[gnugo-devel] 261 games 3.1.31 vs 3.1.32

[bump]

Since shortly after the 3.1.31 release, we've had a precipitous
rating slide on NNGS. I did some testing to see if 3.1.31 showed
signs of being broke.

I've been playing 3.1.31 against 3.1.32 and have collected 116 games with
3.1.32 as white, and 145 games with 3.1.32 as black.

In the 116 games with 3.1.32 as white, black won 68, a win rate of 58%. 
In the remaining 145 games with 3.1.32 as black, black won 84, again
a win rate of 58%.

Perhaps these results are too small to be statistically significant.
They suggest that 3.1.31 and 3.1.32 are evenly matched, but that
5.5 points is not sufficient komi. Perhaps the correct komi is
about 7.5 points.

If anyone wants to look at these 261 games, they are available at:

http://match.stanford.edu/gnugo/games_32_31.tar.gz

I also played 200 games against 3.0 with 3 stone handicap. 3.1.32
won 101 of these, a win rate of 50 %. I did see a higher win rate
than that at one point. It is about the same as I got in a 50 game
series involving 3.1.29.

Dan