Re: [Bug-gnubg] Alternative weights files and call for benchmarkers

[Joseph Heled]

Hi again,

While the benchmark shows the new net(s) are better than the old, one must remember the improvement in playing ability is not large.

In 1 million 5 point matches between the two, the new net wins in 50.084% of the matches. This result is not even strongly significant, at about 2 sigma.

Also, the new crashed net is still hopeless at advancing a prime from the deep. Until it does a reasonable job there, I see little hope for gnubg advancing to the next level in the crashed area.

Here is an example. Unfortunately gnubg can't import this file because the Board: is unsupported. Perhaps someone can look into this? (My own fibs2html handles this fine. but it is very old by now)

-Joseph
 
Score is 0-0 in a 1 point match.
gnu0plyX is X - gnu0plyO is O
Board: -10 0 0 2 2 2 2 2 2 1 1 1 0 0 0 0 0 0 0 0 0 0 -1 0
O: (5 4)  21-16  20-16
X: (1 5)  23-22  22-17
O: (1 2)  16-15  16-14
X: (5 5)  17-12
O: (5 3)  25-20  20-17
X: (4 6)  can't move
O: (5 1)  21-16  17-16
X: (6 6)  can't move
O: (3 3)  19-16  19-16  14-11  11-8
X: (2 3)  can't move
O: (1 1)  20-19  19-18  18-17  17-16
X: (2 3)  can't move
O: (6 4)  14-10  8-2
X: (3 6)  12-6  6-3
O: (2 4)  15-13  15-11
X: (6 1)  3-0  1-0
O: (5 2)  17-15  16-11
X: (1 3)  1-0  1-0
O: (6 2)  17-15  13-7
X: (6 4)  1-0  1-0
O: (2 2)  16-14  11-9  10-8  8-6
X: (4 4)  1-0  1-0  1-0  1-0
O: (5 2)  16-14  6-1
X: (6 1)  1-0
X: wins 2 points