DoubleCheck 2.3 is out

[lucasart]

If anyone is interested in a 2300 elo engine these days
http://wbec-ridderkerk.nl/html/download.htm

As usual, source code available and licensed under GNU GPL v3. It is still portable and should compile on any plateform (Mac OSX, and generally any UNIX compatible system, also Google Android). The official compiles are 64-bit only, for both Linux and Windows, but Jim Ablett kindly provided other compiles.

If you're not sure which compile is fastest on your computer, or want to be certain that the compile you have is correct, run the program from a command line with an argument (whichever argument you want doesn't matter, so long ar there is at least one)
* It will run a benchmark test, searching several positions in a row at given depth, and return a time and node count
* The node count is a signature that ensures that the compile is correct. The correct node count should be 5784711.
* The time of course depends on your computer. If you want to know, mine does it in about 14.1 seconds.

In terms of changes, there hasn't been very much. I mostly tweaked search reductions and search extensions, as well as introduced a rather "uncommon" way of doing eval pruning and razoring.

I'd be happy if any of the rating lists could test it, because I don't have any precise idea of its level. This is because I almost only do self testing at very fast time control. And this tends to overestimate any improvement.

Enjoy!
Lucas

[Graham Banks]

If anyone is interested in a 2300 elo engine these days
http://wbec-ridderkerk.nl/html/download.htm.........

Of course there will be those interested!
Thanks Lucas. Below are some ratings for various versions of DoubleCheck from the CCRL 40/40 list.

40/40

Code: Select all

Rank                Engine                 Elo   +    -   Score  AvOp  Games
1 DoubleCheck 2.1 32-bit               2186  +41  -41  52.3%  -18.1   219
  DoubleCheck 1.3.1 64-bit             2185  +40  -40  47.4%  +21.4   234
  DoubleCheck 1.0 64-bit               1949  +44  -46  23.0% +222.6   239

[zullil]

As usual, source code available and licensed under GNU GPL v3. It is still portable and should compile on any plateform (Mac OSX, and generally any UNIX compatible system, also Google Android).
Enjoy!
Lucas

Compiled on OSX by invoking the following from the command line, after making /DoubleCheck_23/src the working directory:

Code: Select all

gcc -O3 -DPOSIX -I/./*.h ./*.c -std=c99 -o DoubleCheck2.3

[lucasart]

As usual, source code available and licensed under GNU GPL v3. It is still portable and should compile on any plateform (Mac OSX, and generally any UNIX compatible system, also Google Android).
Enjoy!
Lucas

Compiled on OSX by invoking the following from the command line, after making /DoubleCheck_23/src the working directory:

Code: Select all

gcc -O3 -DPOSIX -I/./*.h ./*.c -std=c99 -o DoubleCheck2.3

Good news
For a fast executable you also need to put a -DNDEBUG (removes all the assert code that slows down massively), and you can add an -flto (link time optimization)

[lucasart]

If anyone is interested in a 2300 elo engine these days
http://wbec-ridderkerk.nl/html/download.htm.........

Of course there will be those interested!
Thanks Lucas. Below are some ratings for various versions of DoubleCheck from the CCRL 40/40 list.

40/40

Code: Select all

Rank                Engine                 Elo   +    -   Score  AvOp  Games
1 DoubleCheck 2.1 32-bit               2186  +41  -41  52.3%  -18.1   219
  DoubleCheck 1.3.1 64-bit             2185  +40  -40  47.4%  +21.4   234
  DoubleCheck 1.0 64-bit               1949  +44  -46  23.0% +222.6   239

Thanks!
I ran some blitz test in 1min+1sec/move and got the following results:
* DC 2.3 64bit vs DC 1.3.1 64bit: +160 elo after 200 games
* DC 2.3 64bit vs Jazz 5.01 64bit JA: +0 elo after 200 games (score was exactly equal!)
So I can reasonably hope that DC 2.3 64bit will break the 2300 elo "barrier" on CCRL. Although I'm curious to see the results against a varied population of engines.
On a technical level there is no workaround in the code to make the 32bit fast, so you should find the 32-bit versions much slower than the 64-bit ones.

[lucasart]

I did some more testing

* DC 2.3 64bit vs Jazz 5.01 64bit: 200 games
* DC 2.3 64bit vs Sungorus 1.4 64bit: 200 games
* Sungorus 1.4 64bit vs Jazz 5.01 64bit: 100 games

Bayeselo rating table

Code: Select all

Rank Name                     Elo    +    - games score oppo. draws 
   1 Sungorus 1.4               6   23   23   300   51%    -3   21% 
   2 DoubleCheck 2.3           -3   20   20   400   50%     1   22% 
   3 Jazz v5.01 JA (x86_64)    -3   23   23   300   50%     0   26% 

So the three engines are fairly equal.

PS: all games are played in 1min+1sec/move, on my laptop. It's a recent laptop, quite fast: it runs the DoubleCheck benchmark test in about 7 sec.

[Graham Banks]

I did some more testing

* DC 2.3 64bit vs Jazz 5.01 64bit: 200 games
* DC 2.3 64bit vs Sungorus 1.4 64bit: 200 games
* Sungorus 1.4 64bit vs Jazz 5.01 64bit: 100 games

Bayeselo rating table

Code: Select all

Rank Name                     Elo    +    - games score oppo. draws 
   1 Sungorus 1.4               6   23   23   300   51%    -3   21% 
   2 DoubleCheck 2.3           -3   20   20   400   50%     1   22% 
   3 Jazz v5.01 JA (x86_64)    -3   23   23   300   50%     0   26% 

So the three engines are fairly equal.

PS: all games are played in 1min+1sec/move, on my laptop. It's a recent laptop, quite fast: it runs the DoubleCheck benchmark test in about 7 sec.

Looks promising.

[lucasart]

I did some more testing

* DC 2.3 64bit vs Jazz 5.01 64bit: 200 games
* DC 2.3 64bit vs Sungorus 1.4 64bit: 200 games
* Sungorus 1.4 64bit vs Jazz 5.01 64bit: 100 games

Bayeselo rating table

Code: Select all

Rank Name                     Elo    +    - games score oppo. draws 
   1 Sungorus 1.4               6   23   23   300   51%    -3   21% 
   2 DoubleCheck 2.3           -3   20   20   400   50%     1   22% 
   3 Jazz v5.01 JA (x86_64)    -3   23   23   300   50%     0   26% 

So the three engines are fairly equal.

PS: all games are played in 1min+1sec/move, on my laptop. It's a recent laptop, quite fast: it runs the DoubleCheck benchmark test in about 7 sec.

Looks promising.

Added Pawny 0.3.1 to the list. I've just downloaded and compiled BayesElo, and learning about it, fascinating stuff! I'm starting to do a little rating list actually. Testing conditions are:
* performance.bin by Marc Lacrosse, limited to depth 10 (20 half moves)
* 1min+1sec/move: longer time control = better games. But for a given amount of CPU time, it's certainly better to play 10 times more games than 10 times longer games. There's little evidence supporting the idea that longer games give more accurate ratings. On the other hand ridiculous time controls (line 10sec+0.1sec/move or so) may lead to many time losses or engines having issues finishing the shallowest depth in some situations. 1+1 seems a good compromise!
* no resign or adjudication. These weak programs are far to buggy for this feature to be trustworthy (they might not be able to win a dead won endgame, or give large scores when it's a draw etc.)
* Only engines that have a Linux version, that actually *works* (XBoard or UCI). And only open source ones. This is a good way to drastically reduce the list !
* Only 64 bit engines: we're almost in 2012 for god sake, who uses a 32 bit CPU these days ? (except for these windows users with 32bit Windows XP running on a 64 bit machine...)
* And, of course, elo are computed by BayesElo. As a statistician myself, I don't believe in the EloStat method.
* ELO are meaningless as such, only an ELO difference means anything. So an arbitrary offset is put, but really it doesn't matter. For now, I just put Sungorus 1.4 at 2402 ELO, but it may change depending on my mood

Code: Select all

Rank Name                     Elo    +    - games score oppo. draws 
   1 Pawny 0.3.1 (x64)       2518   34   33   200   66%  2397   17% 
   2 Sungorus 1.4            2402   24   24   350   48%  2413   20% 
   3 Jazz v5.01 JA (x86_64)  2400   24   24   350   48%  2414   25% 
   4 DoubleCheck 2.3         2394   20   20   500   46%  2424   21% 

Running Beowulf at the moment: 100 games against DC 2.3 50 games against other engines.

Lucas

[abik]

If anyone is interested in a 2300 elo engine these days

Congrats on the stronger new release. Just like Jim, I also compiled your engine for ARM, and put the binary on UCI and XBoard engines for Andrdoid.

Results of a quick 1-second per move tournament on a Nexus S, played from both sides of all lines in the Noomen test suite 2012 (games available on request):

Code: Select all

                     1         2          3                                                            
1   DoubleCheck 2.3      *     36.5-23.5  39.5-20.5  76.0/120
2   DoubleCheck 2.0  23.5-36.5     *      35.5-24.5  59.0/120
3   DoubleCheck 1.3  20.5-39.5 24.5-35.5      *      45.0/120

[ZirconiumX]

Just get a few warnings on OSX with an ancient gcc (4.0.1):

Code: Select all

$ gcc -O3 -DPOSIX -DNDEBUG -I/./*.h ./*.c -std=c99 -o doublecheck
./search.c:49: warning: 'adjust_tt_score' declared inline after being called
./search.c:49: warning: previous declaration of 'adjust_tt_score' was here
./search.c:48: warning: 'can_return_tt' declared inline after being called
./search.c:48: warning: previous declaration of 'can_return_tt' was here
./search.c:47: warning: 'update_killers' declared inline after being called
./search.c:47: warning: previous declaration of 'update_killers' was here

Nothing major I don't think.

Also, do I dare see someone breaking the UCI rules?!

Move format:
------------

The move format is in long algebraic notation.
A nullmove from the Engine to the GUI should be send as 0000.
Examples: e2e4, e7e5, e1g1 (white short castling), e7e8q (for promotion)

info score cp 83 depth 11 nodes 246624 time 1720 pv Rc5d5 Rd1xd5 e6xd5 Nc4b2 d5d4 Qa1c1 d4xe3 f2xe3 Bb7e4

Tut tut, you should be ashamed of yourself.

Seriously though, brilliant engine, it is lightning fast!

Matthew:out