Arasan 23.0

[jdart]

Arasan 23.0 is released and is avaialable from http://www.arasanchess.org.

Changes in Arasan 23.0:
1) Add NNUE evaluation
2) Changes to Makefile: use BUILD_TYPE to select instruction set(s)
3) Code cleanup: put globals into namespace. Remove TbType because we only support Syzygy tablebases now.
4) Code cleanup: qualify imports from std namespace (remove "using namespace std").
5) Code cleanup: remove ASSERT macro, replace with standard assert.
6) Fixes/updates to selfplay utility for NN position generation.
7) Fix unit tests to account for score variance in draw positions.
8) Fix for loading .rc file, book and network when program executable is found via PATH traversal.

Score of Arasan-23.0 vs Arasan-22.3: 238 - 27 - 135 [0.764] 400
Elo difference: 203.83 +/- 29.33
(your mileage may vary).

Some other notes:
- The Windows installer that includes the Arasan GUI does not ship with a network and has NNUE disabled by default, because it is a 32-bit executable and NNUE is not likely to perform well in that environment.
- Only the AVX2 builds have SIMD support for the NNUE eval, so non-AVX2 builds will not perform nearly as well with NNUE enabled.
- If building from source please see the build instructions at https://github.com/jdart1/arasan-chess/ ... c/BUILD.md. Note in particular: 1) C++17 is required now, 2) shallow "clone" + "make" will not work: you need to fetch the submodules, and 3) "make" with no BUILD_TYPE specified will build a very "generic" executable with no special x86 instruction set support.
- Mac M1 is theoretically supported but I have done no compilation/testing on that platform: the supplied Mac binaries are for x86-64.

[Rebel]

Games : 1000

Elo-pool : 3249

http://rebel13.nl/a/grl.htm

[connor_mcmonigle]

Games : 1000

Elo-pool : 3249

http://rebel13.nl/a/grl.htm

I would guess you've not set the set the "NNUE file" option given the poor performance you're observing.

[Rebel]

Indeed, something is definitely wrong.

When you start arasan from the command line, uci -> go depth 5 -> it says the NNUE file is loaded and so I ran it without setting the NNUE option. When I saw the bad results (similar to the current) I added the NNUE option and went to bed. Still not good and so I aborted the tournament.

I noticed that cutechess complained alot with : Warning: EngineProcess: Destroyed while process is still running.

[Gabor Szots]

For me AVX2 is almost 50 % faster than BMI2.

[Werner]

For me AVX2 is almost 50 % faster than BMI2.

Hi, same here: there must be something wrong with the bmi2 compile as here normally avx2 is included:
startpos bmi2 120k nps
startpos avx2 820k nps

[Rebel]

The popcount version only gets 84K nps.

Switching off NNUE the NPS is 806K.

[jswaff]

Score of Arasan-23.0 vs Arasan-22.3: 238 - 27 - 135 [0.764] 400
Elo difference: 203.83 +/- 29.33
(your mileage may vary).

That is seriously impressive, especially for an engine that is already very strong. Congrats Jon!

[dkappe]

Any details about the net: data, training, architecture, etc?

[jdart]

When you start arasan from the command line, uci -> go depth 5 -> it says the NNUE file is loaded and so I ran it without setting the NNUE option. When I saw the bad results (similar to the current) I added the NNUE option and went to bed. Still not good and so I aborted the tournament.

Not clear what is going on there but I notice when loading it as a UCI engine in Fritz the network name is blank in the UCI options - so w/o setting that it may not be loading the network.