|From:||Eberlein, Robert (REBERLEI)|
|Subject:||[Bug-gnubg] gnubg backgame failings|
|Date:||Thu, 8 Jan 2004 13:29:28 -0500|
Backgame analysis by gnubg is showing two severe problems: 1) The equity discrepancy between even and odd ply is very large. 2) Having borne checkers off, the non-backgame side is losing gammons and backgammons in the rollout! I am using 0.14-dev3D 1.1490 040104 (build Jan 4 2004), Gnu Backgammon for windows.
Position ID: YEYF4TnO7w4AAA
Match ID: EQH2AAAAAAAA
gnubg is on roll with a 5-4 to play. The two legal plays are 6/2 5/off, and 5/off 4/0ff. The evaluations of the equity of these choices vary hugely with the ply:
Play 0-ply 1-ply 2-ply 3-ply
6/2 5/off 1.114 0.303 1.007 0.269
5/off 4/off 1.027 0.278 0.939 0.219
I tried doing 0-ply rollouts of the position to learn more, but stopped when the side bearing off lost gammons and backgammons! The illegal losses occurred both with and without variance reduction.
I'll assume the gammon and backgammon loss is a bug that can be fixed. I think that the large equity swings as a function of evaluation depth reveal a serious problem with the outputs of the neural network for this type of position, and I wonder if anyone with Snowie could report what the Snowie evaluations are here. If Snowie does not show the same problem, then gnubg is broken in this area.
|[Prev in Thread]||Current Thread||[Next in Thread]|