Test30 overpass Test20 ?

[David Bigler]

I know the controversy of self ELO but at least right now, Test30 is showing good progress and overpass Test20 on self ELO.
Not sure however if Test30 is now really stronger than Test20.....?

[Tryfon Gavriel]

It does seem as though technically on the "Networks" page : http://www.lczero.org/networks/ - that test30 progress is much more solid and assured! 4242.29 right now as of 3rd Nov 9:55 AM (UK time) for Number 30882

Please can we start a new thread called "Test 30 games for commenting" - I will be keen to video annotate them as there is a lot of attacking skill shown e.g. this game:

Leela Chess reacts enthusiastically to provocative moves!

https://www.youtube.com/watch?v=NJ3z5EXqFq8&t=0s&list=PL9JCz2Gsbqe5YuXT3wyqa0B0k9uK0Uxkw&index=2

Or just post the games to the "Games for commenting" and indicate Test 30 games. I am keen to video annotate them!

Cheers, K

[Paul Berger]

Test30 is still on its first LR drop when Test20 is already on the third and final one. The outcome is clear to me ;-)

[David Bigler]

I understand that dev wanted to go with Test20 to the end to have a final conclusion...
I am wondering if it makes sense to go further than where we are now.... I mean, what is the point to prove/say that test20 can overpass test10 with 100 million games?

IMHO - They should restrict resources spent to what was done on test10 +/- 30% and when it’s obvious that a test is not promising, just end it.

I however understand that we are at the very beginning and that there was a lot to learn/understand, but talking here of going onwards

Long live to LC0, great chess project!

Rgds

[David Vonderhar]

Well i just ran a little 10 games match(2+0) against test 20 and test 20 completely crushed test 30

lc0 ID 22093(4246.12) vs lc0 ID 30882(4242.29)      With opening book    

5/4/0 +200 ELO for 22093  

Here's a game

[Event "2 min/game"]

[Site "lc0 ID 22093(4246.12)  vs lc0 ID 30882(4242.29), Leela"]

[Date "2018.11.03"]

[Round "10.1"]

[White "Lc0 v0.18.1 30882"]

[Black "Lc0 v0.18.1 22093"]

[ECO "C50"]

[Result "0-1"]

1. e4 {book 0s} e5 {book 0s} 2. Nf3 {book 0s} Nc6 {book 

0s} 3. Bc4 {book 0s} Bc5 {book 0s} 4. O-O {book 0s} d6 

{book 0s} 5. b4 {book 0s} Bb6 {book 0s} 6. c3 {+0.21/6 

2s} Nf6 {book 0s} 7. d3 {+0.21/7 2s} h6 {book 0s (O-O)} 

8. a4 {-0.01/6 2s} a6 {+0.09/5 2s (a5)} 9. Nbd2 {book 0s} 

O-O {+0.05/5 2s} 10. h3 {+0.17/5 2s (Ab3)} Ne7 {+0.06/5 2s} 

11. Re1 {+0.08/6 3s (Ab3)} Ng6 {+0.10/5 2s} 12. Nf1 

{+0.13/6 3s (d4)} a5 {+0.01/5 3s (c6)} 13. b5 {+0.10/6 2s} 

Be6 {-0.06/5 3s} 14. Be3 {+0.28/6 2s (Axe6)} Bxc4 {-0.31/5 

2s (Axe3)} 15. Bxb6 {-0.08/7 2s (dxc4)} Bxd3 {-0.37/6 2s} 

16. Qxd3 {-0.11/6 1s} Nf4 {-0.48/7 2s (cxb6)} 17. Qd2 

{+0.05/7 1s} cxb6 {-0.59/7 2s} 18. Rad1 {0.00/7 2s} Qc7 

{-0.54/7 2s} 19. Qxd6 {-0.03/7 1s} Qxc3 {-0.46/8 1s} 

20. Ne3 {-0.18/7 3s} Nxe4 {-0.59/6 3s} 21. Qxb6 {-0.26/7 

2s} Rac8 {-0.60/7 3s (Cc5)} 22. Qxb7 {+0.02/6 2s} Nc5 

{-0.70/8 1s} 23. Qa7 {-0.04/6 2s (De7)} Ncd3 {-0.71/6 3s 

(e4)} 24. Rf1 {-0.12/5 1s} e4 {-1.03/6 2s} 25. Nd4 {+0.16/5 

2s (Ch4)} Qxd4 {-3.23/7 2s (Ta8)} 26. Qxd4 {-4.86/6 2s} 

Ne2+ {-3.18/8 1s} 27. Kh2 {-4.21/7 2s} Nxd4 {-3.64/7 1s} 

28. f3 {-3.97/6 1s} f5 {-3.93/7 2s} 29. fxe4 {-4.57/6 2s 

(Cd5)} fxe4 {-4.08/6 1s} 30. Rxf8+ {-4.61/6 2s (Cd5)} Kxf8 

{-4.03/5 2s} 31. Rb1 {-4.19/5 1s} Ke7 {-4.41/5 3s} 32. b6 

{-4.87/6 2s} Rb8 {-4.49/5 2s} 33. Nc4 {-4.54/6 2s} Ke6 

{-4.93/5 2s (Cc6)} 34. Nxa5 {-6.10/6 1s} Kd5 {-6.24/5 2s} 

35. b7 {-6.94/6 1s} e3 {-6.52/6 2s} 0-1

On Saturday, November 3, 2018 at 9:43:06 AM UTC, David Bigler wrote:

[garrykli...@gmail.com]

they said it would take more games with this settings they put in on test 20. possibly twice as muchl.

[Cscuile]

Within the next few weeks, Test30 should surpass Test20 in real Elo. Given progress continues. 

[David Vonderhar]

Hey, leela just trolled stockfish in chess.com/cccc  game 306

could you make a video of that?

[MathAndreas]

I have run some matches on a GTX 1060. Results are

Lc0 ID22097 - Lc0 ID30891 : TC 10s/0.1s Win/Loss/Draw 50-16-34, Winrate 67%  Elodiff +123
Lc0 ID22097 - Lc0 ID30891 : TC 72s/0.6s Win/Loss/Draw 47-6-47,   Winrate 71%  Elodiff +156

[Enqwert]

I have run 2+2, GTX 1060 

1: Lc0 22102 17,5/28  +85 Elo dif
2: Lc0 30893 10,5/28

For 1+1 test, it was around +130, so test 30 seem to benefit more from longer TC.

[Enqwert]

David I ran a new test with 22102 s 30913 at 1+1  with no tb, 2 move opening and they seem to be closer. Your results are pretty lopsided compared to mine, do you enable table base?