Crafty-22.2 is available

[Peter Skinner]

I would like to state that 22.2 is _significantly_ better than the previous best version of 21.7.

In my own testing and confirmed by the way we internally test, 22.2 has become stronger than Fruit 2.1, and can we are certain that it can go much higher. Since moving to a "Fruit" like scoring structure, where material is evaluated in the middlegame and endgame phases (If you look at the source, you will see new mg/eg evaluation), we feel Crafty is moving in a very positive direction.

I have played over 1000 games with the latest versions on ICC and FICS, so I know it is very stable for ICS play. Robert and Tracy also play consistently on ICC using the accounts Crafty/Regal1 respectively. My account is obviously Somnus on ICC.

If you are part of a testing team and use the "own book" method of testing, please PM me here or email me and I can provide books that are better than the released versions on my website.

I would prefer (and I am sure that the rest of the team are in agreement) that you use our books in testing to gain the most accurate rating possible. I have made these books, and while not quite as good as Ted's books (our book maker), they do well on ICS and in local testing.

As always, if you have any issues, please feel free to email me and I can usually sort things out, including making specialized builds for those with older hardware.

Peter

[Peter Skinner]

Here is my rc file, and logging _is_ turned off:

mt=2
egtb
hash=192M
hashp=128M
cache=32M
log off
whisper 2
exit

There will still be a game.xxx produced, but no log.xxx file.

Peter

Do you still need to use smpnice=1 in conjunction with ponder=off for ponder off use , and what is the "whisper 2" command ?

.

Since I have never really used the smpnice=1 feature, possibly Robert can answer that for you. I always use ICS with two computers (both dual core) to play matches, or I just match an opponent on FICS/ICC for a few hundred games. We have Robert's cluster that we use for matches, and since that produces results much faster than I can attain them locally, I really only test the opening books that I/we make online.

The whisper 2 command is purely for ICS play where Crafty will "whisper" it's eval to observers. You can also use the kibitz command, but that can become an annoyance to human opponents.

Peter

[Zach Wegner]

Do you still need to use smpnice=1 in conjunction with ponder=off for ponder off use...

I sure hope not. Bob, if you are reading this, get rid of the way you handle this with fork() every move. Here's what I do, which took about 2 minutes to code:

Set up a wait_pipe[2] for each processor with pipe()
When you want the processors to stop, send them a message so they do a blocked read on the pipe
When going active, write a byte to the pipe from the master processor.

This has zero overhead (just a context switch), and is actually portable. I use this for Windows too, with a slightly modified pipe(). So there is no need to bother with an option anymore.

[Werner]

Hi Werner,

Can you tell us if this new version has the same Windows problems (using Wb2uci) as all the former Crafty22.x versions had?

Hi Ernest,
I do not remember which problems you mean with wb2uci adapter.
I had only some with smp under Arena. - But a short test here showed me: with ponder off the smp-engine (2CPU) uses no cpu time after the move now (that´s good).

best

[Werner]

I would like to state that 22.2 is _significantly_ better than the previous best version of 21.7.

In my own testing and confirmed by the way we internally test, 22.2 has become stronger than Fruit 2.1,

Whow, that´s a surprise: about 130 elos better than version 22.1

and here an example from my running tournament:

[Event "Bug-1"]
[Site "AMD64-X2"]
[Date "2008.11.16"]
[Round "2"]
[White "Crafty 22.2 x64 1CPU"]
[Black "BugChess2 v.1.6.3"]
[Result "1-0"]
[BlackElo "2200"]
[Time "21:32:29"]
[WhiteElo "2200"]
[TimeControl "40/900:40/900:40/900"]
[Termination "adjudication"]
[PlyCount "58"]
[WhiteType "program"]
[BlackType "program"]

1. e4 c5 2. Nf3 d6 3. d4 cxd4 4. Nxd4 Nf6 5. Nc3 a6 6. Be3 e6 7. f3 b5 8.
g4 h6 9. Qd2 Nbd7 10. O-O-O Bb7 11. a3 {+0.15/13 17} d5 {0.00/13 20} 12.
exd5 {+0.20/15 17} Nxd5 {0.00/13 30} 13. Nxd5 {+0.18/14 20} Bxd5 {-0.04/13
19} 14. h4 {+0.07/13 17} e5 {+0.11/13 30} 15. Re1 {+0.22/12 18} Rc8
{-0.15/13 76} 16. Bd3 {+0.61/13 17} Nc5 {-0.24/13 33} 17. Bf4 {-0.13/12 19}
f6 {+0.03/13 28} 18. Bxe5 {+0.23/13 17} fxe5 {0.00/13 22} 19. Rxe5+
{+0.01/11 18} Kf7 {0.00/14 40} 20. Bf5 {+0.01/12 18} Rc7 {0.00/13 26} 21.
b4 {+0.01/12 18} Bxf3 {0.00/13 28} 22. bxc5 {+0.01/12 23} Bxh1 {-3.89/13
151} 23. Qd3 {+1.53/12 18} Kf6 {-4.33/11 129} 24. Re1 {+6.30/11 18} g5
{-7.34/11 25} 25. Bh7 {+9.91/12 37} Rhxh7 {-7.91/12 19} 26. hxg5+
{+10.82/12 18} hxg5 {-7.94/13 5} 27. Rf1+ {+12.61/14 18} Kg7 {-12.26/14 24}
28. Ne6+ {+15.72/15 18} Kg8 {-15.81/14 85} 29. Qxd8 {+15.89/15 21} Rhf7
{-19.77/13 20} 30. ... {1-0 Arena Entscheidung} 1-0

[Peter Skinner]

I would like to state that 22.2 is _significantly_ better than the previous best version of 21.7.

In my own testing and confirmed by the way we internally test, 22.2 has become stronger than Fruit 2.1,

Whow, that´s a surprise: about 130 elos better than version 22.1

22.2 changed quite significantly over 22.1.

I am very confident in my statement that Crafty has become stronger than Fruit 2.1 due to the way we internally test and my own results.

Then again setups can change anything.... but I am still confident

Peter

[Spock]

But a short test here showed me: with ponder off the smp-engine (2CPU) uses no cpu time after the move now (that´s good).

best

So smpnice=1 is no longer necessary. Very good. I'll try it myself later

[glorfindel]

It is good news that crafty is getting stronger, and fortunately with 22.2 I can take back moves in analysis mode with 4 threads running without a crash and with smpnice=1.

My problem is now in positions from the endgame TBs:
1. In analysis mode (in XBoard) I can see no output about the best move or the score. I only see depth 0 score +0.00.
2. If I play out a position from the endgame TBs against crafty and I take back a move, crafty crashes immediately like this:

Code: Select all

79230 >first : remove
xboard: Error: first chess program (/usr/games/bin/crafty) exited unexpectedly
GameEnds(0, (null), 2)

I managed to crash Crafty 2 times out of 2 in a K+2P v K+P ending, 2 times out of 2 in a K+B+P v K+N ending, but in a K+R+B v K+R ending it refused to crash and worked normally.

I used the Makefile provided, target linux-amd64, only with NUMA disabled.

[mike860]

that's the source, but where is the engine ?

I just compiled Win32 and Win64 exe's on my website:

http://www.webkikr.net

Thanks, Peter.

[bob]

Do you still need to use smpnice=1 in conjunction with ponder=off for ponder off use...

I sure hope not. Bob, if you are reading this, get rid of the way you handle this with fork() every move. Here's what I do, which took about 2 minutes to code:

Set up a wait_pipe[2] for each processor with pipe()
When you want the processors to stop, send them a message so they do a blocked read on the pipe
When going active, write a byte to the pipe from the master processor.

This has zero overhead (just a context switch), and is actually portable. I use this for Windows too, with a slightly modified pipe(). So there is no need to bother with an option anymore.

with threads, there is effectively zero overhead anyway. Doing anything to make the threads "wait" is painful because there is no single solution that is portable across every platform.

If you use linux, you can now create about 150,000 threads per minute on a fairly slow box... Creating 3 or 7 takes nothing...