With all those nets I get confused.
Which are exactly the "old" in "new" (in a nngnu ready format too), and more important, where can I download them now.
In my gnubg directory I see
$ md5sum gnubg.weights*
d182a162a011aac839bc48879c349b14 gnubg.weights
bb76fbaa6256f79dacc17749045d38ee gnubg.weights.newold
54d68d84633443825ff8c346b0283d55 gnubg.weights.old
and various nngnubg.weights around :)
-Joseph
On 18 March 2013 03:57, Philippe Michel
<address@hidden> wrote:
On Sat, 16 Mar 2013, Neural Gnat wrote:
I've just re-analysed a 1000-game money session that I did about a week ago with 2012's World Class versus Casual. This new test version has found 1839 doubtful moves, 304 bad moves and 247 very bad moves, knocking the mainstream version down from Supernatural to World Class (-4.0).
On the other hand, if old has, say, an error rate of 4 vs. perfect play and new has 3 due to different mistakes, they may well be 4 away from each other.
The question is, how do you determine which of those opinions are correct? Dare I mention XG? ;o)
Roll out the disagreements. All of them would take time, of course, but only a few games' worth or the largest ones should give some idea of what is happening. Analysing these with XGR++ instead could be a reasonable shortcut and allow to look at more of them in a given time.
Another question is, how do I get these two versions playing each other? I tried the "socket" players a few years ago but, with no instructions, no result and no feedback from GnuBg, I soon gave up.
I don't know if the file is shipped in gnubg's Windows installation, but the comments a the start of matchseries.py here should help :
http://cvs.savannah.gnu.org/viewvc/gnubg/gnubg/scripts/
But don't expect to play the two versions against each other and get anything better than an anecdotal gross result. You would need a session *much* longer than 1000 games for a statistically significant result, and stock gnubg isn't suited to this. It keeps the whole session in memory and would likely get slower and slower and crash at some point.
You could still analyse the short session with XG at the highest level of luck analysis you can afford and get a useful variance-reduced result by a neutral third party.