Arasan 23.3

[jdart]

Arasan 23.3 is available from https://www.arasanchess.org.

Changes in Arasan 23.3:
1) Convert from using a separate polling thread to polling from the
search itself, as was done earlier in Arasan development.
2) Fix build problem with utility programs
3) Fixes for errors and warnings in build of tuner.
4) Some code cleanup. Remove spinlocks, replace with std::mutex.
5) Change how extensions/reductions are handled.
6) Use regex to parse .rc file. Tolerate/ignore blank lines in the file.
7) Shift to using pure NNUE rather than mixed classical/NNUE eval in imbalanced positions.
8) Changes to selfplay utility: -n sets position count not game
count. Add -a flag to append to existing file.
9) Add AVX512 support
10) New network
11) Changes/tuning to strength reduction feature.

Match results, Windows AVX2 build, 1:0+0.6 time control:

Score of Arasan23.3 vs Arasan23.2: 528 - 302 - 1170 [0.556] 2000
... Arasan23.3 playing White: 305 - 124 - 571 [0.591] 1000
... Arasan23.3 playing Black: 223 - 178 - 599 [0.522] 1000
... White vs Black: 483 - 347 - 1170 [0.534] 2000
Elo difference: 39.4 +/- 9.8, LOS: 100.0 %, DrawRatio: 58.5 %

[Gabor Szots]

Thanks Jon.

A note: the avx2 and the modern compile seems a bit slower than the simple 64-bit one, which surprises me. I have a Haswell CPU.

[Modern Times]

Thanks Jon.

A note: the avx2 and the modern compile seems a bit slower than the simple 64-bit one, which surprises me. I have a Haswell CPU.

Odd. The AVX2 one is very much faster than the standard x64 compile on my Ryzen laptop

[Gabor Szots]

Thanks Jon.

A note: the avx2 and the modern compile seems a bit slower than the simple 64-bit one, which surprises me. I have a Haswell CPU.

Odd. The AVX2 one is very much faster than the standard x64 compile on my Ryzen laptop

My bad. I used them from the command line without setting a net. Now the avx2 compile is indeed much faster than the other two.

[Modern Times]

I just double-clicked the exe, typed "bench" and it displayed text that the net had been loaded and then it ran the benchmark.

[pohl4711]

Arasan 23.3 is available from https://www.arasanchess.org.

Changes in Arasan 23.3:
1) Convert from using a separate polling thread to polling from the
search itself, as was done earlier in Arasan development.
2) Fix build problem with utility programs
3) Fixes for errors and warnings in build of tuner.
4) Some code cleanup. Remove spinlocks, replace with std::mutex.
5) Change how extensions/reductions are handled.
6) Use regex to parse .rc file. Tolerate/ignore blank lines in the file.
7) Shift to using pure NNUE rather than mixed classical/NNUE eval in imbalanced positions.
8) Changes to selfplay utility: -n sets position count not game
count. Add -a flag to append to existing file.
9) Add AVX512 support
10) New network
11) Changes/tuning to strength reduction feature.

Match results, Windows AVX2 build, 1:0+0.6 time control:

Score of Arasan23.3 vs Arasan23.2: 528 - 302 - 1170 [0.556] 2000
... Arasan23.3 playing White: 305 - 124 - 571 [0.591] 1000
... Arasan23.3 playing Black: 223 - 178 - 599 [0.522] 1000
... White vs Black: 483 - 347 - 1170 [0.534] 2000
Elo difference: 39.4 +/- 9.8, LOS: 100.0 %, DrawRatio: 58.5 %

And still the annoying warning in cutechess-cli. Engine process destroyed while engine is still running.

[Chessqueen]

Arasan 23.3 is available from https://www.arasanchess.org.

Changes in Arasan 23.3:
1) Convert from using a separate polling thread to polling from the
search itself, as was done earlier in Arasan development.
2) Fix build problem with utility programs
3) Fixes for errors and warnings in build of tuner.
4) Some code cleanup. Remove spinlocks, replace with std::mutex.
5) Change how extensions/reductions are handled.
6) Use regex to parse .rc file. Tolerate/ignore blank lines in the file.
7) Shift to using pure NNUE rather than mixed classical/NNUE eval in imbalanced positions.
8) Changes to selfplay utility: -n sets position count not game
count. Add -a flag to append to existing file.
9) Add AVX512 support
10) New network
11) Changes/tuning to strength reduction feature.

Match results, Windows AVX2 build, 1:0+0.6 time control:

Score of Arasan23.3 vs Arasan23.2: 528 - 302 - 1170 [0.556] 2000
... Arasan23.3 playing White: 305 - 124 - 571 [0.591] 1000
... Arasan23.3 playing Black: 223 - 178 - 599 [0.522] 1000
... White vs Black: 483 - 347 - 1170 [0.534] 2000
Elo difference: 39.4 +/- 9.8, LOS: 100.0 %, DrawRatio: 58.5 %

You posted Elo difference, but NOT estimated Final Elo with this version and New Network?

[Gabor Szots]

I just double-clicked the exe, typed "bench" and it displayed text that the net had been loaded and then it ran the benchmark.

I did not know that method. Anyway, I had an older arasan.rc in use in which the network had not been set. You can't use the same rc file for different versions, if you want them to use different networks you have to install them in different folders.

[jdart]

[quote
You posted Elo difference, but NOT estimated Final Elo with this version and New Network?

You only get relative ELO with a selfplay test. And any absolute ELO depends on the rating pool. I expect this will get tested by CCRL and/or CEGT eventually and then you will get a number that is conformant with their pool, testing and rating conditions.

[Wolfgang]

https://cegt.forumieren.com/t1706-testi ... n-x64-1cpu