Zevra v2.6 (NNUE) released!

[sovaz1997]

Hello!
After a long break I release new version of Zevra with NNUE.
The net has size (768->256)x2->1 and fully trained from Zero.
First net has 768->8 size and trained on self-play games with random weights net.
Last net (11th gen) trained on 1 billion self-play games of previous net.


Changes
- NNUE eval (weights integrated to binary)
- UCI options for changing NNUE eval to HCE and Temperature for randomizing moves
- Small search improvements to make NNUE working correctly
- Remove hash probing in qsearch
- Remove Hash buckets, because didn't show improvements

Regression test (on Mac M2 Pro):

Code: Select all

tc=60+0.6, hash=256
Score of Zevra Self v11 vs Zevra Classic: 2363 - 1004 - 523  [0.675] 3890
...      Zevra Self v11 playing White: 1113 - 608 - 224  [0.630] 1945
...      Zevra Self v11 playing Black: 1250 - 396 - 299  [0.720] 1945
...      White vs Black: 1509 - 1858 - 523  [0.455] 3890
Elo difference: 126.7 +/- 10.7, LOS: 100.0 %, DrawRatio: 13.4 %

The performance on avx2 is better as I see, but don't have full test results for now.


Minimum requirements

- CPU: AVX-2 support. Intel Haswell or newer. Also Apple silicon architecture supported with using NEON instructions.

Download
https://github.com/sovaz1997/Zevra2/releases/tag/v2.6

[sovaz1997]

If something doesn't work, tell me please
I uploaded updated binaries for Windows, because it was not fully static. Now should be ok

[Rowen]

Thanks for sharing. I can see that temperature randomises moves but how exactly does it work? I have only seen it being used with regard to Lc0, and what is the best way of weakening the engine, to play against. Increasing temperature with or without HCE? Thanks

[sovaz1997]

Yeah, of course!

First: I generate offsets before search. Offsets is random values from 0 to temperature.

After when I run search, I run it with some bias, like

Code: Select all

 eval = -search(board, searchInfo, -beta + bias[moveNumber], -alpha + bonus[moveNumber], depth + 1) - bonus[moveNumber];

Bias is different on any move.

I did the temperature, because was need to generate more randomized games + also if I test with fixed nodes (I did it with NN), I wanted to have randomization, because otherwise I will have a lot of similar games.

[sovaz1997]

Ah yeah, one mistake
I trained on 1 billion positions, not 1 billion games of course

[JohnW]

I am not sure if I am doing something wrong but it keeps losing on time when the time control is 1 minute with 1 second increment.

[sovaz1997]

Where do you test, in cutechess or somewhere else? I'll take a look
In theory there shouldn't be any problems, is it specifically this tc (1 minute + 1 second) that's the problem?

[sovaz1997]

I do my tests on different, on 60+0.6 also and don't have timelosses. Maybe I can have time-losses, if I load more threads than I have
Maybe I need to count time differently in that case.

[JohnW]

Where do you test, in cutechess or somewhere else? I'll take a look
In theory there shouldn't be any problems, is it specifically this tc (1 minute + 1 second) that's the problem?

I am using Arena. I haven't tried other time controls

[JohnW]

I just tried it in the Fritz 19 software and it's working fine there.