Nemorino 6 (NNUE)

[dkappe]

...and the likes of Andrew Grant have been complaining about for the past few weeks here and in the TCEC and OpenBench communities about engines copying Stockfish code and using Stockfish to help train their nets.

It would be noteworthy if Andy didn’t complain about something. Count down to Andy’s snark response...3...2...1

[Angle]

Nemorino 6:
- Uses a net (nemorino_net_6.nnue) trained with training data generated with Nemorino partly using nets from Stockfish

Nemorino 6.03:
- Uses a net (nemorino_net_13.nnue) trained with training data generated exclusively from Nemorino classic eval
- Uses flip instead of rotate
- Is much weaker than Nemorino 6.00 (but much stronger than any Nemorino 5.* version)
- Was created in relatively short time for TCEC FRC event

Hello, Florentino!
May I use version 6.03 with net_13 together with ver. 6.00 with net_6 in my tournaments and report results on this forum? Also I have a problem with downloading your dev. builds from bitbucket page: every time I get a message ''Project not found or access denied'' at AppVeyor page, Two months ago everything was good with downloading.

[jshriver]

I enjoyed watching this engine play. Here is a game I just ran against Fat Fritz and Nemorino won! i5 4 threads 8gig memory, Fat Fritz was using my nvidia 1650 gpu.

[pgn]
[Event "ARTEMIS, Blitz 15.0min"]
[Site "Morgantown"]
[Date "2020.12.07"]
[Round "1"]
[White "Nemorino"]
[Black "Fat Fritz"]
[Result "1-0"]
[ECO "C45"]
[TimeControl "900"]
[Annotator "0.29;0.24"]
[PlyCount "133"]

1.e4 e5 2.Nf3 Nc6 3.d4 exd4 4.Nxd4 Bb4+ 5.c3 Bc5 6.Be3 Bb6 7.g3 Nf6 8.Bg2 O-O 9.O-O Re8 10.Nd2 d5 11.Nxc6 bxc6 12.Re1 Bg4 13.f3 Be6 14.e5 Nd7 15.f4 f6 16.exf6 Nxf6 17.Bxb6 axb6 18.Nf3 Bg4 19.a4 Qd6 20.Rxe8+ Rxe8 21.Qd3 Bxf3 22.Bxf3 c5 23.Rd1 c6 24.Kg2 Qe6 25.b4 h5 26.a5 bxa5 27.bxa5 h4 28.a6 c4 29.Qd4 h3+ 30.Kf2 Ra8 31.a7 Qd7 32.Ra1 Qb7 33.Ra2 Qb3 34.Qd2 Qb7 35.Qe3 Qb3 36.Ra6 Qb2+ 37.Qe2 Qxc3 38.g4 Kf8 39.g5 Qd4+ 40.Qe3 Qxe3+ 41.Kxe3 Ne8 42.f5 Nc7 43.Rxc6 Rxa7 44.Kf4 Ne8 45.Bxd5 Ra2 46.Kg3 Rd2 47.Bxc4 Rg2+ 48.Kh4 Rxh2 49.Bf1 Ke7 50.Re6+ Kf8 51.Bxh3 Rh1 52.Rb6 Kf7 53.Rb7+ Kf8 54.Rd7 Rh2 55.Kg4 Rc2 56.Bf1 Rc7 57.Rd4 Rc6 58.Bb5 Rb6 59.Bxe8 Kxe8 60.Kh5 Ke7 61.Ra4 Kf7 62.Ra7+ Kf8 63.Ra8+ Kf7 64.g6+ Ke7 65.Rg8 Kf6 66.Rf8+ Ke5 67.f6 1-0

[/pgn]

[Chessqueen]

I enjoyed watching this engine play. Here is a game I just ran against Fat Fritz and Nemorino won! i5 4 threads 8gig memory, Fat Fritz was using my nvidia 1650 gpu.

[pgn]
[Event "ARTEMIS, Blitz 15.0min"]
[Site "Morgantown"]
[Date "2020.12.07"]
[Round "1"]
[White "Nemorino"]
[Black "Fat Fritz"]
[Result "1-0"]
[ECO "C45"]
[TimeControl "900"]
[Annotator "0.29;0.24"]
[PlyCount "133"]

1.e4 e5 2.Nf3 Nc6 3.d4 exd4 4.Nxd4 Bb4+ 5.c3 Bc5 6.Be3 Bb6 7.g3 Nf6 8.Bg2 O-O 9.O-O Re8 10.Nd2 d5 11.Nxc6 bxc6 12.Re1 Bg4 13.f3 Be6 14.e5 Nd7 15.f4 f6 16.exf6 Nxf6 17.Bxb6 axb6 18.Nf3 Bg4 19.a4 Qd6 20.Rxe8+ Rxe8 21.Qd3 Bxf3 22.Bxf3 c5 23.Rd1 c6 24.Kg2 Qe6 25.b4 h5 26.a5 bxa5 27.bxa5 h4 28.a6 c4 29.Qd4 h3+ 30.Kf2 Ra8 31.a7 Qd7 32.Ra1 Qb7 33.Ra2 Qb3 34.Qd2 Qb7 35.Qe3 Qb3 36.Ra6 Qb2+ 37.Qe2 Qxc3 38.g4 Kf8 39.g5 Qd4+ 40.Qe3 Qxe3+ 41.Kxe3 Ne8 42.f5 Nc7 43.Rxc6 Rxa7 44.Kf4 Ne8 45.Bxd5 Ra2 46.Kg3 Rd2 47.Bxc4 Rg2+ 48.Kh4 Rxh2 49.Bf1 Ke7 50.Re6+ Kf8 51.Bxh3 Rh1 52.Rb6 Kf7 53.Rb7+ Kf8 54.Rd7 Rh2 55.Kg4 Rc2 56.Bf1 Rc7 57.Rd4 Rc6 58.Bb5 Rb6 59.Bxe8 Kxe8 60.Kh5 Ke7 61.Ra4 Kf7 62.Ra7+ Kf8 63.Ra8+ Kf7 64.g6+ Ke7 65.Rg8 Kf6 66.Rf8+ Ke5 67.f6 1-0

[/pgn]

Here is a position in which most human GMs will probably move its White King TO b6 and later promote, here comes the Big Surprise Nemorino promote to Queen and Ethereal resigned after thinking for 1 minute

[d]2r5/P6k/3R4/1K6/4P3/8/8/5B2 w - - 0 1

[JohnW]

Funny that when I register the engine in the Fritz 17 gui it says the name is "6.00 (AVX2/PEXT)" instead of "Nemorino 6 (NNUE)".
Even if I edit the uci file it still includes "6.00 (AVX2/PEXT)" in the name.

[jshriver]

Funny that when I register the engine in the Fritz 17 gui it says the name is "6.00 (AVX2/PEXT)" instead of "Nemorino 6 (NNUE)".
Even if I edit the uci file it still includes "6.00 (AVX2/PEXT)" in the name.

Does the same for me. I played it against Fat Frit and Nemorino crushed it. Maybe I have a weak gpu (GTX 1660) but at one point Nemorino forked Fat Fritz with a knight against queen, rook and a bishop. Talk about a blunder.

[jshriver]

Willing to donate some GPU power to nemorino's development, have a 1660 and 960 at home and access to 2 V100's on google cloud.

[Paloma]

Here is a position in which most human GMs will probably move its White King TO b6 and later promote, here comes the Big Surprise Nemorino promote to Queen and Ethereal resigned after thinking for 1 minute

[d]2r5/P6k/3R4/1K6/4P3/8/8/5B2 w - - 0 1

Yes, 65.a8=Q is a bad move.
[pgn][Event "?"]
[Site "?"]
[Date "????.??.??"]
[Round "?"]
[White "?"]
[Black "?"]
[Result "*"]
[SetUp "1"]
[FEN "r7/2K4k/P2R4/8/4P3/8/2P5/5B2 w - - 0 61"]
[PlyCount "11"]

61. Kb6 $2 {15#} (61. Rd8 {13#} Ra7+ 62. Kb8 Rf7 63. a7 Rxa7 64. Kxa7 Kg7 65.
Rd6 {9#}) 61... Rb8+ 62. Kc6 Rc8+ 63. Kb5 $2 {16#} (63. Kb7 $1 {13#}) 63...
Rxc2 $2 64. a7 {12# (7men)} Rc8 65. a8=Q $4 {20#} (65. e5 $1 {11#} Ra8 66. e6 {
10#}) 65... Rxa8 66. e5 {19# (6men)} *
[/pgn]
61. Kb6? (15#)
(61. Rd8 (13#) Ra7+ 62. Kb8 Rf7 63. a7 Rxa7 64. Kxa7 Kg7 65. Rd6 (9#))
...Rb8+ 62. Kc6 Rc8+ 63. Kb5? 16#
(63. Kb7! (13#))
...Rxc2? 64. a7 (12#, (7men)) Rc8 65. a8=Q ?? (20#)
(65. e5! (11#) Ra8 66. e6 (10#))
65... Rxa8 66. e5 (19#, (6men)

[Florentino]

a8=Q is not a bad move it's simply the best move as it immediately ends the game with a win. As long as tablebases and tablebase adjudication is used, there might be shorter ways to a win than mating.

[carldaman]

a8=Q is not a bad move it's simply the best move as it immediately ends the game with a win. As long as tablebases and tablebase adjudication is used, there might be shorter ways to a win than mating.

...it immediately ends the game with an adjudication win.

Most of us are used to checkmate being needed to win. The flaw is with the TBs lacking DTM, but also with the engine that's unable to search out a checkmate on its own, one that would take precedence over the TBs.

Are we sure that Nemorino v6 isn't buggy in some ways?
CCRL is showing Nemorino v6 4CPU (after 90 games) being close to 200 Elo(!) below v6 single-core and about equal to v5 on 4CPU.

Nemorino 6.00 64-bit 3334 +16 −16 54.8%
Nemorino 5.00 64-bit 4CPU 3148 +18 −18 48.1%
Nemorino 6.00 64-bit 4CPU 3146 +64 −68 22.8%