)Self play elo was ~43 through 10k games.
Bench: 6612352
ELO | 43.13 +- 3.82 (95%)
CONF | 32.0+0.32s Threads=1 Hash=64MB
GAMES | N: 10024 W: 2229 L: 991 D: 6804
https://github.com/jhonnold/berserk/releases/tag/4.5.0
New engine releases & news 2021
Moderators: hgm, Rebel, chrisw
-
jhonnold
- Posts: 117
- Joined: Wed Feb 17, 2021 3:16 pm
- Full name: Jay Honnold
Berserk 4.5.0
Berserk 4.5.0 available. Apologies to testers for another rapid release, but wanted to get this one out before I take a break for a bit (Summer and Moving )
Self play elo was ~43 through 10k games.
Bench: 6612352
ELO | 43.13 +- 3.82 (95%)
CONF | 32.0+0.32s Threads=1 Hash=64MB
GAMES | N: 10024 W: 2229 L: 991 D: 6804
https://github.com/jhonnold/berserk/releases/tag/4.5.0
)Self play elo was ~43 through 10k games.
Bench: 6612352
ELO | 43.13 +- 3.82 (95%)
CONF | 32.0+0.32s Threads=1 Hash=64MB
GAMES | N: 10024 W: 2229 L: 991 D: 6804
https://github.com/jhonnold/berserk/releases/tag/4.5.0
-
j.t.
- Posts: 239
- Joined: Wed Jun 16, 2021 2:08 am
- Location: Berlin
- Full name: Jost Triller
Re: New engine releases & news 2021
Finally, I can release version 1.11 of Nalwald 
Nalwald 1.11 is approximately 100 Elo points better than Nalwald 1.10 in self-play. I didn't test against other engines, so I only speculate that true improvement is around 60 Elo.
- I made quite a few changes to futility reductions
- I introduced hash result futility pruning
- null move and late move reductions are now calculated differently
- I added a new evaluation bonus for a knight that attacks a bishop, rook, or queen
- I added some kind of relativity to the history heuristic: for moves that fail low, I reduce the history value
- thanks to Martin and John Merlino I found out about optimizing the K factor of the sigmoid function, so I did that
- the optimization algorithm got a small redesign
The issue that caused some crashes in previous version has been investigated but not fixed, as the solution (switching to a different memory management) would lose around 10 Elo points. The cause to why the alternative memory management is slower has been found and is expected to be fixed at some point.
Downloads are here.
Nalwald 1.11 is approximately 100 Elo points better than Nalwald 1.10 in self-play. I didn't test against other engines, so I only speculate that true improvement is around 60 Elo.
- I made quite a few changes to futility reductions
- I introduced hash result futility pruning
- null move and late move reductions are now calculated differently
- I added a new evaluation bonus for a knight that attacks a bishop, rook, or queen
- I added some kind of relativity to the history heuristic: for moves that fail low, I reduce the history value
- thanks to Martin and John Merlino I found out about optimizing the K factor of the sigmoid function, so I did that
- the optimization algorithm got a small redesign
The issue that caused some crashes in previous version has been investigated but not fixed, as the solution (switching to a different memory management) would lose around 10 Elo points. The cause to why the alternative memory management is slower has been found and is expected to be fixed at some point.
Downloads are here.
-
Gabor Szots
- Posts: 1362
- Joined: Sat Jul 21, 2018 7:43 am
- Location: Szentendre, Hungary
- Full name: Gabor Szots
Re: New engine releases & news 2021
Thanks Jost. I have started a test so as far as I am concerned you don't have to hurry releasing a new version.
Gabor Szots
CCRL testing group
CCRL testing group
-
Gabor Szots
- Posts: 1362
- Joined: Sat Jul 21, 2018 7:43 am
- Location: Szentendre, Hungary
- Full name: Gabor Szots
Re: New engine releases & news 2021
Gabor Szots
CCRL testing group
CCRL testing group
-
Sylwy
- Posts: 4465
- Joined: Fri Apr 21, 2006 4:19 pm
- Location: IASI - the historical capital of MOLDOVA
- Full name: SilvianR
Re: New engine releases & news 2021
NNUE or NN ? The executable is small .......and nothing impressive in the play of this Rust engine. Seems just a small integrated NN.Gabor Szots wrote: ↑Sat Jul 24, 2021 10:44 am Velvet 2.0.0
New is that it uses NNUE.
https://github.com/mhonert/velvet-chess ... tag/v2.0.0
Ktulu 9 is an old 32-bit HCE chess engine.
[pgn]
[Event "NN Test121"]
[Site "ISR 3"]
[Date "2021.07.24"]
[Round "2"]
[White "Ktulu 9 w32"]
[Black "Velvet-v2.0.0-x86_64-avx2"]
[Result "1-0"]
[BlackElo "2200"]
[ECO "B90"]
[Opening "Sicilian"]
[Time "22:00:22"]
[Variation "Najdorf, 6.Be3 e5 7.Nb3 Be6 8.f3 Nbd7"]
[WhiteElo "2200"]
[TimeControl "240+2"]
[Termination "adjudication"]
[PlyCount "69"]
[WhiteType "program"]
[BlackType "program"]
1. Nc3 c5 2. e4 d6 3. Nf3 Nf6 4. d4 cxd4 5. Nxd4 a6 6. Be3 e5 7. Nb3 Be6 8.
f3 Nbd7 9. g4 Be7 {(Bf8-e7 g4-g5 Nf6-h5 Qd1-d2 O-O O-O-O Ra8-c8 Kc1-b1
Nh5-f4 h2-h4 Qd8-c7 Nc3-d5 Be6xd5 e4xd5 Nd7-b6 Be3xf4) +0.03/17 8} 10. Qd2
O-O {(O-O Rh1-g1 Qd8-c7 O-O-O Rf8-c8 g4-g5 Nf6-h5 Nc3-d5 Be6xd5 e4xd5
Nh5-f4 c2-c3 Nd7-b6 Be3xf4 e5xf4 Qd2xf4) +0.07/16 3} 11. O-O-O Qc7 {(Qd8-c7
g4-g5 Nf6-h5 Nc3-d5 Be6xd5 Qd2xd5 Nh5-f4 Qd5-a5 Qc7-c6 Qa5-c3 Qc6-a4 Qc3-a5
Qa4-c6 Qa5-c3 Qc6-a4 Qc3-a5 Qa4-c6 Qa5-c3) 0.00/18 5} 12. Kb1 Nb6 {(Nd7-b6
h2-h4 Rf8-d8 g4-g5 Nf6-h5 Nc3-d5 Nb6xd5 e4xd5 Be6-f5 Rh1-g1 Rd8-e8 Bf1-d3
Bf5-g6 Bd3-e4 Nh5-f4 Be4xg6) -0.12/17 8} 13. h4 {(h2-h4 g7-g6 Qd2-f2 Nb6-c4
Bf1xc4 Be6xc4 a2-a3 Nf6-d7 Nc3-d5 Bc4xd5 Rd1xd5 a6-a5 h4-h5 a5-a4 h5xg6
f7xg6 Qf2-h2 Nd7-f6) +0.44/13 6} Rfd8 {(Rf8-d8 Bf1-d3 Nb6-c4 Bd3xc4 Be6xc4
g4-g5 Nf6-h5 Qd2-f2 Rd8-c8 Nc3-a4 Ra8-b8 Na4-b6 Rc8-d8 Rh1-g1 Bc4-e6 Nb6-d5
Be6xd5 Rd1xd5 Rd8-c8) -0.14/19 7} 14. g5 {(g4-g5 Nf6-d7) +0.34/13 15} Nfd7
{(Nf6-d7 Nc3-d5 Nb6xd5 e4xd5 Be6-f5 Rd1-c1 Rd8-c8 h4-h5 Nd7-b6 h5-h6 Nb6-c4
Bf1xc4 Qc7xc4 h6xg7) -0.36/21 38} 15. Qe1 {(Qd2-e1 f7-f6 Qe1-f2 Nb6-c4
Bf1xc4 Qc7xc4 Nc3-d5 Be6xd5 Rd1xd5 b7-b5 c2-c3 a6-a5 Qf2-g3 a5-a4 Nb3-d2
Qc4-c6 g5-g6 Nd7-b6 Be3xb6 Qc6xb6 h4-h5) +0.62/13 9} Rdc8 {(Rd8-c8 Bf1-d3
a6-a5 a2-a4 Be6-c4 Qe1-e2 Qc7-c6 Be3xb6) +0.12/16 3} 16. h5 {(h4-h5 Nb6-c4
Bf1xc4 Qc7xc4 Rh1-h2 a6-a5 g5-g6 f7xg6 h5xg6 h7-h6 Nc3-d5 Be7-d8 Rd1-d3
Be6xd5 Rd3xd5 a5-a4 Nb3-a5 Bd8xa5) +0.49/13 5} Nc4 {(Nb6-c4 Bf1xc4 Qc7xc4
g5-g6 b7-b5 g6xh7+ Kg8xh7 Qe1-d2 a6-a5 Rd1-g1 b5-b4 Qd2-g2 Qc4xb3 c2xb3
b4xc3 b2xc3 Be6xb3 a2xb3 a5-a4) -0.11/21 32} 17. Bxc4 {(Bf1xc4) +0.70/13
11} Qxc4 {(Qc7xc4 Qe1-f1 Qc4xf1 Rh1xf1 a6-a5 Nc3-d5 Be7-f8 Nb3-d2 a5-a4
a2-a3 Ra8-a5 c2-c4 b7-b5 Rd1-c1 Nd7-c5 Be3xc5 d6xc5 c4xb5 Ra5xb5) -0.24/19
3} 18. g6 {(g5-g6 f7xg6 h5xg6 h7xg6 Qe1-g3 Nd7-f8 Be3-g5 Be7xg5 Qg3xg5
Qc4-c7 Qg5-d2 Qc7-d8 Qd2xd6 Qd8xd6 Rd1xd6 Rc8-c6 Rd6-d3 Be6xb3 a2xb3 Kg8-f7
Kb1-a2 g6-g5) +0.86/13 8} hxg6 {(h7xg6 h5xg6 f7xg6 Qe1-g3 Nd7-f8 Qg3-h2
b7-b5 a2-a3 Ra8-b8 Qh2-h8+ Kg8-f7 Rd1-g1 b5-b4 Nb3-d2 Qc4-c7 a3xb4 Rb8xb4
Nd2-b3) -0.16/18 5} 19. hxg6 {(h5xg6 f7xg6 Qe1-g3 Nd7-f8 Qg3-h2 b7-b5 a2-a3
Kg8-f7 f3-f4 Nf8-d7) +0.71/14 10} fxg6 {(f7xg6 Qe1-g3 Nd7-f8 Qg3-h2 b7-b5
f3-f4 e5xf4 Qh2-h8+ Kg8-f7 Be3-d4 b5-b4 Qh8xg7+ Kf7-e8 Nb3-d2 Qc4-c7 Rh1-h6
b4xc3 Qg7xe7+) -0.64/21 29} 20. Qg3 {(Qe1-g3 Nd7-f8 Qg3-h2 Nf8-d7 Qh2-g2
Be6-f7 Nb3-d2 Qc4-e6 Qg2-h2 Kg8-f8 Qh2-h8+ Bf7-g8 Nd2-b3 Nd7-f6 a2-a3
Kf8-f7 Qh8-h2 Kf7-f8 Qh2-g2 Bg8-h7 Nb3-a5) +0.72/15 15} Nf8 {(Nd7-f8 Qg3-h2
b7-b5 f3-f4 e5xf4 Qh2-h8+ Kg8-f7 Be3-d4 b5-b4 Qh8xg7+ Kf7-e8 Nb3-d2 Qc4-c7
Rh1-h6 b4xc3 Qg7xe7+) -0.64/18 5} 21. Qh2 {(Qg3-h2 Kg8-f7 Rd1-f1 Nf8-d7
Qh2-g3) +0.82/14 11} Kf7 {(Kg8-f7 f3-f4 Be7-f6 Rd1xd6 e5xf4 Be3-d4 Bf6-g5
Qh2-h8 Qc4xd4 Nb3xd4 Rc8xc3 b2xc3 Be6xa2+ Kb1xa2) -0.56/17 4} 22. Rdf1
{(Rd1-f1 Nf8-d7 Qh2-g3 a6-a5 Nb3-d2 Qc4-c7 f3-f4 e5xf4 Rf1xf4+ Be7-f6
Nd2-f3 Kf7-g8 Nf3-d4 g6-g5 Rf4xf6 Be6xa2+ Kb1xa2) +0.72/13 5} Bf6 {(Be7-f6
f3-f4 Kf7-e8 f4-f5 Be6-f7 Qh2-g2 g6xf5 Rf1xf5 Qc4-b4 Nc3-d5 Bf7xd5 e4xd5
Qb4xb3 c2xb3) -0.41/16 3} 23. f4 {(f3-f4 Kf7-e7 Qh2-g3 Nf8-d7 Qg3xg6 e5xf4
Be3xf4 Nd7-e5 Bf4xe5 Bf6xe5 Nb3-d2 Qc4-c6 Nc3-d5+ Be6xd5 e4xd5 Qc6xd5
Nd2-e4 Ke7-d8 Rf1-f7) +1.33/13 16} Ke8 {(Kf7-e8 f4-f5 Be6-f7 f5xg6 Nf8xg6
Nb3-d2 Qc4-c6 Qh2-h7 Ng6-f8 Qh7-h3 b7-b5 a2-a3 Nf8-e6 Qh3-h7 Ne6-f8 Qh7-h8
Ke8-d7) -0.35/17 4} 24. fxe5 {(f4xe5 d6xe5 Qh2-h8 Qc4-b4 Rf1-f2 a6-a5 a2-a3
Qb4-e7 Nc3-a4 Be6xb3 c2xb3 Qe7-d6 Rf2-d2 Qd6-c6 Na4-c3 Ke8-f7 Rd2-d5)
+1.46/13 7} Bxe5 {(Bf6xe5 Be3-f4 Be5xc3 b2xc3 Rc8-c7 Bf4xd6 Rc7-f7 Rf1xf7
Ke8xf7 Qh2-h8 Qc4xe4 Nb3-c5 Qe4-c4 Nc5xe6) -0.68/15 3} 25. Bf4 {(Be3-f4
Be5xc3 b2xc3 Rc8-c7 Bf4xd6 Rc7-f7 Qh2-g2 Ra8-c8 Bd6-b4 g6-g5 Rf1-d1 g5-g4
Rd1-d4 Qc4-c7 Rd4-d3 Nf8-g6 Bb4-d6 Qc7-c4 Kb1-a1) +2.02/13 5} Bxc3 {(Be5xc3
b2xc3 Rc8-c7 Bf4xd6 Rc7-f7 Qh2-g2 Rf7xf1+ Rh1xf1 Be6-f7 Rf1-h1 Ra8-d8
Bd6xf8 Ke8xf8 Qg2-g5 Rd8-e8 Rh1-h8+ Bf7-g8 Qg5xg6) -0.83/21 24} 26. bxc3
{(b2xc3) +2.32/14 6} Rc7 {(Rc8-c7 Bf4xd6 Rc7-f7 Qh2-g2 Rf7xf1+ Rh1xf1
Be6-f7 Rf1-h1 Ra8-d8 Bd6xf8 Ke8xf8 Qg2-g5 Rd8-e8 Rh1-h8+ Bf7-g8 Qg5xg6)
-0.83/20 4} 27. Bxd6 {(Bf4xd6 Rc7-f7 Rf1xf7 Be6xf7 Qh2-h8 Bf7-g8 Qh8xg7
Qc4-f7 Bd6xf8 Qf7xf8 Qg7xb7 Ra8-d8 Qb7-c6+ Ke8-e7 Qc6xg6 Qf8-f3 Qg6-g7+
Bg8-f7 Rh1-h7 Qf3-f6 Qg7-g3 Qf6-b6) +4.22/14 4} Rf7 {(Rc7-f7 Rf1xf7 Ke8xf7
Qh2-f4+ Kf7-g8 Bd6-e5 Be6-f5 Rh1-h8+ Kg8-f7 e4xf5 Qc4xf4 Be5xf4 g6xf5
Nb3-d4 g7-g6 Kb1-c1 Ra8-d8 Kc1-b2 Nf8-e6) -3.38/19 2} 28. Rxf7 {(Rf1xf7)
+4.81/14 9} Kxf7 {(Ke8xf7 Qh2-f4+ Kf7-g8 Bd6-e5 Be6-f5 Rh1-h8+ Kg8-f7 e4xf5
Qc4xf4 Be5xf4 g6xf5 Nb3-a5 Ra8-d8 Na5xb7 Rd8-d1+ Kb1-b2 Nf8-g6 Rh8-d8
Rd1xd8) -3.41/19 3} 29. Qf4+ {(Qh2-f4+ Be6-f5 Nb3-d4 Nf8-e6 Nd4xe6 Qc4xe6
e4xf5 g6xf5 Rh1-h5 Kf7-g8 Rh5xf5 Ra8-d8 Bd6-c7 Rd8-e8 Rf5-e5 Qe6-c6 Re5xe8+
Qc6xe8 Bc7-e5 Qe8-g6 Qf4-c4+ Kg8-h7 Qc4-h4+ Kh7-g8 Qh4-d8+ Kg8-h7 Qd8-e7
Qg6-g2) +5.11/14 5} Kg8 {(Kf7-g8 Bd6-e5 Be6-f5 Rh1-h8+ Kg8-f7 e4xf5 Qc4xf4
Be5xf4 g6xf5 Nb3-a5 Ra8-d8 Na5xb7 Rd8-d1+ Kb1-b2 Nf8-g6 Rh8-d8 Rd1xd8
Nb7xd8+ Kf7-e8) -3.41/20 4} 30. Be5 {(Bd6-e5 Nf8-h7 Qf4-h2 Kg8-f7 Qh2xh7
Ra8-g8 Rh1-h4 Qc4-b5 Rh4-f4+ Kf7-e7 Be5xg7 Ke7-d8 Qh7xg6 Qb5-e8 Qg6-h7
Qe8-d7 Bg7-f6+) +6.74/15 8} Bf5 {(Be6-f5 Rh1-h8+ Kg8-f7 e4xf5 Qc4xf4 Be5xf4
g6xf5 Nb3-a5 Ra8-d8 Na5xb7 Rd8-d1+ Kb1-b2 Nf8-g6 Rh8-h5 Rd1-d5 Nb7-d6+
Kf7-e6) -3.41/18 4} 31. Nd4 {(Nb3-d4) +8.16/14 5} Qxc3 {(Qc4xc3 Nd4xf5
Qc3-b4+ Be5-b2 g6xf5 Qf4-e5) -5.50/20 20} 32. exf5 {(e4xf5) +8.20/13 3}
Qb4+ {(Qc3-b4+ Kb1-c1 g6-g5 Qf4xg5 Kg8-f7 Qg5xg7+ Kf7-e8 Nd4-e6 Ra8-c8
Ne6xf8 Qb4-e7 Qg7xe7+ Ke8xe7 Nf8-g6+ Ke7-d7 Rh1-h7+ Kd7-e8) -9.42/18 18}
33. Kc1 {(Kb1-c1) +9.41/14 5} Qe7 {(Qb4-e7 f5-f6 g7xf6 Be5xf6 Qe7-h7 Rh1xh7
Nf8xh7 Nd4-e6 Ra8-c8 Qf4-e4 Kg8-f7 Ne6-g5+ Kf7xf6 Ng5xh7+ Kf6-g7 Qe4xb7+
Kg7-h6 Qb7xc8) -11.22/18 12} 34. f6 {(f5-f6) +12.77/13 3} gxf6 {(g7xf6
Be5xf6 Qe7-a3+ Kc1-d1 Ra8-d8 Rh1-h8+ Kg8-f7 Bf6xd8+ Kf7-e8 Qf4-e5+ Ke8xd8
Nd4-e6+ Kd8-d7 Qe5-d5+ Kd7-e7 Ne6xf8 Qa3-f3+ Qd5xf3 Ke7-d6) -M16/19 15} 35.
Bxf6 {(Be5xf6) +14.11/15 4 Arena Adjudication} 1-0
[/pgn]
-
connor_mcmonigle
- Posts: 530
- Joined: Sun Sep 06, 2020 4:40 am
- Full name: Connor McMonigle
Re: New engine releases & news 2021
What do you mean by NNUE? I believe you're using "uses NNUE" to mean "copies NN code from Stockfish" while Gabor is using "uses NNUE" to mean "uses an Efficiently Updated Neural Network". However, you've previously referred to Seer as an NNUE engine. Seer does not copy any code from Stockfish and it relies on a network with a somewhat different topology. In any case, I'm hopeful to see the term "NNUE" die as there's no consensus as to what it means. The term "NNUE"'s continued usage serves only to induce confusion within many of the individuals on this forum.Sylwy wrote: ↑Sat Jul 24, 2021 9:19 pmNNUE or NN ? The executable is small .......and nothing impressive in the play of this Rust engine. Seems just a small integrated NN.Gabor Szots wrote: ↑Sat Jul 24, 2021 10:44 am Velvet 2.0.0
New is that it uses NNUE.
https://github.com/mhonert/velvet-chess ... tag/v2.0.0
Ktulu 9 is an old 32-bit HCE chess engine.
[pgn]
[Event "NN Test121"]
[Site "ISR 3"]
[Date "2021.07.24"]
[Round "2"]
[White "Ktulu 9 w32"]
[Black "Velvet-v2.0.0-x86_64-avx2"]
[Result "1-0"]
[BlackElo "2200"]
[ECO "B90"]
[Opening "Sicilian"]
[Time "22:00:22"]
[Variation "Najdorf, 6.Be3 e5 7.Nb3 Be6 8.f3 Nbd7"]
[WhiteElo "2200"]
[TimeControl "240+2"]
[Termination "adjudication"]
[PlyCount "69"]
[WhiteType "program"]
[BlackType "program"]
1. Nc3 c5 2. e4 d6 3. Nf3 Nf6 4. d4 cxd4 5. Nxd4 a6 6. Be3 e5 7. Nb3 Be6 8.
f3 Nbd7 9. g4 Be7 {(Bf8-e7 g4-g5 Nf6-h5 Qd1-d2 O-O O-O-O Ra8-c8 Kc1-b1
Nh5-f4 h2-h4 Qd8-c7 Nc3-d5 Be6xd5 e4xd5 Nd7-b6 Be3xf4) +0.03/17 8} 10. Qd2
O-O {(O-O Rh1-g1 Qd8-c7 O-O-O Rf8-c8 g4-g5 Nf6-h5 Nc3-d5 Be6xd5 e4xd5
Nh5-f4 c2-c3 Nd7-b6 Be3xf4 e5xf4 Qd2xf4) +0.07/16 3} 11. O-O-O Qc7 {(Qd8-c7
g4-g5 Nf6-h5 Nc3-d5 Be6xd5 Qd2xd5 Nh5-f4 Qd5-a5 Qc7-c6 Qa5-c3 Qc6-a4 Qc3-a5
Qa4-c6 Qa5-c3 Qc6-a4 Qc3-a5 Qa4-c6 Qa5-c3) 0.00/18 5} 12. Kb1 Nb6 {(Nd7-b6
h2-h4 Rf8-d8 g4-g5 Nf6-h5 Nc3-d5 Nb6xd5 e4xd5 Be6-f5 Rh1-g1 Rd8-e8 Bf1-d3
Bf5-g6 Bd3-e4 Nh5-f4 Be4xg6) -0.12/17 8} 13. h4 {(h2-h4 g7-g6 Qd2-f2 Nb6-c4
Bf1xc4 Be6xc4 a2-a3 Nf6-d7 Nc3-d5 Bc4xd5 Rd1xd5 a6-a5 h4-h5 a5-a4 h5xg6
f7xg6 Qf2-h2 Nd7-f6) +0.44/13 6} Rfd8 {(Rf8-d8 Bf1-d3 Nb6-c4 Bd3xc4 Be6xc4
g4-g5 Nf6-h5 Qd2-f2 Rd8-c8 Nc3-a4 Ra8-b8 Na4-b6 Rc8-d8 Rh1-g1 Bc4-e6 Nb6-d5
Be6xd5 Rd1xd5 Rd8-c8) -0.14/19 7} 14. g5 {(g4-g5 Nf6-d7) +0.34/13 15} Nfd7
{(Nf6-d7 Nc3-d5 Nb6xd5 e4xd5 Be6-f5 Rd1-c1 Rd8-c8 h4-h5 Nd7-b6 h5-h6 Nb6-c4
Bf1xc4 Qc7xc4 h6xg7) -0.36/21 38} 15. Qe1 {(Qd2-e1 f7-f6 Qe1-f2 Nb6-c4
Bf1xc4 Qc7xc4 Nc3-d5 Be6xd5 Rd1xd5 b7-b5 c2-c3 a6-a5 Qf2-g3 a5-a4 Nb3-d2
Qc4-c6 g5-g6 Nd7-b6 Be3xb6 Qc6xb6 h4-h5) +0.62/13 9} Rdc8 {(Rd8-c8 Bf1-d3
a6-a5 a2-a4 Be6-c4 Qe1-e2 Qc7-c6 Be3xb6) +0.12/16 3} 16. h5 {(h4-h5 Nb6-c4
Bf1xc4 Qc7xc4 Rh1-h2 a6-a5 g5-g6 f7xg6 h5xg6 h7-h6 Nc3-d5 Be7-d8 Rd1-d3
Be6xd5 Rd3xd5 a5-a4 Nb3-a5 Bd8xa5) +0.49/13 5} Nc4 {(Nb6-c4 Bf1xc4 Qc7xc4
g5-g6 b7-b5 g6xh7+ Kg8xh7 Qe1-d2 a6-a5 Rd1-g1 b5-b4 Qd2-g2 Qc4xb3 c2xb3
b4xc3 b2xc3 Be6xb3 a2xb3 a5-a4) -0.11/21 32} 17. Bxc4 {(Bf1xc4) +0.70/13
11} Qxc4 {(Qc7xc4 Qe1-f1 Qc4xf1 Rh1xf1 a6-a5 Nc3-d5 Be7-f8 Nb3-d2 a5-a4
a2-a3 Ra8-a5 c2-c4 b7-b5 Rd1-c1 Nd7-c5 Be3xc5 d6xc5 c4xb5 Ra5xb5) -0.24/19
3} 18. g6 {(g5-g6 f7xg6 h5xg6 h7xg6 Qe1-g3 Nd7-f8 Be3-g5 Be7xg5 Qg3xg5
Qc4-c7 Qg5-d2 Qc7-d8 Qd2xd6 Qd8xd6 Rd1xd6 Rc8-c6 Rd6-d3 Be6xb3 a2xb3 Kg8-f7
Kb1-a2 g6-g5) +0.86/13 8} hxg6 {(h7xg6 h5xg6 f7xg6 Qe1-g3 Nd7-f8 Qg3-h2
b7-b5 a2-a3 Ra8-b8 Qh2-h8+ Kg8-f7 Rd1-g1 b5-b4 Nb3-d2 Qc4-c7 a3xb4 Rb8xb4
Nd2-b3) -0.16/18 5} 19. hxg6 {(h5xg6 f7xg6 Qe1-g3 Nd7-f8 Qg3-h2 b7-b5 a2-a3
Kg8-f7 f3-f4 Nf8-d7) +0.71/14 10} fxg6 {(f7xg6 Qe1-g3 Nd7-f8 Qg3-h2 b7-b5
f3-f4 e5xf4 Qh2-h8+ Kg8-f7 Be3-d4 b5-b4 Qh8xg7+ Kf7-e8 Nb3-d2 Qc4-c7 Rh1-h6
b4xc3 Qg7xe7+) -0.64/21 29} 20. Qg3 {(Qe1-g3 Nd7-f8 Qg3-h2 Nf8-d7 Qh2-g2
Be6-f7 Nb3-d2 Qc4-e6 Qg2-h2 Kg8-f8 Qh2-h8+ Bf7-g8 Nd2-b3 Nd7-f6 a2-a3
Kf8-f7 Qh8-h2 Kf7-f8 Qh2-g2 Bg8-h7 Nb3-a5) +0.72/15 15} Nf8 {(Nd7-f8 Qg3-h2
b7-b5 f3-f4 e5xf4 Qh2-h8+ Kg8-f7 Be3-d4 b5-b4 Qh8xg7+ Kf7-e8 Nb3-d2 Qc4-c7
Rh1-h6 b4xc3 Qg7xe7+) -0.64/18 5} 21. Qh2 {(Qg3-h2 Kg8-f7 Rd1-f1 Nf8-d7
Qh2-g3) +0.82/14 11} Kf7 {(Kg8-f7 f3-f4 Be7-f6 Rd1xd6 e5xf4 Be3-d4 Bf6-g5
Qh2-h8 Qc4xd4 Nb3xd4 Rc8xc3 b2xc3 Be6xa2+ Kb1xa2) -0.56/17 4} 22. Rdf1
{(Rd1-f1 Nf8-d7 Qh2-g3 a6-a5 Nb3-d2 Qc4-c7 f3-f4 e5xf4 Rf1xf4+ Be7-f6
Nd2-f3 Kf7-g8 Nf3-d4 g6-g5 Rf4xf6 Be6xa2+ Kb1xa2) +0.72/13 5} Bf6 {(Be7-f6
f3-f4 Kf7-e8 f4-f5 Be6-f7 Qh2-g2 g6xf5 Rf1xf5 Qc4-b4 Nc3-d5 Bf7xd5 e4xd5
Qb4xb3 c2xb3) -0.41/16 3} 23. f4 {(f3-f4 Kf7-e7 Qh2-g3 Nf8-d7 Qg3xg6 e5xf4
Be3xf4 Nd7-e5 Bf4xe5 Bf6xe5 Nb3-d2 Qc4-c6 Nc3-d5+ Be6xd5 e4xd5 Qc6xd5
Nd2-e4 Ke7-d8 Rf1-f7) +1.33/13 16} Ke8 {(Kf7-e8 f4-f5 Be6-f7 f5xg6 Nf8xg6
Nb3-d2 Qc4-c6 Qh2-h7 Ng6-f8 Qh7-h3 b7-b5 a2-a3 Nf8-e6 Qh3-h7 Ne6-f8 Qh7-h8
Ke8-d7) -0.35/17 4} 24. fxe5 {(f4xe5 d6xe5 Qh2-h8 Qc4-b4 Rf1-f2 a6-a5 a2-a3
Qb4-e7 Nc3-a4 Be6xb3 c2xb3 Qe7-d6 Rf2-d2 Qd6-c6 Na4-c3 Ke8-f7 Rd2-d5)
+1.46/13 7} Bxe5 {(Bf6xe5 Be3-f4 Be5xc3 b2xc3 Rc8-c7 Bf4xd6 Rc7-f7 Rf1xf7
Ke8xf7 Qh2-h8 Qc4xe4 Nb3-c5 Qe4-c4 Nc5xe6) -0.68/15 3} 25. Bf4 {(Be3-f4
Be5xc3 b2xc3 Rc8-c7 Bf4xd6 Rc7-f7 Qh2-g2 Ra8-c8 Bd6-b4 g6-g5 Rf1-d1 g5-g4
Rd1-d4 Qc4-c7 Rd4-d3 Nf8-g6 Bb4-d6 Qc7-c4 Kb1-a1) +2.02/13 5} Bxc3 {(Be5xc3
b2xc3 Rc8-c7 Bf4xd6 Rc7-f7 Qh2-g2 Rf7xf1+ Rh1xf1 Be6-f7 Rf1-h1 Ra8-d8
Bd6xf8 Ke8xf8 Qg2-g5 Rd8-e8 Rh1-h8+ Bf7-g8 Qg5xg6) -0.83/21 24} 26. bxc3
{(b2xc3) +2.32/14 6} Rc7 {(Rc8-c7 Bf4xd6 Rc7-f7 Qh2-g2 Rf7xf1+ Rh1xf1
Be6-f7 Rf1-h1 Ra8-d8 Bd6xf8 Ke8xf8 Qg2-g5 Rd8-e8 Rh1-h8+ Bf7-g8 Qg5xg6)
-0.83/20 4} 27. Bxd6 {(Bf4xd6 Rc7-f7 Rf1xf7 Be6xf7 Qh2-h8 Bf7-g8 Qh8xg7
Qc4-f7 Bd6xf8 Qf7xf8 Qg7xb7 Ra8-d8 Qb7-c6+ Ke8-e7 Qc6xg6 Qf8-f3 Qg6-g7+
Bg8-f7 Rh1-h7 Qf3-f6 Qg7-g3 Qf6-b6) +4.22/14 4} Rf7 {(Rc7-f7 Rf1xf7 Ke8xf7
Qh2-f4+ Kf7-g8 Bd6-e5 Be6-f5 Rh1-h8+ Kg8-f7 e4xf5 Qc4xf4 Be5xf4 g6xf5
Nb3-d4 g7-g6 Kb1-c1 Ra8-d8 Kc1-b2 Nf8-e6) -3.38/19 2} 28. Rxf7 {(Rf1xf7)
+4.81/14 9} Kxf7 {(Ke8xf7 Qh2-f4+ Kf7-g8 Bd6-e5 Be6-f5 Rh1-h8+ Kg8-f7 e4xf5
Qc4xf4 Be5xf4 g6xf5 Nb3-a5 Ra8-d8 Na5xb7 Rd8-d1+ Kb1-b2 Nf8-g6 Rh8-d8
Rd1xd8) -3.41/19 3} 29. Qf4+ {(Qh2-f4+ Be6-f5 Nb3-d4 Nf8-e6 Nd4xe6 Qc4xe6
e4xf5 g6xf5 Rh1-h5 Kf7-g8 Rh5xf5 Ra8-d8 Bd6-c7 Rd8-e8 Rf5-e5 Qe6-c6 Re5xe8+
Qc6xe8 Bc7-e5 Qe8-g6 Qf4-c4+ Kg8-h7 Qc4-h4+ Kh7-g8 Qh4-d8+ Kg8-h7 Qd8-e7
Qg6-g2) +5.11/14 5} Kg8 {(Kf7-g8 Bd6-e5 Be6-f5 Rh1-h8+ Kg8-f7 e4xf5 Qc4xf4
Be5xf4 g6xf5 Nb3-a5 Ra8-d8 Na5xb7 Rd8-d1+ Kb1-b2 Nf8-g6 Rh8-d8 Rd1xd8
Nb7xd8+ Kf7-e8) -3.41/20 4} 30. Be5 {(Bd6-e5 Nf8-h7 Qf4-h2 Kg8-f7 Qh2xh7
Ra8-g8 Rh1-h4 Qc4-b5 Rh4-f4+ Kf7-e7 Be5xg7 Ke7-d8 Qh7xg6 Qb5-e8 Qg6-h7
Qe8-d7 Bg7-f6+) +6.74/15 8} Bf5 {(Be6-f5 Rh1-h8+ Kg8-f7 e4xf5 Qc4xf4 Be5xf4
g6xf5 Nb3-a5 Ra8-d8 Na5xb7 Rd8-d1+ Kb1-b2 Nf8-g6 Rh8-h5 Rd1-d5 Nb7-d6+
Kf7-e6) -3.41/18 4} 31. Nd4 {(Nb3-d4) +8.16/14 5} Qxc3 {(Qc4xc3 Nd4xf5
Qc3-b4+ Be5-b2 g6xf5 Qf4-e5) -5.50/20 20} 32. exf5 {(e4xf5) +8.20/13 3}
Qb4+ {(Qc3-b4+ Kb1-c1 g6-g5 Qf4xg5 Kg8-f7 Qg5xg7+ Kf7-e8 Nd4-e6 Ra8-c8
Ne6xf8 Qb4-e7 Qg7xe7+ Ke8xe7 Nf8-g6+ Ke7-d7 Rh1-h7+ Kd7-e8) -9.42/18 18}
33. Kc1 {(Kb1-c1) +9.41/14 5} Qe7 {(Qb4-e7 f5-f6 g7xf6 Be5xf6 Qe7-h7 Rh1xh7
Nf8xh7 Nd4-e6 Ra8-c8 Qf4-e4 Kg8-f7 Ne6-g5+ Kf7xf6 Ng5xh7+ Kf6-g7 Qe4xb7+
Kg7-h6 Qb7xc8) -11.22/18 12} 34. f6 {(f5-f6) +12.77/13 3} gxf6 {(g7xf6
Be5xf6 Qe7-a3+ Kc1-d1 Ra8-d8 Rh1-h8+ Kg8-f7 Bf6xd8+ Kf7-e8 Qf4-e5+ Ke8xd8
Nd4-e6+ Kd8-d7 Qe5-d5+ Kd7-e7 Ne6xf8 Qa3-f3+ Qd5xf3 Ke7-d6) -M16/19 15} 35.
Bxf6 {(Be5xf6) +14.11/15 4 Arena Adjudication} 1-0
[/pgn]
I looked through the source code and Velvet indeed uses a neural network which is updated efficiently. Surely their engine would be significantly stronger had they had somehow copied thousands of lines of code from Stockfish, but the author opted to take an alternative path. Kudos to the author for doing something original!
-
mvanthoor
- Posts: 1784
- Joined: Wed Jul 03, 2019 4:42 pm
- Location: Netherlands
- Full name: Marcel Vanthoor
Re: New engine releases & news 2021
Dude. Be fair. I didn't know on top of my head where Ktulu is in the CCRL list, and I never heard of Velvet. I just downloaded it and started a test against current Rustic dev (~2150) and Minimalchess 0.5.0, because you make it sound as if this engine is some sort of crappy startup.
Code: Select all
Rank Name Elo +/- Games Score Draw
0 Velvet 2.0.0 720 nan 96 98.4% 3.1%
1 MinimalChess 0.5.0 -672 nan 49 2.0% 4.1%
2 Rustic Alpha 3.4.100 -787 nan 47 1.1% 2.1%
96 of 2000 games finished.
Thanks for mentioning this not-so-impressive engine. I'm going to add it to my list of engines to use for future testing.
Needless to say that I didn't leave the test to finish. It's no use having two engines steam-waltzed for 4000 games.
Can't even copy code without rewriting it... this engine is written in Rust. (One could probably use Daniel Shawul's probing code, and then interface with that.)connor_mcmonigle wrote: ↑Sat Jul 24, 2021 10:30 pm I looked through the source code and Velvet indeed uses a neural network which is updated efficiently. Surely their engine would be significantly stronger had they had somehow copied thousands of lines of code from Stockfish, but the author opted to take an alternative path. Kudos to the author for doing something original!
-
Sylwy
- Posts: 4465
- Joined: Fri Apr 21, 2006 4:19 pm
- Location: IASI - the historical capital of MOLDOVA
- Full name: SilvianR
Re: New engine releases & news 2021
1.-Here are listed by chessprogramming the engines they use NNUE (Efficiently Updatable Neural Networks):connor_mcmonigle wrote: ↑Sat Jul 24, 2021 10:30 pm
1.-What do you mean by NNUE? I believe you're using "uses NNUE" to mean "copies NN code from Stockfish" while Gabor is using "uses NNUE" to mean "uses an Efficiently Updated Neural Network". However, you've previously referred to Seer as an NNUE engine. Seer does not copy any code from Stockfish and it relies on a network with a somewhat different topology. In any case, I'm hopeful to see the term "NNUE" die as there's no consensus as to what it means. The term "NNUE"'s continued usage serves only to induce confusion within many of the individuals on this forum.
2.-I looked through the source code and Velvet indeed uses a neural network which is updated efficiently. Surely their engine would be significantly stronger had they had somehow copied thousands of lines of code from Stockfish, but the author opted to take an alternative path. Kudos to the author for doing something original!
https://www.chessprogramming.org/Category:NNUE
Of course, there are enough original approaches. SF is just one. As you can see there are no engines listed like Winter or SlowChess Blitz. I know Winter uses a neural network with two main parts. The first part is a non-standard convolutional neural network which uses sparsity similarly to NNUE. The second part is a fully connected network which has as input the output of the convolutional network as well as a set of handcrafted features standard to classical engines. It was not clear to me what type of approach Velvet uses.
2.-Thank you for clarification !
-
connor_mcmonigle
- Posts: 530
- Joined: Sun Sep 06, 2020 4:40 am
- Full name: Connor McMonigle
Re: New engine releases & news 2021
I've released a new version of Seer here: https://github.com/connormcmonigle/seer ... tag/v2.2.0
Many thanks to all those who tested Seer v2.1.0. This release comes just two weeks after the previous release as I've set my criterion for releasing a new version at +100 elo in self play with the 4moves_noob opening book. While this book considerably inflates elo, as Seer continues to progress, satisfying this criterion will become increasingly difficult hopefully yielding a more reasonable release cycle.
This version is expected to be about 70 elo stronger than v2.1.0 under CCRL conditions and tested at > +100 elo with the 4moves_noob opening book on OpenBench.
Many thanks to all those who tested Seer v2.1.0. This release comes just two weeks after the previous release as I've set my criterion for releasing a new version at +100 elo in self play with the 4moves_noob opening book. While this book considerably inflates elo, as Seer continues to progress, satisfying this criterion will become increasingly difficult hopefully yielding a more reasonable release cycle.
This version is expected to be about 70 elo stronger than v2.1.0 under CCRL conditions and tested at > +100 elo with the 4moves_noob opening book on OpenBench.
-
AlexChess
- Posts: 1486
- Joined: Sat Feb 06, 2021 8:06 am
- Full name: Alex Morales
Re: New engine releases & news 2021
Wow!
Just finished to test 2.1.0 (not 2.0.1) together with other promising engines not Stockfish derivatives like Minic, Stash and Zahak.
https://www.chess.com/forum/view/game-a ... e-fritz-17
I'll update the results soon
Best regards, AlexChess
Just finished to test 2.1.0 (not 2.0.1) together with other promising engines not Stockfish derivatives like Minic, Stash and Zahak.
https://www.chess.com/forum/view/game-a ... e-fritz-17
I'll update the results soon
Best regards, AlexChess
Chess engines and dedicated chess computers fan since 1981 Mac mini M1 8GB-256GB, Windows 11 & Ubuntu ARM64.
ProteusSF Dev Forum: https://shorturl.at/bpvI1
Mac mini M1 8GB-256GB, Windows 11 & Ubuntu ARM64.ProteusSF Dev Forum: https://shorturl.at/bpvI1