Not logged inRybka Chess Community Forum
Up Topic The Rybka Lounge / Computer Chess / Latest each: Stockfish +99 Elo over Lc0 (Nooman LDRv2)
- - By MrKris (****) Date 2021-06-20 19:29
The newest Stockfish & net vs. the newest Lc0 and recent net with Nooman Low Draw Rate v2 book ( http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=32336 ):

Stockfish_21061823 143.0-81.0 Lc0-0.28.0rc1-69454
Score of Stockfish_21061823 vs Lc0-0.28.0rc1-69454: +81 -19 =124 63.8%
...      Stockfish_21061823 playing White: 76 - 0 - 36  [0.839] 112
...      Stockfish_21061823 playing Black: 5 - 19 - 88  [0.438] 112
...      White vs Black: 95 - 5 - 124  [0.701] 224
Elo difference: 98.7 +/- 30.0, LOS: 100.0 %, DrawRatio: 55.4 %  224 games.


---compare with same book: http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?pid=588497
Stockfish_21060401  125.5-98.5 Lc0-0.27-net-69272
Score of Stockfish_21060401 vs Lc0-0.27-net-69272: 64 - 37 - 123 [0.560]
...      Stockfish_21060401 playing White: 58 - 5 - 49  [0.737] 112
...      Stockfish_21060401 playing Black: 6 - 32 - 74  [0.384] 112
...      White vs Black: 90 - 11 - 123  [0.676] 224
Elo difference: 42.1 +/- 30.5, LOS: 99.6 %, DrawRatio: 54.9 %  224 games.

---

In the new match (top), opening reversed game pairs:
- Stockfish very commonly won with the opening advantage (ussually white) and commonly was able to draw with an opening disadvantage (usually black);
- On the other hand Lc0 commonly only drew with an advantage and lost with a disadvantage.

Game #122, Lc0 0-1 Sf, white has a disadvantage then it steps its K into big problems for no reason and calls it only -0.51 disadvantage.
[FEN "3r2k1/2b2pp1/r1b4p/np1p4/3Qn3/1PP3P1/P4P1P/3RRBK1 w - - 2 25 "]

{After 24...Ne4 +1.22/26 2.8s} 25. Kg2 {?? -0.51/14 1.1s, Harmon 25.a4} Bb6 {+1.68/24 3.0s} 26. Qb4 {-0.72/17 1.9s}
Nxc3 {+1.91/25 0.75s} 27. Qxc3 {-0.81/17 0.34s} d4+ {+2.05/24 1.2s}
28. Qxc6 {-0.78/17 0.25s} Nxc6 {+1.93/27 2.6s} 29. Bxb5 {-0.79/18 0.22s}
Nb4 {+2.03/26 1.8s} 30. Bxa6 {-0.87/15 0.51s} Nxa6 {+2.13/24 1.1s -+, 0-1 move 53}

Game #127, Sf 1-0 Lc0, evidentially -0.71 means "Lost" for Lc0 because Stockfish is over +3 and quickly wins:
[FEN "2r2r2/3bk1p1/1p2p3/p2p2bP/1q1P4/1P2QR2/P4PK1/RB6 w - - 0 30"]

30. Qxg5+ {+2.77/26 3.4s} Kd6 {-0.66/27 0.37s} 31. Rxf8 {+3.16/26 2.0s}
Rxf8 {-0.71/25 0.40s} 32. Qxg7 {+3.20/27 2.0s} Rxf2+ {-0.70/25 0.35s}
33. Kxf2 {+3.35/26 1.3s} Qd2+ {-0.70/24 0.30s} 34. Kg3 {+3.46/27 3.5s}
Qe1+ {-0.67/21 3.4s} 35. Kg2 {+3.62/25 2.0s} Qd2+ {-0.67/21 0.61s}
36. Kh1 {+3.77/27 4.0s} Qd1+ {-0.68/20 0.80s} 37. Kh2 {+3.83/25 1.4s}
Qd2+ {-0.68/19 1.2s} 38. Qg2 {+3.84/28 3.2s} Qxd4 {-0.74/18 0.84s}
39. Bd3 {+3.92/24 0.89s} Qxa1 {-0.82/19 1.1s} 40. h6 {+3.98/22 1.1s}
e5 {-0.86/15 1.2s} 41. h7 {+5.17/23 1.2s} e4 {-1.56/16 5.3s}
42. Qg8 {+5.77/23 1.7s} Qb2+ {-2.85/12 1.3s} 43. Bc2 {+6.50/27 2.9s}
e3 {-8.04/12 18s} 44. h8=Q {+7.08/22 0.85s}
Qxc2+ {-10.49/1 0s, White wins by adjudication, 1-0} 1-0

Game #135 Sf 1-0 Lc0, black is lost a clear 2 pawns down (saying only -1.03) then 35...Ba4?? throws away that B to tactics:
[FEN "1rr2k2/3b1pbR/1q4p1/3Pp3/4P1P1/pP3P2/P2QB3/1KN4R w - - 7 35"]

35. Bf1 {+3.11/26 1.2s}
Ba4 {?? -1.03/10 1.2s minus 1.03 games?} 36. d6 {+6.57/21 0.66s} Qc6 {-1.56/16 3.6s}
37. Rh8+ {+8.00/26 1.0s} Bxh8 {-2.81/1 0s} 38. Rxh8+ {+8.19/23 0.66s}
Kg7 {-3.17/1 0s} 39. Qh6+ {+8.25/23 0.72s} Kf6 {-3.77/1 0s}
40. Qh4+ {+8.61/24 3.2s} Kg7 {-1.98/14 0.19s} 41. Rh7+ {+8.54/22 0.81s}
Kg8 {-2.11/16 0.070s} 42. Rh8+ {+8.69/27 0.60s} Kg7 {-2.45/1 0s}
43. Qh6+ {+8.83/24 0.76s} Kf6 {-3.04/1 0s} 44. g5+ {+8.90/23 0.71s}
Ke6 {-4.14/1 0s} 45. Bh3+ {+9.14/24 3.2s} Kxd6 {-5.17/1 0s}
46. Rxc8 {+9.17/22 0.29s} Rxc8 {-14.90/11 2.7s, White wins by adjudication, 1-0} 1-0
Attachment: Sfnewnew.pgn - Games (598k)
Parent - - By sarona (**) Date 2021-06-21 04:33
Stockfish_21061823 143.0-81.0 Lc0-0.28.0rc1-69454
Score of Stockfish_21061823 vs Lc0-0.28.0rc1-69454: +81 -19 =124 63.8%
...      Stockfish_21061823 playing White: 76 - 0 - 36  [0.839] 112
...      Stockfish_21061823 playing Black: 5 - 19 - 88  [0.438] 112
...      White vs Black: 95 - 5 - 124  [0.701] 224
Elo difference: 98.7 +/- 30.0, LOS: 100.0 %, DrawRatio: 55.4 %  224 games.


I realize that a very good low draw rate book was used, but to see Lc0 crushed +0 -76 =36 (with Black) was unexpected. Especially when Stockfish was able to secure 88 draws with the Black pieces. Even if you support the idea of a Leela Ratio, this was still a pretty severe beating.

I remember reading a thread back in January of 2019 at the other forum (TC) where the poster was talking about blind spots in the networks Lc0 used. Now I know that Leela (Lc0 v0.20.1) and the networks have advanced since the time of that writing, but it seems that this still happens too often.  Some of the fairly recent CCC tournament games come to mind. . .  http://talkchess.com/forum3/viewtopic.php?f=2&t=69783&hilit=Lc0+blind+spots

Thanks for the link to those games.
Parent - - By MrKris (****) Date 2021-06-21 07:55 Upvotes 1

> I realize that a very good low draw rate book was used, but to see Lc0 crushed +0 -76 =36 (with Black) was unexpected. Especially when Stockfish was able to secure 88 draws with the Black pieces.


And: Stockfish_21061823 143.0-81.0 Lc0-0.28.0rc1-69454 =63.8%
vs. my last (same book) Stockfish_21060401  125.5-98.5 Lc0-0.27-net-69272 =56%

If Stockfish's new nets are a factor there is still another 21061923 at abrok.

> Even if you support the idea of a Leela Ratio, this was still a pretty severe beating.


https://www.sp-cc.de/nn-vs-sf-testing.htm
"Hardware: i7-8750H 2.6GHz (Hexacore) Notebook, RTX 2060 GPU, Windows 10 64bit, 16GB RAM
Speed:  Stockfish (running on 11 hyperthreading-threads, Intel Turbo-Mode off): 9000 kn/s, Lc0 (with old 32930 20x256 net): 16000 n/s in starting position. Since Lc0 0.26.3, Lc0 uses Cuda 11.1, which makes it around +37% faster. To keep the Leela-Ratio on 1.0, I slowed down the GPU with the Afterburner-Tool.
...
Thinking time: Lc0 2'+1'' and Stockfish 3'+1.5'' (means a perfect Leela-Ratio of 1.0). Average game-duration: 8 minutes, one 750 games-testrun takes around 4 days. "


My hardware: R7 2700X 8-core/16-threads desktop, RTX 2060 (desktop not laptop version)
My speed: Stockfish (16 SMT threads) 11006 kn/s,
Lc0 (LS15 20x256 net - sorry I don't have the old one handy) 24000 n/s,
CPU and GPU kept at stock - no adjustments/slow-down of any kind.
...
My T/C: both G/1 min. + 1 sec., about 15 games/hour, 224 Noomnan LDRv2 col. rev. = about 15 hours.

I see I forgot to mention my T/C in my post. >> Thanks for the link to those games.
Though its in the .pgn :smile: .
Parent - - By sarona (**) Date 2021-06-21 18:09
Not doubting your results. I followed your threads at the old Immortal Forum. I respect your testing methods. I was just surprised at how Stockfish won so many games with White. Also, Stefan Pohl is one of my favorite testers.

I have seen the elo jump with the new Stockfish nets trained on Leela data. Hopefully Lc0 can make some progress. Stockfish needs competition in engine competitions. Maybe Dragon might step up. I am skeptical that it can, though.
Parent - By MrKris (****) Date 2021-06-21 20:26

> Not doubting your results.


Thanks! Just an interesting point: Stefan Pohl's CPU (being Intel) is a little better than one might think even though a laptop version, and his RTX 2060 is maybe a little weaker than one might think being a laptop version (my 2060 is 160 watts, any laptop would melt*).
So, I think (but just a computer amateur) that I have at least as good, if not better, Leela Ratio than he does, that is to my Lc0's good - definitely not worse than his for Lc0 I believe.

> Stefan Pohl is one of my favorite testers.


Mine also! He makes excellent opening books, maintains his site and lists all the time. (In contrast I just try to post something interesting occasionally, much less work than he is doing.)

> I followed your threads at the old Immortal Forum.


Perhaps an interesting story about someone I know: They liked the public Immortal section for non-members and joined. The for members sections had 'controversial' aspects that was a surprise compared to the public section. This person thought "okay, if I never do anything 'controversial' then my ISP, computer, etc. can not have any trace of what never existed". So they stayed a member and ignored any 'controversial' aspects.

> I have seen the elo jump with the new Stockfish nets trained on Leela data. Hopefully Lc0 can make some progress. Stockfish needs competition in engine competitions. Maybe Dragon might step up. I am skeptical that it can, though.


Stockfish seems to be coming up with original ideas. The free engines can have a disadvantage in that the authors are donating their spare time, the commercial engines can have a disadvantage of being 'vs. a team of authors'.

 
*Years ago I bought a refurbished, but major brand, laptop with a short warranty. I foolishly did not look into extending it and it broke a very short time after it ended. I wish him good luck but I suspect laptops are not really up to chess (on TalkChess he related that he needed to utilize his warranty) -  chess equivalent to a stress test 100% of the time vs. even the most demanding video games being much less so. This was before Lc0: I don't think he uses ponder so in his CPU vs GPU games one or the other gets somewhat of a break while the other is moving which I did not have.
Parent - - By Ozymandias (****) Date 2021-06-23 08:44
At that TC, smaller nets might do better. Have you tried LS 15 (20x256), 751236 (15x192) or even J104.1-30 (10x128)?
Parent - - By MrKris (****) Date 2021-06-25 05:50
Thanks but the githhub advise has changed:

Best Nets for Lc0 (official Lc0 github)
https://github.com/LeelaChessZero/lc0/wiki/Best-Nets-for-Lc0
Size versus Recommended Purpose
    30b: Recommended for RTX cards (any time control), GTX cards (analysis, long time control)

Examples:

fastgm rating list:
http://fastgm.de/16-60-0.6.html
60 seconds / game + 0.6 seconds / move
16 Cores vs GeForce RTX 2070
4   Lc0 0.27.0 68864   3519  
24   Lc0 0.25.1 LS 15   3480
52   Lc0 0.25.1 703042   3403

Scaling: Stockfish thread count and Lc0 net size
http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=33879
G/90"+1"  R7 2700X | RTX 2060 
Each 100 games (same openings) Sf 1th, 4th & 16th
10x128 net                     Elo  +-  Games
   0 Lc0-0.27+_J104.1-30       -56  17   300
20x256 net
   0 Lc0-0.27--LS15            -35  16   300
30x384 net           
   0 Lc0-0.27--J94-100          2   14   300
Parent - - By sarona (**) Date 2021-06-27 00:33
Perhaps net size/time control could be the problem.

[Event "G/1'+1'' 1Gh Sy.6 NoomanLDRv2"]
[Site "R7 2700X 16threads | RTX 2060"]
[Date "2021.06.19"]
[Round "41"]
[White "Stockfish_21061823"]
[Black "Lc0-0.28.0rc1-69454"]
[Result "1-0"]
[ECO "B06"]
[GameDuration "00:02:43"]
[GameEndTime "2021-06-19T04:55:04.329"]
[GameStartTime "2021-06-19T04:52:21.104"]
[Opening "Robatsch (Modern) defense"]
[Termination "adjudication"]
[SetUp "1"]
[FEN "r2q1k2/4p1bB/p4p2/2pnR3/pr2N1Pp/5Q1P/2P5/5RK1 b - - 0 28"]
[TimeControl "60+1"]
[PlyCount "82"]

28... Nc7 {-0.38/15 2.0s} 29. c3 {+1.55/25 1.1s}
Rab8 {+0.02/15 2.3s} 30. Nxf6 {+1.75/25 1.2s} Bxf6 {0.00/12 0.24s}
31. g5 {+2.08/25 1.7s} Qd6 {-1.78/16 3.0s} 32. Rfe1 {+2.15/25 1.5s}
Rb1 {-2.03/16 0.30s} 33. Bxb1 {+2.52/23 1.2s} Rxb1 {-2.01/16 0.53s}
34. Rxe7 {+2.54/23 1.1s} Rxe1+ {-2.15/13 1.9s} 35. Rxe1 {+2.71/22 1.5s}
a3 {-1.82/12 3.5s} 36. gxf6 {+2.83/22 2.3s} Ne6 {-1.83/13 0.30s}
37. Rf1 {+3.81/22 1.1s} a2 {-2.50/14 6.1s} 38. Qa8+ {+4.96/20 0.89s}
Kf7 {-2.63/13 0.24s} 39. Qb7+ {+5.49/21 1.1s} Kf8 {-3.13/12 0.74s}
40. Qe7+ {+6.08/23 0.91s} Qxe7 {-5.48/14 1.9s} 41. fxe7+ {+6.54/24 1.2s}
Kxe7 {-5.66/13 0.30s, White wins by adjudication} 1-0

After 28...Nc7, Lc0's evals become inconsistent between moves 28 through 31.

28...Nc7 (0.38/15 2.0s)
29...Rab8 (0.02/15 2.3s)
30...Bxf6 (0.00/12 0.24s)
31...Qd6 (-1.78/16 3.0s)

32...Rb1 (-2.03/16 0.30s)

Meanwhile, Stockfish's evals are running 1.55 -2.00 throughout. It seems that only after 31...Qd6, does Leela see problems.

There are numerous examples of this in the match.
Parent - - By sarona (**) Date 2021-06-27 03:44 Edited 2021-06-27 03:59
Here is another one - Game 46.

Lc0 has the position as 0.00 from 17...Rxb2 all the way to 22...Kf8.  Then realizes that disaster awaits after 23...Nf3+ with an eval of 2.28. By move 33, the eval is over 5.00

Event "G/1'+1'' 1Gh Sy.6 NoomanLDRv2"]
[Site "R7 2700X 16threads | RTX 2060"]
[Date "2021.06.19"]
[Round "46"]
[White "Lc0-0.28.0rc1-69454"]
[Black "Stockfish_21061823"]
[Result "0-1"]
[ECO "B48"]
[GameDuration "00:02:35"]
[GameEndTime "2021-06-19T05:29:06.831"]
[GameStartTime "2021-06-19T05:26:31.022"]
[Opening "Sicilian"]
[PlyCount "81"]
[Termination "adjudication"]
[TimeControl "60+1"]
[Variation "Taimanov Variation"]

1. e4 {book} c5 {book} 2. Nf3 {book} e6 {book} 3. d4 {book} cxd4 {book}
4. Nxd4 {book} Nc6 {book} 5. Nc3 {book} Qc7 {book} 6. Be3 {book} a6 {book}
7. a3 {book} Nf6 {book} 8. f4 {book} d6 {book} 9. Qf3 {book} Be7 {book}
10. O-O-O {book} O-O {-0.68/24 2.8s} 11. Nxc6 {+0.38/14 4.3s}
bxc6 {-0.86/25 2.4s} 12. g4 {+0.39/13 0.30s} d5 {-0.77/26 5.3s}
13. g5 {+0.48/16 2.7s} Nd7 {-0.98/28 19s} 14. h4 {+0.56/12 2.7s}
Rb8 {-0.69/24 2.8s} 15. h5 {+0.59/10 0.88s} Qa5 {-0.62/23 1.00s}
16. Bf2 {+0.64/10 2.9s} Rd8 {-0.78/24 2.7s} 17. g6 {+0.92/13 3.8s}
Rxb2 {0.00/31 2.1s} 18. Kxb2 {+1.29/14 2.7s} Qxa3+ {0.00/34 1.2s}
19. Kb1 {+1.27/1 0s} Qb4+ {0.00/36 1.5s} 20. Kc1 {+1.88/14 2.0s}
Qa3+ {0.00/37 1.1s} 21. Kd2 {+1.94/12 0.38s} Ne5 {0.00/37 1.7s}
22. gxf7+ {+1.00/16 6.1s} Kf8 {0.00/38 1.4s} 23. Qh3 {+1.22/17 0.24s}
Nf3+ {+2.28/23 1.3s} 24. Ke2 {-1.22/20 9.4s} Qxc3 {+2.37/26 2.2s}
25. e5 {-1.21/1 0.001s} Qxc2+ {+2.09/29 3.0s} 26. Kxf3 {-1.26/1 0s}
Qxd1+ {+2.27/28 1.9s} 27. Kg2 {-1.26/17 0.30s} Qd2 {+2.38/27 2.0s}
28. Bd3 {-1.19/15 0.30s} Qxf4 {+3.49/21 1.5s} 29. h6 {-1.32/16 1.6s}
Qg5+ {+3.66/21 1.2s} 30. Kf1 {-1.41/16 3.4s} g6 {+4.41/23 1.6s}
31. Qf3 {-1.55/19 2.5s} d4 {+4.68/23 1.4s} 32. Rg1 {-1.53/27 0.36s}
Qc1+ {+4.91/23 1.4s} 33. Kg2 {-2.37/14 7.2s} Qd2 {+5.03/25 4.3s}
34. Kf1 {-1.37/13 0.62s} a5 {+5.12/23 3.3s} 35. Rg3 {-2.18/12 1.8s}
Qc1+ {+5.14/23 1.3s} 36. Kg2 {-2.63/15 0.48s} Qxh6 {+5.27/25 1.9s}
37. Rh3 {-2.57/11 0.40s} Qg5+ {+5.38/24 1.6s} 38. Kh1 {-3.72/13 15s}
Kg7 {+6.15/22 1.7s} 39. Bg3 {-4.09/11 3.7s} Qe3 {+6.92/25 1.4s}
40. Qf1 {-5.09/10 0.30s} Bg5 {+7.01/24 1.2s}
41. Bh4 {-6.13/9 0.69s, Black wins by adjudication} 0-1

It actually took Stockfish awhile to get an eval over (-) 2.00
Parent - By MrKris (****) Date 2021-06-27 05:46
This is extraordinarily common.

Here are four decisive games from TCEC 20 SuperFinal:

1) https://tcec-chess.com/#div=sf&game=6&season=20 Wh.Lc0 1-0 Bk.Stockfish
2) https://tcec-chess.com/#div=sf&game=52&season=20 Wh.Lc0 1-0 Bk.Stockfish
3) https://tcec-chess.com/#div=sf&game=5&season=20 Wh.Stockfish 1-0 Bk.Lc0
4) https://tcec-chess.com/#div=sf&game=33&season=20 Wh.Stockfish 1-0 Bk.Lc0
 
Parent - By MrKris (****) Date 2021-06-27 05:26

> Perhaps net size/time control could be the problem.


net size - in this thread two days before your post I already answered that:
___________________________________________________________________
MrKris (***) Date 2021-06-24 21:50
Thanks but the githhub advise has changed:

Best Nets for Lc0 >>> OFFICIAL Lc0 GITHUB <<<
https://github.com/LeelaChessZero/lc0/wiki/Best-Nets-for-Lc0
Size versus Recommended Purpose
    30b: Recommended for RTX cards (any time control), GTX cards (analysis, long time control)
___________________________________________________________________

/time control - see http://fastgm.de/16-60-0.6.html
"16 Cores vs GeForce RTX 2070"  [2x Xeon, circa 2014]
his list shows all of the recent largest nets doing much better than all of the smaller nets at his G/60"+0.6".

Also:
https://www.sp-cc.de/nn-vs-sf-testing.htm
1 Stockfish 13 bmi2                : 3723   
2 Lc0 0.27.0 68695 (30x384)   : 3661
https://www.sp-cc.de/index.htm
   1 Stockfish 210615 avx2    : 3755    7    7  7000    75.4 %   3542   47.9 %
   2 Stockfish 210619 avx2    : 3751    7    7  7000    75.1 %   3542   48.7 %
   3 Stockfish 13 210218       : 3723

S. Pohl has Sf13 +62 Elo over this best Lc0 and his latest Sf 6/19 +28 Elo over Sf13
equals S. Pohl: his latest Stockfish + 90 Elo over Lc0

could be the problem. - I do not see one.

- - - - - - - -    - - - - - - - -

> After 28...Nc7, Lc0's evals become inconsistent between moves 28 through 31.


> There are numerous examples of this in the match.


Typical (not every game of course): the loser, not realizing its losing, taking time to 'catch up' to the winner's evals. which often where already rising.

TCEC S21 Div. P recently started but Lc0 has two excellent examples already:
https://tcec-chess.com/#div=p&game=52&season=21 Stoofvlees 53. 1.67 Lc0 53... 0.99 | Stoofvlees 57. 2.04 Lc0 57... 4.73 , Sv 1-0 Lc
https://tcec-chess.com/#div=p&game=16&season=21 Dragon 48. 4.14 Lc0 48... 1.02 | Dragon 57. 6.05 Lc0 57...1.99 | Dragon 63. 10.43 Lc0 63... 9.09 , Dr 1-0 Lc
One already from AllieStein:
https://tcec-chess.com/#div=p&game=1&season=21 Stockfish 33. 3.07 AllieStein 33... 1.14 | Stockfish 35. 3.29 AllieStein 4.75
 
- By MrKris (****) Date 2021-07-08 04:00
1st try (no Ponder) old Sf net
G/30"+0.5"  256Mh  Sy.6
R7 2700X 16ths | RTX 2060
Stockfish_21060401  125.5-98.5  Lc0-0.27-net-69272
Score of Stockfish_21060401 vs Lc0-0.27-net-69272: 64 - 37 - 123 [0.560]
...      Stockfish_21060401 playing White: 58 - 5 - 49  [0.737] 112
...      Stockfish_21060401 playing Black: 6 - 32 - 74  [0.384] 112
...      White vs Black: 90 - 11 - 123  [0.676] 224
Elo difference: 42.1 +/- 30.5, LOS: 99.6 %, DrawRatio: 54.9 %
224 of 224 games finished.

Previous, this thread - new Sf net!
G/60"+1" 1Gh Sy.6
R7 2700X 16ths | RTX 2060
Stockfish_21061823 143.0-81.0 Lc0-0.28.0rc1-69454
Score of Stockfish_21061823 vs Lc0-0.28.0rc1-69454: +81 -19 =124 63.8%
...      Stockfish_21061823 playing White: 76 - 0 - 36  [0.839] 112
...      Stockfish_21061823 playing Black: 5 - 19 - 88  [0.438] 112
...      White vs Black: 95 - 5 - 124  [0.701] 224
Elo difference: 98.7 +/- 30.0, LOS: 100.0 %, DrawRatio: 55.4 %  224 games.


Current:
G/30"+0.5" Ponder 1Gh Sy.6
R7 2700X 14ths | RTX 2060
Score of Stockfish_21070310 vs Lc0-0.28.0rc1-69637: 78 - 25 - 121 [0.618]
...      Stockfish_21070310 playing White: 74 - 1 - 37  [0.826] 112
...      Stockfish_21070310 playing Black: 4 - 24 - 84  [0.411] 112
...      White vs Black: 98 - 5 - 121  [0.708] 224
Elo difference: 83.8 +/- 30.7, LOS: 100.0 %, DrawRatio: 54.0 %  224 games.


Maybe the Lc0 net has finally improved, and/or its just the +/- 30 Elo from the only 224 games.
Also, I reduced Sf's threads to 14 to be able to use ponder.
__________________________________________________________

I also tried this, but it did not work out well:
SfClassic = 'use NNUE' false (unchecked)
G/30"+0.5" Ponder 1Gh Sy.6
R7 2700X 14ths | RTX 2060
SfClassic_21070310  31.5-48.5  Lc0-0.28.0rc1-69637
Score of SfClassic_21070310 vs Lc0-0.28.0rc1-69637: 14 - 31 - 35 [0.394]
...      SfClassic_21070310 playing White: 11 - 9 - 20  [0.525] 40
...      SfClassic_21070310 playing Black: 3 - 22 - 15  [0.263] 40
...      White vs Black: 33 - 12 - 35  [0.631] 80
Elo difference: -75.0 +/- 57.9, LOS: 0.6 %, DrawRatio: 43.8 %
80 of 224 games finished - match abandoned (Lc0 99.4% LOS)


I saw this game, #80 Rd. 40, and decided to end it.
1st white pawn sacrifice 12. h4!
White pawn offer 16. g4!, declined.
2nd white pawn sacrifice 28. c4!
After 30. Rb4 {+4.53} b5 {-(bad for black)1.73} Lc0 has clear win from its 2 pawn sacrifices.
Black has an extra pawn each side but white has connected passed d and e pawns.
4brr1/p4k2/6p1/1p1PPp1p/1Rp2P1N/P7/3K4/6R1 w - - 0 31
 
[Event "Sf Classic vs Lc0 | NoomanLDRv2 | 1Gh Sy.6"]
[Site "G/30''+.5'' Ponder R7 2700X | RTX 2060"]
[Date "2021.07.06"]
[Round "40"]
[White "Lc0-0.28.0rc1-69637"]
[Black "SfClassic_21070310"]
[Result "1-0"]
[ECO "C19"]
[GameDuration "00:01:25"]
[GameEndTime "2021-07-06T20:01:53.413"]
[GameStartTime "2021-07-06T20:00:27.775"]
[Opening "French"]
[PlyCount "84"]
[Termination "adjudication"]
[TimeControl "30+0.5"]
[Variation "Winawer, Advance, poisoned pawn Variation"]

1. e4 {book} e6 {book} 2. d4 {book} d5 {book} 3. Nc3 {book} Bb4 {book}
4. e5 {book} c5 {book} 5. a3 {book} Bxc3+ {book} 6. bxc3 {book} Ne7 {book}
7. Qg4 {book} O-O {book} 8. Bd3 {book} Nbc6 {book} 9. Nf3 {+0.39/11 3.4s}
f5 {-0.60/25 3.8s} 10. Qg3 {+0.49/12 0.33s} Qb6 {-0.66/24 2.1s}
11. Bh6 {+0.52/16 2.1s} Ng6 {+0.30/21 0.72s} 12. h4 {+0.76/13 1.2s}
gxh6 {-1.23/29 8.0s} 13. h5 {+0.81/18 0.97s} Kg7 {-1.15/27 0.28s}
14. hxg6 {+0.81/18 0.33s} hxg6 {-1.19/26 0.36s} 15. Qh4 {+0.85/17 0.38s}
h5 {-1.23/28 0.55s} 16. g4 {+0.90/15 0.29s} Qd8 {-1.03/28 0.40s}
17. Ng5 {+0.92/16 0.30s} Rh8 {-1.21/21 1.8s} 18. gxf5 {+0.89/1 0.088s}
exf5 {-1.10/21 0.86s} 19. Rg1 {+0.92/1 0.036s} Qe7 {-1.30/22 4.1s}
20. Rb1 {+1.05/1 0.015s} b6 {-1.30/24 2.1s} 21. Bb5 {+1.09/1 0.045s}
Bb7 {-1.29/21 1.3s} 22. Kd2 {+1.15/12 3.0s} cxd4 {-1.45/23 1.3s}
23. cxd4 {+1.57/12 1.2s} Raf8 {-1.82/23 0.24s} 24. Bxc6 {+1.67/12 0.76s}
Bxc6 {-1.78/21 0.78s} 25. Nf3 {+1.74/12 0.27s} Qxh4 {-2.14/25 4.6s}
26. Nxh4 {+2.90/13 0.75s} Be8 {-2.03/24 0.002s} 27. f4 {+3.16/11 0.67s}
Rhg8 {-2.41/25 4.1s} 28. c4 {+4.25/12 0.90s} dxc4 {-2.63/21 0.44s}
29. d5 {+4.50/11 0.34s} Kf7 {-2.56/23 1.2s} 30. Rb4 {+4.56/11 2.7s}
b5 {-1.73/19 0.28s} 31. Nf3 {+5.29/11 0.39s} Kg7 {-2.46/21 1.6s}
32. Ng5 {+6.35/10 4.1s} Bd7 {-2.98/26 0s} 33. e6 {+7.03/10 0.30s}
Be8 {-3.08/22 0.005s} 34. Kc3 {+4.44/10 5.9s} Rh8 {-3.38/22 0.81s}
35. Kd4 {+4.56/1 0.055s} h4 {-2.11/17 0.19s} 36. Ke5 {+7.60/9 1.4s}
a5 {-4.79/19 1.5s} 37. Rb2 {+8.39/11 0.66s} h3 {-4.87/20 0.002s}
38. Rh1 {+9.04/10 1.0s} Rh4 {-5.20/19 0.066s} 39. Rhh2 {+9.42/11 2.4s}
b4 {-6.00/18 0.001s} 40. axb4 {+12.07/9 0.79s} a4 {-6.72/20 0.008s}
41. b5 {+24.77/7 1.3s} c3 {-7.97/18 2.2s} 42. Rbc2 {+26.07/8 0.21s}
Bxb5 {-8.92/17 0.15s, White wins by adjudication} 1-0

Attachment: Sf-Lc69637.pgn - Games (626k)
Attachment: SfCl-Lc69637.pgn - Games (222k)
Up Topic The Rybka Lounge / Computer Chess / Latest each: Stockfish +99 Elo over Lc0 (Nooman LDRv2)

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill