Issue #31 resolved

some feature gammas might be color dependent

dsmic
created an issue

I am running tests with self trained feature gammas compared to original ones. They are run on 19x19 against GnuGo 3.8 L0 with 3000 playouts/s with the attached gomill.cfg (only the loadfeaturegamma line is different).

the black white ratio is different with a significance > 2 sigma?!

with original gamma: oakfoam v gnugo (487/1000 games) board size: 19 komi: 7.5 wins black white avg cpu oakfoam 232 47.64% 107 43.85% 125 51.44% 478.38 gnugo 255 52.36% 118 48.56% 137 56.15% 22.08 225 46.20% 262 53.80% with my feature gammas board size: 19 komi: 7.5 wins black white avg cpu oakfoam 194 48.87% 108 54.27% 86 43.43% 521.22 gnugo 203 51.13% 112 56.57% 91 45.73% 23.71 220 55.42% 177 44.58%

Comments (8)

  1. dsmic reporter

    I found a bug in the experimental circ pattern code, it was probably indirectly effected by the new gammas.

    I will check out further.

  2. dsmic reporter

    It is fixed but not pushed yet: but I talk about circ patterns only. You did not pull it up to now anyway, as it are too experimentel. I will write a mail if I have more information. Your were right, I will need to have some kind of statistics of not played circ patterns, I guess I will take something like 3 random moves in every board position where I take a played one, and hope it is fast enough and good enough to get an idea of how often these patterns exist...

  3. Log in to comment