Where it gets funny is when you use a learning engine !
By default, the learning engine is configured in training mode so even if the engine has evaluated all its moves, there is no evidence that those are the most effective moves in its experience file. The result does not always represent the quality of the experience file. Hence the interest of checking the rate of conformity of the moves played with the experience data.
In tourney mode (experience book, experience readonly, etc.), it's even worse because according to the "eval importance, min depth, etc." settings, the moves that will be transmitted to the GUI will have an evaluation coming from the experience data and it will not necessarily be the same evaluation as that of the engine (merged experience file, another net, etc.).
For example with this game played in training mode :
Code: Select all
[Event "EXP VS BOOK"]
[Site "dual xeon e5-2660v3"]
[Date "2022.07.10"]
[Round "1"]
[White "Depth4_050522 (Eman 8.10)"]
[Black "Solista-010722 (Eman 8.10)"]
[Result "1/2-1/2"]
[ECO "A02"]
[GameDuration "00:02:42"]
[Opening "Bird's Opening"]
[PlyCount "80"]
[Termination "adjudication"]
[TimeControl "60+1"]
1. f4 {book} Nf6 2. Nf3 {-0.30/24 4.1s} e6 3. Nc3 {-0.22/27 2.2s}
d5 {-0.42/30 3.6s} 4. d3 {-0.22/25 0.72s} b6 {-0.27/27 3.5s}
5. e4 {-0.24/25 0.92s} dxe4 {-0.13/30 3.8s} 6. dxe4 {-0.09/26 0.90s}
Qxd1+ {-0.34/29 3.1s} 7. Kxd1 {-0.06/26 1.2s} Bb7 {-0.19/27 1.2s}
8. Bd3 {-0.08/26 1.5s} Nc6 {-0.17/29 1.4s} 9. Ke2 {-0.16/27 3.0s}
O-O-O {-0.35/25 1.5s} 10. a3 {-0.12/25 0.89s} Re8 {-0.29/28 3.4s}
11. e5 {-0.22/30 5.4s} Nd7 {-0.34/25 0.88s} 12. Re1 {-0.13/24 0.76s}
h6 {-0.20/27 1.1s} 13. Kf1 {-0.20/24 1.2s} Rd8 {-0.24/49 10s}
14. h4 {-0.14/26 2.3s} Nc5 {-0.25/23 0.67s} 15. Be2 {-0.25/31 7.0s}
Be7 {-0.38/26 1.5s} 16. Rb1 {-0.29/24 1.9s} Nd7 {-0.38/25 5.0s}
17. Bd3 {-0.24/28 1.2s} Nc5 {-0.31/30 3.5s} 18. Be2 {-0.16/22 1.3s}
a5 {-0.17/24 1.4s} 19. g3 {-0.14/28 1.7s} Rhg8 {-0.29/29 5.2s}
20. Kf2 {-0.21/23 1.6s} f6 {-0.33/27 0.95s} 21. Bc4 {-0.12/28 1.7s}
a4 {-0.31/29 0.97s} 22. exf6 {-0.17/30 2.1s} Bxf6 {-0.24/27 1.1s}
23. Ne4 {0.00/27 1.1s} Bd4+ {-0.31/29 1.2s} 24. Kf1 {-0.08/31 1.4s}
Rge8 {-0.25/30 1.9s} 25. Nxc5 {-0.07/31 1.3s} Bxc5 {-0.19/31 1.0s}
26. Bxe6+ {-0.13/32 1.8s} Kb8 {-0.14/31 1.3s} 27. b4 {-0.07/32 2.9s}
axb3 {-0.09/32 1.2s} 28. Rxb3 {0.00/31 1.2s} Nd4 {0.00/38 4.9s}
29. Nxd4 {0.00/36 1.2s} Bxd4 {0.00/38 0.83s} 30. f5 {0.00/35 2.1s}
Ba6+ {-0.28/26 1.1s} 31. Kg2 {0.00/36 1.5s} Bc4 {-0.15/27 1.7s}
32. Bf4 {-0.08/34 4.8s} Bxb3 {-0.13/25 1.1s} 33. cxb3 {0.00/32 1.3s}
Bc5 {-0.21/25 1.2s} 34. Be5 {0.00/32 2.5s} Re7 {0.00/36 3.8s}
35. a4 {-0.06/32 1.4s} h5 {0.00/35 1.1s} 36. Bf4 {0.00/37 1.2s}
Ree8 {0.00/45 2.3s} 37. Rf1 {0.00/42 2.2s} Rf8 {0.00/42 1.1s}
38. Kh3 {0.00/45 1.4s} Kb7 {0.00/36 1.3s} 39. Rc1 {0.00/43 1.4s}
Bd6 {0.00/40 1.2s} 40. Bg5 {0.00/46 1.5s}
Ba3 {0.00/44 1.6s, Draw by adjudication} 1/2-1/2
So now,
Who believe "Eman 8.10" has evaluated and played "1. f4" ?
Who believe "1. f4" was played by the GUI ?
Which evaluated moves by "Depth4_050522 (Eman 8.10)" really represent the quality of the experience data of the Depth4_050522.exp file ?
Thanks for watching
