Minic version 3

[xr_a_y]

Second year anniversary for Minic ! and a new version number 3.00

Porting Stockfish NNUE implementation inside Minic was fun and instructive. This has clearly raised my interest in NN technologies and in this sense is a success.
As you already know, I was never comfortable with the big copy/paste of SF code involved, nor with the use of SF based nets of course.
This is why I first trained a net with my own data (but using Nodchip learner), namely the "napping nexus" net and then merged the learner directly inside Minic in order to train the pure Minic net named "nascent nutrient". Still, this was not enough to make me feel this technology and code can be inside Minic this way and in no way this can be considered as an "official" version of Minic as soon as a NNUE net is in use. This is why I decided to not use SF implementation anymore. This means that, starting from version 3, Minic is no more compatible with SF nets. Only versions 2.47 to 2.53 can read and use SF nets (and only using NNUE version 0x7AF32F16u nets).

In the meantime, I discovered Seer engine by Connor McMonigle (https://github.com/connormcmonigle/seer-nnue), and was very impressed by its code quality. This engine propose a very clear and simple to understand and use NNUE implementation. This implementation allows to very easily change architecture of the net and implements an engine-independent training code using pytorch. I decided to borrow this code base and adapt it to Minic (was very easy). This is a good compromise for me, a code base I can understand and work with, a very good starting point.

I then trained a first net for it, named "nefarious nucleus" (available here : https://github.com/tryingsomestuff/NNUE-Nets) that performs even with standard Minic2.53 eval at short TC (10s+0.1). Maybe better at longer TC...

My hope is of course to train better nets, probably try other architectures and also speed-up the code that is for now quite slow.

Please note that Minic 3.00 with no net loaded is weaker than version 2.53 because it has less nps. So that there is probably no need for testing for now ...

So for now this is a just a new forum thread for a new story

[Guenther]

Second year anniversary for Minic :D ! and a new version number 3.00

Porting Stockfish NNUE implementation inside Minic was fun and instructive. This has clearly raised my interest in NN technologies and in this sense is a success.
As you already know, I was never comfortable with the big copy/paste of SF code involved, nor with the use of SF based nets of course.
This is why I first trained a net with my own data (but using Nodchip learner), namely the "napping nexus" net and then merged the learner directly inside Minic in order to train the pure Minic net named "nascent nutrient". Still, this was not enough to make me feel this technology and code can be inside Minic this way and in no way this can be considered as an "official" version of Minic as soon as a NNUE net is in use. This is why I decided to not use SF implementation anymore. This means that, starting from version 3, Minic is no more compatible with SF nets. Only versions 2.47 to 2.53 can read and use SF nets (and only using NNUE version 0x7AF32F16u nets).

In the meantime, I discovered Seer engine by Connor McMonigle (https://github.com/connormcmonigle/seer-nnue), and was very impressed by its code quality. This engine propose a very clear and simple to understand and use NNUE implementation. This implementation allows to very easily change architecture of the net and implements an engine-independent training code using pytorch. I decided to borrow this code base and adapt it to Minic (was very easy). This is a good compromise for me, a code base I can understand and work with, a very good starting point.

I then trained a first net for it, named "nefarious nucleus" (available here : https://github.com/tryingsomestuff/NNUE-Nets) that performs even with standard Minic2.53 eval at short TC (10s+0.1). Maybe better at longer TC...

My hope is of course to train better nets, probably try other architectures and also speed-up the code that is for now quite slow.

Please note that Minic 3.00 with no net loaded is weaker than version 2.53 because it has less nps. So that there is probably no need for testing for now ...

So for now this is a just a new forum thread for a new story ;-)

Very interesting!

I guess I cannot work around a new 'field' anymore in my XB/UCI chronology for 'nn type' and/or 'nn file(s) itself' and perhaps how it was learnt,
what I avoided so far. (Of course this is not meant as criticism on your decisions - more a general remark)
This will be probably a challenge for rating lists in the future too.

Any suggestions welcome...

[xr_a_y]

Fun game of Minic3 using "nefarious nucleus" net versus Weiss. Minic giving everything for the king side attack

[pgn]
[Event "My Tournament"]
[Site "?"]
[Date "2020.11.03"]
[Round "395"]
[White "minic_3.00_nefarious nucleus"]
[Black "weiss"]
[Result "1-0"]
[ECO "B25"]
[GameDuration "00:00:17"]
[GameEndTime "2020-11-03T17:07:04.166 CET"]
[GameStartTime "2020-11-03T17:06:46.869 CET"]
[Opening "Sicilian"]
[PlyCount "59"]
[TimeControl "10+0.1"]
[Variation "Closed"]

1. e4 {book} c5 {book} 2. Nc3 {book} Nc6 {book} 3. g3 {book} g6 {book}
4. Bg2 {book} Bg7 {book} 5. d3 {book} d6 {book} 6. Nge2 {book}
Nf6 {+0.13/16 0.49s} 7. O-O {+0.26/14 0.69s} O-O {+0.14/14 0.32s}
8. h3 {+0.28/13 0.45s} Bd7 {+0.15/15 0.38s} 9. f4 {+0.48/12 0.42s}
e5 {+0.13/14 0.50s} 10. f5 {+0.53/12 0.67s} h6 {+0.10/13 0.50s}
11. g4 {+0.61/13 0.48s} Nd4 {+0.10/15 0.33s} 12. g5 {+0.74/12 0.60s}
hxg5 {+0.24/13 0.57s} 13. Bxg5 {+0.84/12 0.48s} Qc8 {+0.16/13 0.28s}
14. Kh1 {+1.72/12 0.43s} a5 {+0.06/14 0.57s} 15. Bxf6 {+2.16/14 0.29s}
Bxf6 {-0.17/15 0.34s} 16. Nd5 {+2.19/15 0.33s} Qd8 {-0.21/16 0.30s}
17. c3 {+1.84/14 0.53s} Nxe2 {-0.04/14 0.26s} 18. Qxe2 {+1.50/13 0.32s}
b5 {-0.04/13 0.49s} 19. Rf3 {+3.53/11 0.31s} Bh4 {-0.41/15 0.26s}
20. Raf1 {+3.43/13 0.28s} Kh7 {-0.37/13 0.28s} 21. fxg6+ {+7.28/13 0.37s}
Kxg6 {-0.42/14 0.29s} 22. Rf5 {+7.26/15 0.41s} Bxf5 {-4.10/15 0.33s}
23. exf5+ {+8.07/15 0.43s} Kh6 {-3.84/13 0.32s} 24. f6 {+8.79/13 0.31s}
Qd7 {-4.91/14 0.99s} 25. Nb6 {+9.13/15 0.41s} Qd8 {-4.57/14 0.29s}
26. Rf5 {+13.80/15 0.21s} Bg5 {-7.37/13 0.18s} 27. Rxg5 {+99.93/32 0.18s}
Kxg5 {-M6/127 0.030s} 28. Qg4+ {+99.95/43 0.16s} Kh6 {-M4/127 0.007s}
29. Qh4+ {+99.97/121 0.062s} Kg6 {-M2/127 0.011s}
30. Be4# {+99.99/121 0.008s, White mates} 1-0

[/pgn]

[Werner]

Very interesting!
Now we can compare your net with the one from Seer 1.1

[Terje]

Fun game of Minic3 using "nefarious nucleus" net versus Weiss. Minic giving everything for the king side attack

[pgn]
[Event "My Tournament"]
[Site "?"]
[Date "2020.11.03"]
[Round "395"]
[White "minic_3.00_nefarious nucleus"]
[Black "weiss"]
[Result "1-0"]
[ECO "B25"]
[GameDuration "00:00:17"]
[GameEndTime "2020-11-03T17:07:04.166 CET"]
[GameStartTime "2020-11-03T17:06:46.869 CET"]
[Opening "Sicilian"]
[PlyCount "59"]
[TimeControl "10+0.1"]
[Variation "Closed"]

1. e4 {book} c5 {book} 2. Nc3 {book} Nc6 {book} 3. g3 {book} g6 {book}
4. Bg2 {book} Bg7 {book} 5. d3 {book} d6 {book} 6. Nge2 {book}
Nf6 {+0.13/16 0.49s} 7. O-O {+0.26/14 0.69s} O-O {+0.14/14 0.32s}
8. h3 {+0.28/13 0.45s} Bd7 {+0.15/15 0.38s} 9. f4 {+0.48/12 0.42s}
e5 {+0.13/14 0.50s} 10. f5 {+0.53/12 0.67s} h6 {+0.10/13 0.50s}
11. g4 {+0.61/13 0.48s} Nd4 {+0.10/15 0.33s} 12. g5 {+0.74/12 0.60s}
hxg5 {+0.24/13 0.57s} 13. Bxg5 {+0.84/12 0.48s} Qc8 {+0.16/13 0.28s}
14. Kh1 {+1.72/12 0.43s} a5 {+0.06/14 0.57s} 15. Bxf6 {+2.16/14 0.29s}
Bxf6 {-0.17/15 0.34s} 16. Nd5 {+2.19/15 0.33s} Qd8 {-0.21/16 0.30s}
17. c3 {+1.84/14 0.53s} Nxe2 {-0.04/14 0.26s} 18. Qxe2 {+1.50/13 0.32s}
b5 {-0.04/13 0.49s} 19. Rf3 {+3.53/11 0.31s} Bh4 {-0.41/15 0.26s}
20. Raf1 {+3.43/13 0.28s} Kh7 {-0.37/13 0.28s} 21. fxg6+ {+7.28/13 0.37s}
Kxg6 {-0.42/14 0.29s} 22. Rf5 {+7.26/15 0.41s} Bxf5 {-4.10/15 0.33s}
23. exf5+ {+8.07/15 0.43s} Kh6 {-3.84/13 0.32s} 24. f6 {+8.79/13 0.31s}
Qd7 {-4.91/14 0.99s} 25. Nb6 {+9.13/15 0.41s} Qd8 {-4.57/14 0.29s}
26. Rf5 {+13.80/15 0.21s} Bg5 {-7.37/13 0.18s} 27. Rxg5 {+99.93/32 0.18s}
Kxg5 {-M6/127 0.030s} 28. Qg4+ {+99.95/43 0.16s} Kh6 {-M4/127 0.007s}
29. Qh4+ {+99.97/121 0.062s} Kg6 {-M2/127 0.011s}
30. Be4# {+99.99/121 0.008s, White mates} 1-0

[/pgn]

You'll get him next time, Weiss

[xr_a_y]

You'll get him next time, Weiss

Here is clear Weiss victory, and as Black , maybe Minic pushed it too hard on this one !

[pgn]
[Event "My Tournament"]
[Site "?"]
[Date "2020.11.03"]
[Round "415"]
[White "minic_3.00_nnue"]
[Black "weiss"]
[Result "0-1"]
[ECO "B04"]
[GameDuration "00:00:27"]
[GameEndTime "2020-11-03T17:19:34.409 CET"]
[GameStartTime "2020-11-03T17:19:07.295 CET"]
[Opening "Alekhine's defense"]
[PlyCount "104"]
[TimeControl "10+0.1"]
[Variation "Modern, Larsen Variation"]

1. e4 {book} Nf6 {book} 2. e5 {book} Nd5 {book} 3. d4 {book} d6 {book}
4. Nf3 {book} dxe5 {book} 5. Nxe5 {+1.06/14 0.54s} Nd7 {-0.42/15 0.46s}
6. Nxf7 {+1.59/16 0.64s} Kxf7 {0.00/17 0.30s} 7. Qh5+ {+1.96/16 0.63s}
Ke6 {0.00/18 0.29s} 8. c4 {+1.38/14 0.63s} N5f6 {+1.43/15 0.60s}
9. d5+ {+2.47/16 0.48s} Kd6 {+1.35/15 0.30s} 10. Qf7 {+2.46/18 0.55s}
Ne5 {+2.03/15 0.35s} 11. Bf4 {+2.67/13 0.36s} c5 {+1.78/15 0.32s}
12. Nc3 {+3.62/15 0.52s} a6 {+2.18/15 0.65s} 13. O-O-O {+3.21/14 0.51s}
g6 {+2.33/14 0.37s} 14. Bxe5+ {-0.26/11 0.36s} Kxe5 {+2.12/13 0.25s}
15. d6 {+0.08/13 0.34s} Bh6+ {+2.24/15 0.77s} 16. Kb1 {-1.75/13 0.48s}
Rf8 {+2.45/15 0.37s} 17. Qxe7+ {-1.91/16 0.43s} Qxe7 {+2.61/16 0.32s}
18. dxe7 {-2.51/15 0.43s} Re8 {+2.63/16 0.33s} 19. Nd5 {-2.55/14 0.26s}
Nxd5 {+4.46/16 0.28s} 20. Rxd5+ {-3.54/13 0.40s} Kf6 {+4.57/15 0.21s}
21. Bd3 {-4.07/16 0.36s} Rxe7 {+4.56/16 0.39s} 22. Rxc5 {-4.19/15 0.34s}
Be6 {+5.06/16 0.24s} 23. Be4 {-4.25/14 0.19s} Rd8 {+4.72/16 0.32s}
24. Bd5 {-4.72/14 0.35s} Bf5+ {+5.04/16 0.21s} 25. Ka1 {-4.35/13 0.35s}
Bd2 {+4.53/16 0.19s} 26. a3 {-3.87/12 0.29s} Be1 {+5.67/15 0.22s}
27. f4 {-4.59/13 0.32s} Bf2 {+6.93/18 0.23s} 28. Ra5 {-4.70/12 0.19s}
Bb6 {+7.10/18 0.23s} 29. Ra4 {-5.14/12 0.24s} a5 {+7.23/17 0.19s}
30. c5 {-5.61/14 0.26s} Bxc5 {+7.67/17 0.33s} 31. Bf3 {-6.39/15 0.28s}
b6 {+7.75/18 0.34s} 32. h4 {-7.42/12 0.27s} Re3 {+9.28/17 0.41s}
33. Rc4 {-8.15/11 0.15s} Rd2 {+9.64/14 0.18s} 34. Rc3 {-8.15/14 0.15s}
Rxc3 {+11.01/16 0.28s} 35. bxc3 {-8.62/14 0.24s} Bxa3 {+11.20/15 0.33s}
36. Re1 {-8.55/13 0.24s} Bb2+ {+12.42/15 0.17s} 37. Ka2 {-9.03/13 0.20s}
Bxc3+ {+14.33/15 0.19s} 38. Kb3 {-8.83/11 0.22s} Bb4 {+14.83/15 0.21s}
39. Rc1 {-10.95/13 0.20s} b5 {+15.27/14 0.18s} 40. Rc6+ {-11.35/20 0.17s}
Be6+ {+15.56/14 0.15s} 41. Rxe6+ {-11.62/17 0.18s} Kxe6 {+15.61/13 0.19s}
42. Bc6 {-12.02/16 0.18s} Kf5 {+16.08/16 0.23s} 43. g3 {-12.11/14 0.12s}
Rd3+ {+16.43/16 0.22s} 44. Kb2 {-12.60/15 0.18s} a4 {+20.08/15 0.14s}
45. Bxb5 {-12.82/12 0.12s} a3+ {+27.50/14 0.17s} 46. Kc1 {-14.61/13 0.11s}
Rxg3 {+27.88/17 0.19s} 47. Bd7+ {-14.41/13 0.14s} Kf6 {+28.50/16 0.14s}
48. h5 {-19.99/12 0.17s} a2 {+M21/18 0.24s} 49. Kb2 {-26.89/13 0.15s}
Ra3 {+M7/24 0.13s} 50. Be6 {-99.94/23 0.074s} a1=Q+ {+M5/127 0.032s}
51. Kc2 {-99.96/81 0.053s} Qc3+ {+M3/127 0.014s} 52. Kb1 {-99.98/121 0.022s}
Ra1# {+M1/127 0.013s, Black mates} 0-1
[/pgn]

[xr_a_y]

Very interesting!
Now we can compare your net with the one from Seer 1.1

Yes !

Seer net is better inside Minic at short TC at least.
This look logical as the Seer net is based on SF data and is trained longer.
Minic one is using same data as Nascent Nutrient.
I'm trying to make "nefarious nucleus" stronger currently

[Madeleine Birchfield]

You'll get him next time, Weiss

I remember you said at TCEC that you had plans to implement NNUE into your engine.

[Madeleine Birchfield]

Second year anniversary for Minic ! and a new version number 3.00

Porting Stockfish NNUE implementation inside Minic was fun and instructive. This has clearly raised my interest in NN technologies and in this sense is a success.
As you already know, I was never comfortable with the big copy/paste of SF code involved, nor with the use of SF based nets of course.
This is why I first trained a net with my own data (but using Nodchip learner), namely the "napping nexus" net and then merged the learner directly inside Minic in order to train the pure Minic net named "nascent nutrient". Still, this was not enough to make me feel this technology and code can be inside Minic this way and in no way this can be considered as an "official" version of Minic as soon as a NNUE net is in use. This is why I decided to not use SF implementation anymore. This means that, starting from version 3, Minic is no more compatible with SF nets. Only versions 2.47 to 2.53 can read and use SF nets (and only using NNUE version 0x7AF32F16u nets).

I'm glad that people are moving away from Stockfish's unoptimised NNUE architecture and trainer code and trying to implement and optimise NNUE on their own.

[connor_mcmonigle]

Congrats and thanks for consulting with me on this! I'm excited to see what kind of optimizations and architecture improvements you're able achieve. IMHO, the barrier to entry for making improvements/exploring different architectures is significantly lowered with my NN implementation as the C++ inference code totals a mere ~300 LOC (Compared to 5000 LOC for the NNUE implementation found in Stockfish). (The Pytorch training scripts are a bit more sizable, but still pretty light). Ultimately, I'm just pleased my networks can produce competitive results. I'll definitely be rooting for Minic at TCEC knowing that I contributed to its development. Rather amusingly, some improvements to the architecture since 1.0 now found both in Minic 3.0 and Seer 1.1 make the play far more aggressive I've found (the change is largely inspired by this paper: https://arxiv.org/abs/1608.06993 and involves concatenating encodings from separate layers to form skip connections to stabilize and improve training of deeper networks).