cheng4 0.38 release

[Frank Quisinsky]

Hi Martin,

thanks!!
With delay (a big tourney is still running on my site) I will test it around middle of February.

Good news again!

Best
Frank

[petero2]

Hi all,
new cheng4 is available here http://vlasak.biz/cheng as usual.
Most of the gain (I expect some 30 real elo, maybe more) is due to eval tuning using "Texel tuning method" by Peter Osterlund (thanks).
Also thanks to Graham for providing book lines, Lars for testing and Emil for hosting.

Congratulations!

It will be interesting to see how strong this version is on various rating lists.

[mar]

Congratulations!

It will be interesting to see how strong this version is on various rating lists.

Thanks Peter.
Yes I guess we'll know soon.
iirc I measured some 75 elo in selfplay bullet in 3 steps tuning using your method.
EDIT: meaning 25 + 28 + 22 (first attempt, then all params, then retuning)

[petero2]

Congratulations!

It will be interesting to see how strong this version is on various rating lists.

Thanks Peter.
Yes I guess we'll know soon.
iirc I measured some 75 elo in selfplay bullet in 3 steps tuning using your method.
EDIT: meaning 25 + 28 + 22 (first attempt, then all params, then retuning)

For texel 1.03 I measured +112 elo against texel 1.02 at hyper bullet time control. Surprisingly in the grand unified rating list from 2014-11-11 (bayeselo version) the rating difference is 113 elo.

It is possible that different compiler optimizations is responsible for some of this, but I suspect that when testing at hyper-bullet time control, evaluation changes tend to scale better to longer time controls than search changes.

So I hope to see more than 30 elo increase in this version. As I said, it will be interesting to see what happens.

[mar]

So I hope to see more than 30 elo increase in this version. As I said, it will be interesting to see what happens.

So do I But I never got more than half of selfplay so far so I'm trying to use conservative estimates.

[Dr.Wael Deeb]

So I hope to see more than 30 elo increase in this version. As I said, it will be interesting to see what happens.

So do I But I never got more than half of selfplay so far so I'm trying to use conservative estimates.

congratulations on the new release Martin....

I have Cheng4 0.36c at 2773.4 Elo in my private rating list after playing 4420 games at a time control of 40 minutes + 20 seconds increment....

Reference chess engine is Deep Fritz 13 3068 Elo....
Dr.D

[Joerg Oster]

Here are 2 FRC games from Cheng4 0.38 against Stockfish, played with cutechess-cli under Linux. It seems Cheng sometimes has a problem after castling ...

Code: Select all

[Event "?"]
[Site "?"]
[Date "2015.01.18"]
[Round "1"]
[White "SF-Master"]
[Black "Cheng4"]
[Result "1-0"]
[FEN "qrbknnrb/pppppppp/8/8/8/8/PPPPPPPP/QRBKNNRB w GBgb -"]
[PlyCount "19"]
[SetUp "1"]
[TimeControl "8+0.08"]
[Variant "fischerandom"]

1. g3 {+0.10/14 0.43s} c6 {-0.16/13 0.29s} 2. b3 {+0.34/15 0.38s}
g6 {-0.05/13 0.35s} 3. Bb2 {+0.35/16 0.42s} Bxb2 {+0.07/14 0.26s}
4. Qxb2 {+0.27/17 0.34s} a5 {-0.08/13 0.37s} 5. e4 {+0.44/14 0.24s}
d6 {-0.03/12 0.27s} 6. Ne3 {+0.48/14 0.22s} Nf6 {-0.15/12 0.34s}
7. Nd3 {+0.52/14 0.42s} Ne6 {-0.10/13 0.33s} 8. e5 {+0.51/15 0.48s}
dxe5 {-0.18/14 0.32s} 9. Nxe5 {+0.55/15 0.43s} O-O {-0.15/13 0.32s}
10. O-O {+0.55/14 0.13s, Black makes an illegal move: d8g8} 1-0

Code: Select all

[Event "?"]
[Site "?"]
[Date "2015.01.18"]
[Round "22"]
[White "Cheng4"]
[Black "SF-Master"]
[Result "0-1"]
[FEN "rnnkrqbb/pppppppp/8/8/8/8/PPPPPPPP/RNNKRQBB w EAea -"]
[PlyCount "80"]
[SetUp "1"]
[TimeControl "8+0.08"]
[Variant "fischerandom"]

1. Nc3 {+0.20/13 0.30s} g5 {-0.12/14 0.36s} 2. g4 {+0.31/12 0.61s}
Nc6 {-0.20/13 0.25s} 3. Qh3 {+0.15/12 0.37s} Nd6 {+0.06/14 0.58s}
4. Nb3 {+0.29/12 0.28s} O-O-O {+0.24/12 0.46s} 5. O-O-O {+0.19/11 0.26s}
f5 {+0.31/11 0.14s} 6. d3 {+0.05/10 0.28s} Bxc3 {+1.14/13 0.34s}
7. bxc3 {-0.21/13 0.33s} Nb5 {+1.26/13 0.23s} 8. d4 {-0.07/12 0.22s}
fxg4 {+1.27/13 0.19s} 9. Qd3 {0.00/13 0.25s} Na3 {+1.25/12 0.30s}
10. Be4 {0.00/11 0.20s} d5 {+2.46/13 0.44s} 11. Bxh7 {-0.34/13 0.41s}
Bxh7 {+2.84/14 0.28s} 12. Qxh7 {-0.60/14 0.28s} e5 {+2.66/14 0.24s}
13. e3 {-0.70/11 0.25s} Nc4 {+4.44/14 0.38s} 14. Nc5 {-2.34/12 0.31s}
b6 {+4.53/14 0.23s} 15. Qg6 {-2.38/12 0.24s} N6a5 {+4.99/15 0.67s}
16. Qxg5 {-2.78/11 0.26s} bxc5 {+5.22/15 0.30s} 17. Qxg4+ {-3.19/11 0.19s}
Kb7 {+6.53/17 0.29s} 18. dxe5 {-4.34/12 0.35s} Qh6 {+6.76/19 0.29s}
19. Rd2 {-4.58/12 0.16s} Qb6 {+8.32/14 0.15s} 20. Kd1 {-4.70/12 0.35s}
Qb2 {+8.72/14 0.20s} 21. f4 {-5.48/11 0.37s} Rg8 {+11.03/15 0.22s}
22. Qf3 {-5.92/12 0.21s} Nxd2 {+12.80/17 0.15s} 23. Kxd2 {-11.42/13 0.29s}
Nc4+ {+14.29/17 0.22s} 24. Kd3 {-12.14/15 0.24s} Na3 {+14.60/18 0.17s}
25. Ke2 {-12.37/14 0.19s} Nxc2 {+14.80/18 0.16s} 26. Qh1 {-12.72/14 0.19s}
Nxe1+ {+15.28/17 0.24s} 27. Kxe1 {-13.30/15 0.16s} Rg2 {+15.51/16 0.11s}
28. Qxg2 {-14.87/15 0.11s} Qxg2 {+16.15/16 0.24s} 29. Bf2 {-16.70/14 0.21s}
Qxh2 {+16.54/15 0.16s} 30. e6 {-16.17/13 0.21s} Re8 {+17.37/16 0.34s}
31. f5 {-17.62/13 0.31s} Qe5 {+18.00/15 0.22s} 32. e7 {-17.87/12 0.47s}
Qxc3+ {+48.09/16 0.15s} 33. Kd1 {-143.36/13 0.14s} Qd3+ {+M21/20 0.24s}
34. Ke1 {-296.97/13 0.38s} Rxe7 {+M19/20 0.10s} 35. f6 {-M81/12 0.10s}
Rh7 {+M11/23 0.11s} 36. Bg1 {-M10/9 0.090s} Rh1 {+M9/31 0.10s}
37. Kf2 {-M8/31 0.13s} Qd2+ {+M7/42 0.10s} 38. Kf3 {-M6/120 0.042s}
Rxg1 {+M5/127 0.055s} 39. e4 {-M4/120 0.006s} d4 {+M3/127 0.005s}
40. e5 {-M2/120 0.002s} Qe3# {+M1/127 0.004s, Black mates} 0-1

[mar]

Here are 2 FRC games from Cheng4 0.38 against Stockfish, played with cutechess-cli under Linux. It seems Cheng sometimes has a problem after castling ...

Thanks Joerg. I can confirm there's a problem when rook is on G or C file because uci move parser will set incorrect capture flag and cheng refuses the move from GUI as illegal.
The fix was trivial, I wonder if I introduced it somewhere after 0.36a:
Just replace this

Code: Select all

res = MovePack::initCastling( kp, to );

with this

Code: Select all

return MovePack::initCastling( kp, to );

in board.cpp line 1337.

I admit I haven't tested FRC since then.
Oh well. I hate when this happens
So what now, I can prepare a new package but I don't want to set version to 0.38a or something.
So I will either keep the version number as is or simply pretend it's ok and accept losses in such positions

[mar]

congratulations on the new release Martin....

I have Cheng4 0.36c at 2773.4 Elo in my private rating list after playing 4420 games at a time control of 40 minutes + 20 seconds increment....

Reference chess engine is Deep Fritz 13 3068 Elo....
Dr.D

Thanks doc,
would be nice to see your results as well.

[Sylwy]

To be short:

SKIING IS A PLEASURE !



Your Carpathian connexion,
SilvianR

PS: now, this logo is your property !