A bug where gnubg offers to resign a single game from :
GNU Backgammon Position ID: OwAAwL13gAAAAA
Match ID : AQEAABADgAAE
+-1--2--3--4--5--6-------7--8--9-10-11-12-+ O: GNUbg (Cube: 2)
| O O O O | | O | 49 points
| O O O O | | | On roll
| O O O O | | |
| O O | | |
| | | |
| |BAR| |^
XX | | | |
XX | | | |
XX | X | | |
XX | X X | | |
XX | X X | | | 16 points
+24-23-22-21-20-19------18-17-16-15-14-13-+ X: You
Pip counts : O 47, X 8
was reported at bgoline.org.
The problem is a combination of the neural net underestimating gammons
lost by O :
Position ID: OwAAwL13gAAAAA
Match ID: AQEAABADgAAE
Evaluator: Race
Win W(g) W(bg) L(g) L(bg) Equity Cubeful
static: 0.000000 0.000000 0.000000 0.000050 0.000000 -1.000050 -1.000050
1 ply: 0.000000 0.000000 0.000000 0.001895 0.000000 -1.001895 -1.001895
2 ply: 0.000000 0.000000 0.000000 0.001417 0.000000 -1.001418 -1.001418
and the following code in the SanityCheck() function :
float noise = 1 / 10000.0f;
for (i = OUTPUT_WINGAMMON; i < NUM_OUTPUTS; ++i) {
if (unlikely(arOutput[i] < noise)) {
arOutput[i] = 0.0f;
}
}
I don't know the rationale for it. It seems it is from quite early in
gnubg development, maybe it predates the other checks that are in this
function. The latter seem to be enough to handle the certain outcomes in
non-contact positions.
Is it really useful to keep these truncations for contact positions ?
_______________________________________________
Bug-gnubg mailing list
[email protected]
https://lists.gnu.org/mailman/listinfo/bug-gnubg
