RE: [gnugo-devel] remaining patches

[Portela Fernand]

Dan wrote:

> Maybe you explained this and I missed it 

I didn't.

> but how do you measure success rate?

I run regressions with a patched GG which collects "good" and "bad" hits on
the patterns. "good" is defined as "move which doesn't give the opponent a
WIN", "bad" is simply the opposite. For it to be fair, a couple changes
have to be made: all patterns are 'hard-checked' before entering the list
in shape_patterns, and when hits are updated for a pattern, any other
pattern in the list which generates the same move is updated as well. This
produces statistical data which isn't perfect (because of the cutoffs), but
which proved useful and rather reliable. Also, I have a couple other
features in this patch, like the ability to load pattern values from a file
(from the time I was trying to automate the tuning of these). 

I didn't submit a patch for this "tuning-helper", because I wrote it quickly
and quite "dirtily". If somebody is interested though, I'll clean the code
and submit a patch.

/nando