Arasan 24.0 release

[jdart]

Arasan 24.0 is available from https://www.arasachess.org.

Changes in Arasan 24.0:
1) New network architecture based on "Stockfish version 4" structure
2) Cleanup shadow warnings + other misc. code cleanup.
3) Multiple fixes/updates to selfplay utility
4) Implement capture history and futility pruning for capture moves
5) Fix #50. Add an OwnBook option to CECP options.
6) Updates to test tools
7) Fixes to engine debug output, fix problem with Shredder UI

Arasan's network implementation is compatible with the SFNNv4 architecture used in Stockfish 15. Tuning was performed using a slightly modified version of the nnue-pytorch tuner (back-porting some later changes into the SFNNv4 tuning code). Initial weights were tuned based on 1.5B positions from Arasan selfplay games (generated from the "selfplay" utility program). This was followed by two rounds of reinforcement training, each time using about 3.6B selfplay positions. Total elapsed machine time was 7-8 weeks, using about 100 cores.

Fast blitz (2:0+1) results against Arasan 23.5:
2212 games, score: 61.66%%
ELO: 82.57 +-7.8 (95%)
LOS: 100.0%

[smatovic]

Total elapsed machine time was 7-8 weeks, using about 100 cores.

--
Srdja

[jdart]

I have also made some updates to the Arasan test suite, replacing some positions that were known to have multiple solutions, or appeared to have incorrect solutions, or multiple moves somewhat close in score. Latest version here: https://arasanchess.org/arasan2023.epd.

[peter]

I have also made some updates to the Arasan test suite, replacing some positions that were known to have multiple solutions, or appeared to have incorrect solutions, or multiple moves somewhat close in score. Latest version here: https://arasanchess.org/arasan2023.epd.

Thanks for the new version of the engine and of the test suite, I always liked this one too.

Here's one of the latest SF dev.- versions (20230807) single thread (3.5GHz, avx2- compile) with 1"/position, 2 extra plies, Shredder13 GUI:

Code: Select all

Bisher gelöst: 142 von 200  ;  1:20m


         1   2   3   4   5   6   7   8   9  10  11  12  13  14  15  16  17  18  19  20
 -------------------------------------------------------------------------------------
   0 |   0   0   -   0   0   0   -   0   0   0   0   0   0   0   0   0   -   0   0   -
  20 |   0   0   -   0   0   0   -   0   0   0   0   -   -   -   0   0   0   -   0   0
  40 |   -   -   0   -   -   0   0   0   -   0   0   0   -   0   0   0   -   0   -   0
  60 |   0   0   0   0   0   -   0   0   -   0   0   0   -   -   0   0   0   0   0   0
  80 |   0   0   0   0   0   0   0   0   0   0   -   0   0   0   0   -   0   0   -   0
 100 |   0   0   0   0   -   -   0   0   -   0   0   0   0   -   0   -   -   0   0   0
 120 |   -   -   0   -   0   0   0   0   0   0   0   -   -   0   -   0   0   0   0   0
 140 |   -   0   0   0   0   -   -   0   0   0   0   0   -   -   -   -   0   0   0   0
 160 |   0   0   0   -   0   -   0   0   -   0   -   0   -   -   0   -   0   -   -   0
 180 |   0   0   0   -   0   0   0   -   -   0   0   -   0   0   0   0   0   -   0   0

   1 sec -> 142/200
  K/s: 1.266.307  
  TotTime: 3:32m    SolTime: 1:20m

[peter]

Here's one of the latest SF dev.- versions (20230807) single thread (3.5GHz, avx2- compile) with 1"/position, 2 extra plies, Shredder13 GUI:

And here five engines with 3"/pos. single threaded (LC0 2 threads of CPU and 3070ti Nvidia GPU, net 4000M and SugaR had a special parameter- setting), again 2 extra plies, rated with EloStatTS (Frank Schubert):

Code: Select all

    Program                                    Elo   +/-  Matches  Score   Av.Op.   S.Pos.   MST1    MST2   RIndex

  1 SugaRXPrOIccf040823                      : 3526    7    696    54.8 %   3493   171/200    1.2s    1.4s   0.92
  2 Stockfishdev-20230807                    : 3516    7    677    52.8 %   3496   165/200    1.2s    1.5s   0.89
  3 Lc0v0.31.0-dag+git.9be9c142-4000M        : 3502    8    679    50.2 %   3500   152/200    1.2s    1.6s   0.82
  4 Dragon3.2byKomodoChess                   : 3496    8    669    49.2 %   3502   146/200    1.2s    1.7s   0.79
  5 Arasan24.0                               : 3457    9    647    42.4 %   3510   119/200    1.4s    2.0s   0.61



MST1  : Mean solution time (solved positions only)
MST2  : Mean solution time (solved and unsolved positions)
RIndex: Score according to solution time ranking for each position

[chesskobra]

How can I get it to work in xboard on linux? I get the error:

Failed to start first chess program polyglot -noini -ec "arasanx-64" "." -uci NalimovCache 4 -pg ShowTbHits=true on localhost: 1 +0 0 0 stringwarning opening book not found or invalid

I have the book path set correct.

Could you please help. Thanks.

[jdart]

How can I get it to work in xboard on linux? I get the error:

Failed to start first chess program polyglot -noini -ec "arasanx-64" "." -uci NalimovCache 4 -pg ShowTbHits=true on localhost: 1 +0 0 0 stringwarning opening book not found or invalid

First of all, you probably don't want to use the arasanx-64 executable unless you are on very old hardware. I recommend arasanx-64-avx2, or maybe arasanx-64-avx2-bmi2 or arasanx-64-avx512 if on a recent Intel system. Also I'm not clear why you're using Polyglot. Arasan natively supports both CECP (xboard) and UCI protocols, so there's no need for an adapter, and anyway xboard supports UCI engines now via the -fUCI option.

However, the opening book is assumed to be in the Arasan executable directory by default. If you have the book there, and haven't modified the arasan.rc file, there should be no problem loading it.

[Gabor Szots]

Jon, will you please explain this.

In the 24.0 package there is an NNUE file called arasan-d9-SFv4-20230728.nnue. However, Arasan 24 default NNUE file is arasan-d10-20220723.nnue.
Why is that? Which one is to be used?

[jdart]

Jon, will you please explain this.

In the 24.0 package there is an NNUE file called arasan-d9-SFv4-20230728.nnue. However, Arasan 24 default NNUE file is arasan-d10-20220723.nnue.
Why is that? Which one is to be used?

arasan-d9-SFv4-20230728.nnue is the most version that is supposed to be used. But I am wondering why you are saying " Arasan 24 default NNUE file is arasan-d10-20220723.nnue." I believe the only network in the distribution packages is arasan-d9-SFv4-20230728.nnue and that is what is set in the arasan.rc that is in the src directory and in the distribution packages and the installer.

[Gabor Szots]

I am wondering why you are saying " Arasan 24 default NNUE file is arasan-d10-20220723.nnue."

Because this is what I see after issuing 'uci' from the command line even if the only NNUE file in that directory is arasan-d9-SFv4-20230728.nnue.

Arasan 24.0 Copyright 1994-2023 by Jon Dart. All Rights Reserved.
uci
id name Arasan 24.0
id author Jon Dart
option name Hash type spin default 256 min 4 max 64000
option name Ponder type check default true
option name Contempt type spin default 0 min -200 max 200
option name Use tablebases type check default true
option name SyzygyTbPath type string default F:\Sakk\syzygy
option name SyzygyUse50MoveRule type check default true
option name SyzygyProbeDepth type spin default 4 min 0 max 64
option name MultiPV type spin default 1 min 1 max 10
option name OwnBook type check default false
option name Favor frequent book moves type spin default 50 min 0 max 100
option name Favor best book moves type spin default 50 min 0 max 100
option name Favor high-weighted book moves type spin default 100 min 0 max 100
option name Randomize book moves type spin default 50 min 0 max 100
option name Threads type spin default 1 min 1 max 256
option name UCI_LimitStrength type check default false
option name UCI_Elo type spin default 3300 min 1000 max 3300
option name Use NNUE type check default true
option name NNUE file type string default arasan-d10-20220226.nnue
option name Move overhead type spin default 30 min 0 max 1000
uciok

But in the meantime I have found that these were values taken from a previous arasan.rc which I have not updated.