NNUE

[Sidiki]

Hi Bert,

This it's already a great progresse.

Thanks

[Ed Gilbert]

Nice result Bert!
I expect we will see a new version of Damage soon.

-- Ed

[BertTuyt]

Sidiki/Ed thanks.

Yes i want to participate in the next tournament with Damage 16.0 which will be a HUB version and based upon NNUE.

Some expectations/remarks:
* Im optimistic that with further SW progress and better HW, NNUE and the current pattern based eval will be on par.
* Both Intel and AMD will further improve NN HW, and like bitboards (which required 64bit processors), and previous instruction-set improvements (popcnt, pext), the new wave of processor will focus on NN HW acceleration.
* This will improve one of the major drawbacks of NNUE (compared with patterns) the computation efforts, and the reduced search speed and search depth.
* NNUE is very elegant as it uses zero domain knowledge for the network. In my case the input is based upon 191 binary board positions (2 * 45 white/black man, 2 * 50 white/black king, and color to move), and 4 NN layers (input 256, hidden, 32 and 32, output 1).
* As NN and NNUE is very generic, one will see an explosion of development tools for training weights, inference libraries, reinforced learning, .....
* I expect that with the recent introduction of the intel processor i9-11900K one could be on par with 1 min games (against both Scan and KR).
* I will order such a machine, however due to component shortage this might take some time.
* Unfortunately most current processors have not implemented avx-512, VNNI, and DL Boost, so i cant share the most recent versions (for example these instruction sets will only be implemented in the ZEN 4, which will be released late 2022).
* As we might be close to the draw black-hole, i don't expect a huge potential increase of ELO. If in the most recent match (3792 games), Scan nnue would have lost nothing, and the win remains at 27 (so 3765 draws), then the resulting Elo difference would be a minor 2 points

Bert

[BertTuyt]

Most recent match Kingsrow 1.62 - Scan_31nnue.
Conditions Scan_31nnue vs Scan_31, played on an Intel Core i7 8700K, 2-move start positions, TC 65 moves in 1 minute, books off, 6-piece dbs, 6 search thread.

Code: Select all

Kingsrow 1.62 vs Scan_nnue 3.1:  1 wins, 0 losses, 157 draws, 0 unknowns

Bert

[Rein Halbersma]

Most recent match with tmgr (the tool developed by Ed).
Conditions Scan_31nnue vs Scan_31, played on an Intel Core i7 8700K, 2-move start positions, TC 75 moves in 2 minute, books off, 6-piece dbs, 1 search thread.

Code: Select all

[ 1]: 0.502 score,   632 games,    6 wins,    3 losses,   623 draws,   0 unk
[ 2]: 0.502 score,   632 games,    5 wins,    3 losses,   624 draws,   0 unk
[ 3]: 0.501 score,   632 games,    6 wins,    5 losses,   621 draws,   0 unk
[ 4]: 0.498 score,   632 games,    4 wins,    6 losses,   620 draws,   2 unk
[ 5]: 0.497 score,   632 games,    4 wins,    8 losses,   620 draws,   0 unk
[ 6]: 0.498 score,   632 games,    2 wins,    5 losses,   625 draws,   0 unk
total 0.500 score,  3792 games,   27 wins,   30 losses,  3733 draws,   2 unk
elo diff -0.3

Bert

Congrats Bert! That is an impressive result. Will you release the source code for Scan-NNUE?

[BertTuyt]

Rein, I will do tomorrow.

I first wanted to clean up some messy parts, and re-write some sections so that it contains no older traces of the Jonathan code.
But I will not wait for that, so all can have a look.

I will also add source code to translate the python weight.txt file (from the python trainnnet script) into a network file which will be read by the program.

Bert

[BertTuyt]

Herewith a link to scan nnue sources.

https://www.dropbox.com/sh/5sw3zd3jq9z8 ... XGrRa?dl=0

The new/modified files are:
* eval.cpp, eval.hpp
* nnue.cpp, nnue.hpp
* simd.hpp

Bert

[BertTuyt]

And herewith the link to cttgnn, which converts a weights.txt (output python script trainnet) into a .gnn file

https://www.dropbox.com/sh/ahsrgi5j4776 ... ew8va?dl=0

If you have questions , just post them.

Bert

[Yves]

Hello Bert
Please compile the gnn file.
Cordially

[BertTuyt]

Yves, I don't have access now to my computer, apparently I did not include the weights file in the sources, sorry for that.
Will provide a separate link later today.

Bert

[Krzysztof Grzelak]

And if using the DXP can play and whether it works.

[BertTuyt]

Krzysztof,

the test result against Kingsrow was with DXP.
So it seems to work.

Bert

[Krzysztof Grzelak]

I understand and thank you for your answer Bert. I will wait until everything you need to share and then I will play for 1 minute.

[BertTuyt]

Herewith the link with the network file, and .ini file, also including the scan nnue version.

https://www.dropbox.com/sh/2seqmqhn7cuj ... oihla?dl=0

On my computer im able to do test with a stable 4.3 GHz, so i expect that on slower (and older) systems the Elo difference between scan and scan nnue will be around 10 ELO (Scan still stronger).
A boundary condition for the program is that the processor supports avx2.

Bert

[Krzysztof Grzelak]

What libraries (.dlls) do Scan 3.1_nnue.