I've reviewed most of your positions posted on the
GammOnLine forum. While I am no GNUbg expert, I'd say the positions all
looked like they would be evaluated by the "contact" net. If I understand
Joseph, Joern, et al's work correctly, recent and near-future
improvements are to the "race" net, bearoff databases, and the
"crashed" net.
To me that implies most improvements to GNU's
playing strength will be in positions such as bear-in, bear-off, gammon-save,
and post-back-game or post-ace-pt game containment. To be certain as to
which Net is making the evaluation, the Analyze/Evaluate command gives an output
which includes the Net evaluating the position. There was a small bug-fix
which corrected GNU from choosing the "crashed" net in a few positions that
should be evaluated by the "contact" net.
The other significant improvement that could affect
GNU's evaluation of certain positions has been the implementation of
move-filters. I know from your other post that you're aware of this
improvement. It's possible that certain positions could be played more
strongly by GNU using a wide filter and disabling 1-ply evaluations (although
this could have been achieved in the earlier version by setting a large number
of moves, wide tolerance, and setting "no 1-ply pruning").
Ned
----- Original Message -----
Sent: Monday, December 30, 2002 5:06
AM
Subject: [Bug-gnubg] Latest Build
Improvements
As you probably know, I
had GNU and Snowie 3 play a 500 game money session and rolled out all the
disagreements. I used the 15OCT no-gui.exe version and have the 19NOV Windows
build currently installed. Is this the latest build?? I know of the
race net and the cubeful no contact bearoff database but what other neural net
improvements have there been that would affect some of the positions I
analysed? thanks
_______________________________________________ Bug-gnubg mailing
list address@hidden http://mail.gnu.org/mailman/listinfo/bug-gnubg
|