Problem 39246

[palemon]

According to Rybka analysis winning move Ke3 is not even one of 3 best moves:

Analysis by Rybka 2.1o 32-bit:

1. -+ (-10.94): 48...Kc2 49.Be1 Be5 50.Bb4 Kd1 51.Ka8 e1Q 52.Bxe1 Kxe1 53.Ka7 Bc7 54.Ka8 Kf2 55.Ka7
2. -+ (-9.29): 48...Be5 49.b8Q Bxb8+ 50.Kxb8 Kc2 51.Kc7 Kd1 52.Kd7 e1Q 53.Bxe1 Kxe1 54.Ke6 Kf2 55.g4
3. -+ (-6.37): 48...f5 49.Kb6 Bb8 50.Kc6 g4 51.Kd7 g3 52.Kc8 Ba7 53.Bb4 f4 54.Be1 f3 55.gxf3

but all these moves are not even "keep looking" moves

[richard]

After a long time Toga eventually agrees a bit more with Rybka:

+10.15 48...Kc2
+10.02 48...Bf4
+9.89 48...Be5
+6.01 48...Ke3

Unfortunately toga in the short term sees this:
+4.29 48...Ke3
+1.95 48...f5
+1.84 48...Ke4
+1.3 48...Bd6

(evaluation scores are obviously from black's point of view).

I'm a bit surprised it didn't find Kc2 earlier as it doesn't take that long to find.  Some of my generator machines have to do other tasks than generation and it possible one of the generator CPUs got starved for a while so didn't see as much as it should have. I'm hoping to have machines dedicated to only problem generation but in the shorter term I'm going to collect data on depth reached and nodes processed during search so I should be able to detect situations where a CPU didn't get enough processing time during it's search. I'm also nearly finished the changes required to keep the computer evaluations and lines around during the generation phase which would allow to me check exactly what the engine supplied to the generator code after a run is complete.

Regards,
Richard.