On 2012-09-17 16:54, Philippe Michel wrote: > I have uploaded new weights at > http://www.gnubg.org/media/nn-training/pmichel/nets/20120907/ > > The difference since what I posted in June is mostly further training of > the contact net, and this time the result is statistically significantly > better than the current nets. > > new @ 0ply won 50.19% of 5 pts matches against current @ 0ply (1M trials) > > new @ 2ply won 50.69% of 5 pts matches against current @ 2ply (100k trials) > > The latter, if not an outlier, would be a great result. A possible > explanation for the bigger improvement compared to 0ply could be that with > the new nets, 1ply evaluation works adequately (wins 52.8% of 10k 5ptrs > against 0ply) when it is apparently not the case with the current nets. >
Thanks for the work! If this is true, and others sign off that this looks better (I'm sure some will want to verify it), it would seem like a good reason for another Update in the near future. -- Michael Petch CApp::Sysware Consulting Ltd. OpenPGP FingerPrint=D81C 6A0D 987E 7DA5 3219 6715 466A 2ACE 5CAE 3304 _______________________________________________ Bug-gnubg mailing list Bug-gnubg@gnu.org https://lists.gnu.org/mailman/listinfo/bug-gnubg