rémi Coulom]

Jason House wrote:
>
> This may serve as a good test of if there is enough data to assign 
> values to the patterns.

I did not mention this in my paper, but you can rather easily estimate 
uncertainty margins around Elo values. This involves computing the 
second-order derivative of the probability distribution with respect to 
log(gamma). Since the distribution has a shape that looks very much like 
a Gaussian, the second-order derivative at the maximum is a good 
estimation of -1/sigma². That is how I compute confidence intervals in 
bayeselo.

Also, in the conclusion of my paper, I suggest that it may be better to 
avoid using consecutive positions in the same game as learning data, 
because they are not independent, and the whole learning process assumes 
independence. Some quick experiments I made tend to indicate that this 
is important. Using 3 times more games, and keeping only one position 
out of 3 improved learning. I did not investigate this very deeply, so I 
cannot give more precise data, but keep in mind that it is dangerous to 
use consecutive positions because their are too similar, and may lead to 
over-estimation of the confidence in pattern evaluation.

Rémi