Page 1 of 2

32930 Boost network available

Posted: Tue Apr 09, 2019 5:02 am
by dkappe
https://github.com/dkappe/leela-chess-w ... t-Networks

Based on my hypothesis on what was the critical element in the success of Antifish, this is 32930 sharpened with CCRL games at a low learning rate and a very few steps. Seems to give around 30 elo extra than the original net. I'll be doing this for a T40 net as well.

Re: 32930 Boost network available

Posted: Tue Apr 09, 2019 7:59 pm
by Javier Ros
Very good idea! Thanks for share!
I am testing Lc0 Boost 32930, Lc0_32930 and Lc0_41871, all with 0.21.1, on Rtx2070 against Stockfish dev. i7 4790S 8 cores

Re: 32930 Boost network available

Posted: Tue Apr 09, 2019 8:24 pm
by sovaz1997
I don't think that this test can talk about anything. Need more games:

Code: Select all

# PLAYER              :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W    D    L  D(%)
1 32930-boost-1000    :       0     30    33.5      60  55.8      77   12   43    5  71.7
2 32930               :     -25     49    14.0      30  46.7      78    3   22    5  73.3
3 11258               :     -62     50    12.5      30  41.7     ---    2   21    7  70.0

White advantage = 78.53 +/- 24.80
Draw rate (equal opponents) = 82.20 % +/- 7.64
5-3 is not 30 Elo. It is unknown.

Re: 32930 Boost network available

Posted: Tue Apr 09, 2019 8:57 pm
by dkappe
I applied this to my own 256x20-se net called Maddex and ran 2+2 for 800 games and got +30 elo. Right now Iā€™m running tests for 500 to 4500 step nets. 32930 may need more steps. TBD.

Re: 32930 Boost network available

Posted: Thu Apr 11, 2019 6:33 pm
by Javier Ros
Test with Silver suite 50x2 game each match:
Lc0 Boost 32930, Lc0_32930 and Lc0_41871, all with 0.21.1 version, on Rtx2070 against Stockfish dev. i7 4790S at 3.6 MGhz. 8 cores at 5 minutes plus 3 seconds under Fritz gui. First half of the test finished with 150 games played.

Lc0 Boost 32930 scored much better than 32930 and almost as good as 41871.

Lc0_41871SFdev, Blitz 5m+3s-3 2019

Stockfish 090419 64 BMI2 - Lc0 v0.21.1_41871 22.5 - 27.5 +6/=33/-11 45.00%
Stockfish 090419 64 BMI2 - Lc0 v0.21.1Boost32930 23.0 - 27.0 +6/=34/-10 46.00%
Stockfish 090419 64 BMI2 - Lc0 v0.21.1_32930 26.5 - 23.5 +9/=35/-6 53.00%

Program Elo + - Games Score Av.Op. Draws

1 Lc0 v0.21.1_41871 : 2428 57 55 50 55.0 % 2393 66.0 %
2 Lc0 v0.21.1Boost32930 : 2421 56 54 50 54.0 % 2393 68.0 %
3 Stockfish 090419 64 BMI2 : 2393 31 32 150 48.0 % 2407 68.0 %
4 Lc0 v0.21.1_32930 : 2372 52 54 50 47.0 % 2393 70.0 %

Re: 32930 Boost network available

Posted: Sat Apr 13, 2019 3:31 pm
by Raphexon
Would it be possible to distill the latest network to 16x2?

I assume you need an Nvidia GPU for it. (Which I don't have)

Re: 32930 Boost network available

Posted: Mon Apr 15, 2019 5:07 pm
by Javier Ros
Javier Ros wrote: ā†‘Thu Apr 11, 2019 6:33 pm Test with Silver suite 50x2 game each match:
Lc0 Boost 32930, Lc0_32930 and Lc0_41871, all with 0.21.1 version, on Rtx2070 against Stockfish dev. i7 4790S at 3.6 MGhz. 8 cores at 5 minutes plus 3 seconds under Fritz gui. First half of the test finished with 150 games played.

Lc0 Boost 32930 scored much better than 32930 and almost as good as 41871.

Lc0_41871SFdev, Blitz 5m+3s-3 2019

Stockfish 090419 64 BMI2 - Lc0 v0.21.1_41871 22.5 - 27.5 +6/=33/-11 45.00%
Stockfish 090419 64 BMI2 - Lc0 v0.21.1Boost32930 23.0 - 27.0 +6/=34/-10 46.00%
Stockfish 090419 64 BMI2 - Lc0 v0.21.1_32930 26.5 - 23.5 +9/=35/-6 53.00%

Program Elo + - Games Score Av.Op. Draws

1 Lc0 v0.21.1_41871 : 2428 57 55 50 55.0 % 2393 66.0 %
2 Lc0 v0.21.1Boost32930 : 2421 56 54 50 54.0 % 2393 68.0 %
3 Stockfish 090419 64 BMI2 : 2393 31 32 150 48.0 % 2407 68.0 %
4 Lc0 v0.21.1_32930 : 2372 52 54 50 47.0 % 2393 70.0 %
In the second half of the test, the results were

Lc0_41871SFdev, Blitz 5m+3s 2019

Stockfish 090419 64 BMI2 - Lc0 v0.21.1_41871 24.0 - 26.0 +4/=40/-6 48.00%
Stockfish 090419 64 BMI2 - Lc0 v0.21.1Boost32930 22.5 - 27.5 +4/=37/-9 45.00%
Stockfish 090419 64 BMI2 - Lc0 v0.21.1_32930 25.0 - 25.0 +7/=36/-7 50.00%

so that the total test result is

Lc0_41871SFdev, Blitz 5m+3s-3 2019

Stockfish 090419 64 BMI2 - Lc0 v0.21.1_41871 46.5 - 53.5 +10/=73/-17 46.50%
Stockfish 090419 64 BMI2 - Lc0 v0.21.1Boost32930 45.5 - 54.5 +10/=71/-19 45.50%
Stockfish 090419 64 BMI2 - Lc0 v0.21.1_32930 51.5 - 48.5 +16/=71/-13 51.50%

and Lc0 with 32930 Boost gets the best result!
6 men syzygy tablebases for both.

Re: 32930 Boost network available

Posted: Tue Apr 16, 2019 7:53 am
by Modern Times
dkappe wrote: ā†‘Tue Apr 09, 2019 5:02 am https://github.com/dkappe/leela-chess-w ... t-Networks

Based on my hypothesis on what was the critical element in the success of Antifish, this is 32930 sharpened with CCRL games at a low learning rate and a very few steps. Seems to give around 30 elo extra than the original net. I'll be doing this for a T40 net as well.
Which T40 net do you think you'll choose ?

Re: 32930 Boost network available

Posted: Tue Apr 16, 2019 9:19 am
by glennsamuel32
Raphexon wrote: ā†‘Sat Apr 13, 2019 3:31 pm Would it be possible to distill the latest network to 16x2?
The distillation process doesn't work on test 40 and above networks, yet.

Re: 32930 Boost network available

Posted: Wed Apr 17, 2019 11:10 pm
by dkappe
After some additional testing, the boost net at 7000 steps is about the same head 2 head and vs sf10 at 1+1. The style of play is a bit different, however. Would appreciate feedback. https://github.com/dkappe/leela-chess-w ... t-Networks