Deus X Neural Network Progress

[AdminX]

How has the progress of the 'Deus X' neural network been going? Block sizes, Games, Current plans, Playing Style, Elo, etc ... With no selfplay games I would guess that progress may be a little slower in terms of improving strength.

PS: Would be nice if you developed a website / blog.

[Albert Silver]

How has the progress of the 'Deus X' neural network been going? Block sizes, Games, Current plans, Playing Style, Elo, etc ... With no selfplay games I would guess that progress may be a little slower in terms of improving strength.

PS: Would be nice if you developed a website / blog.

Hi Ted,

Things have been going well, and I have since moved to a larger 256x20 size. My experiment with the 'only human games', which was the basis for Deus X 1.0 and 1.1 (both of which played in TCEC), and is now in CCRL on a single core, has since ended. I have learned a lot regarding training, what I want, what I don't, and the current iteration, which I refer to friends as 'Frankenstein' has brought the fruit of its labors. Numerous changes have been made to the binaries and training code, most which would not be considered 'zero' in any way or form, to explore different ideas. The binary changes are all about producing different selfplay games, which are now added to the human and engine games, and not bound to the 'zero' philosophy. A minor example would be that tablebases are supported in the search in the selfplay games. In other words, it is a massive hybrid using all tools and ideas possible, but with only one restriction: no Leela NN weights or games are used in any way or form at any point. The current 20b is pretty strong, but I call it Frankenstein because it was designed to be a testbed for the various experiments to see what works or not. The plan is to use the ones I think work and rebuild from scratch, though this has not started yet.

Albert

[kasinp]

Thanks for the update and good luck with the project. It will be good to see it reenter TCEC when you feel sufficient progress has been made.
There is a lot of interest in different approaches to NN, perhaps specifically about not being as rigid on the self-imposed "zero" aspect.

Cheers,
PK

[yanquis1972]

How has the progress of the 'Deus X' neural network been going? Block sizes, Games, Current plans, Playing Style, Elo, etc ... With no selfplay games I would guess that progress may be a little slower in terms of improving strength.

PS: Would be nice if you developed a website / blog.

Hi Ted,

Things have been going well, and I have since moved to a larger 256x20 size. My experiment with the 'only human games', which was the basis for Deus X 1.0 and 1.1 (both of which played in TCEC), and is now in CCRL on a single core, has since ended. I have learned a lot regarding training, what I want, what I don't, and the current iteration, which I refer to friends as 'Frankenstein' has brought the fruit of its labors. Numerous changes have been made to the binaries and training code, most which would not be considered 'zero' in any way or form, to explore different ideas. The binary changes are all about producing different selfplay games, which are now added to the human and engine games, and not bound to the 'zero' philosophy. A minor example would be that tablebases are supported in the search in the selfplay games. In other words, it is a massive hybrid using all tools and ideas possible, but with only one restriction: no Leela NN weights or games are used in any way or form at any point. The current 20b is pretty strong, but I call it Frankenstein because it was designed to be a testbed for the various experiments to see what works or not. The plan is to use the ones I think work and rebuild from scratch, though this has not started yet.

Albert

Thanks for the update. Do you have any plans yet re keeping the fruits of your labor private or sharing/collaborating with the leela team? (I’m assuming that if test40/50 is a “success” (and possibly regardless) the next step there is to experiment on a broad scale.

[Albert Silver]

How has the progress of the 'Deus X' neural network been going? Block sizes, Games, Current plans, Playing Style, Elo, etc ... With no selfplay games I would guess that progress may be a little slower in terms of improving strength.

PS: Would be nice if you developed a website / blog.

Hi Ted,

Things have been going well, and I have since moved to a larger 256x20 size. My experiment with the 'only human games', which was the basis for Deus X 1.0 and 1.1 (both of which played in TCEC), and is now in CCRL on a single core, has since ended. I have learned a lot regarding training, what I want, what I don't, and the current iteration, which I refer to friends as 'Frankenstein' has brought the fruit of its labors. Numerous changes have been made to the binaries and training code, most which would not be considered 'zero' in any way or form, to explore different ideas. The binary changes are all about producing different selfplay games, which are now added to the human and engine games, and not bound to the 'zero' philosophy. A minor example would be that tablebases are supported in the search in the selfplay games. In other words, it is a massive hybrid using all tools and ideas possible, but with only one restriction: no Leela NN weights or games are used in any way or form at any point. The current 20b is pretty strong, but I call it Frankenstein because it was designed to be a testbed for the various experiments to see what works or not. The plan is to use the ones I think work and rebuild from scratch, though this has not started yet.

Albert

Thanks for the update. Do you have any plans yet re keeping the fruits of your labor private or sharing/collaborating with the leela team? (I’m assuming that if test40/50 is a “success” (and possibly regardless) the next step there is to experiment on a broad scale.

Several ideas, those that might fit the 'zero' approach, have already been discussed with them, but right now their main focus has been to follow Deep Mind where they can. As to what might come after their next projects such as Test40 or 50, I won't claim to know what they might entail, and doubt they do either. I discuss ideas plenty of times with them, but as is my case, we all have our own ideas and desires of what to try next. I will add there are Lc0-compatible projects in the works other than mine, such as Ender (specialized endgame NN weights), Anti-fish, and no doubt others.

[AdminX]

How has the progress of the 'Deus X' neural network been going? Block sizes, Games, Current plans, Playing Style, Elo, etc ... With no selfplay games I would guess that progress may be a little slower in terms of improving strength.

PS: Would be nice if you developed a website / blog.

Hi Ted,

Things have been going well, and I have since moved to a larger 256x20 size. My experiment with the 'only human games', which was the basis for Deus X 1.0 and 1.1 (both of which played in TCEC), and is now in CCRL on a single core, has since ended. I have learned a lot regarding training, what I want, what I don't, and the current iteration, which I refer to friends as 'Frankenstein' has brought the fruit of its labors. Numerous changes have been made to the binaries and training code, most which would not be considered 'zero' in any way or form, to explore different ideas. The binary changes are all about producing different selfplay games, which are now added to the human and engine games, and not bound to the 'zero' philosophy. A minor example would be that tablebases are supported in the search in the selfplay games. In other words, it is a massive hybrid using all tools and ideas possible, but with only one restriction: no Leela NN weights or games are used in any way or form at any point. The current 20b is pretty strong, but I call it Frankenstein because it was designed to be a testbed for the various experiments to see what works or not. The plan is to use the ones I think work and rebuild from scratch, though this has not started yet.

Albert

Thanks for the update Albert!

[Eduard]

I found in my kibitzing database 2 interesting games, played on playchess.com:

The special thing in the games is, that there is nothing very special.
The openings are played a thousend times.

In following game LC0 still found fine moves,
like 19. a5, 33. h4, 34 h5, and others. Very impressive.

[Event "60m+15s, rated"]
[Site "Engine Room"]
[Date "2018.12.11"]
[Round "?"]
[White "Erdo, Lc0 v0.19.1-rc2 Alp"]
[Black "Illumina, Brainfish 091218 64"]
[Result "1-0"]
[ECO "D45"]
[WhiteElo "2450"]
[BlackElo "2435"]
[Annotator "0.37;0.00"]
[PlyCount "102"]
[EventDate "2018.12.01"]
[TimeControl "3600+15"]

1. d4 {0 B/ 0} d5 {0 B/ 0} 2. c4 {0 B/ 0} e6 {0 B/ 0} 3. Nc3 {0 B/ 0} c6 {
0 B/ 0} 4. Nf3 {0 B/ 0} Nf6 {0 B/ 0} 5. e3 {0 B/ 0} Nbd7 {0 B/ 0} 6. Qc2 {
0 B/ 0} Bd6 {0 B/ 0} 7. b3 {0 B/ 0} O-O {0 B/ 0} 8. Be2 {0.37/14 123} b6 {
0 B/ 0} 9. O-O {0.37/15 163} Bb7 {0 B/ 0} 10. Bb2 {0.38/16 17} Re8 {0 B/ 0} 11.
Rad1 {0.43/15 100} Qc7 {0 B/ 0} 12. a4 {0.42/15 113} e5 {0 B/ 0} 13. dxe5 {
0.49/18 115} Nxe5 {0 B/ 0} 14. h3 {0.54/19 57} Nxf3+ {0.00/40 284} 15. Bxf3 {
0.58/19 109} dxc4 {0.00/43 0} 16. bxc4 {0.54/19 55} Rad8 {0.00/40 56} 17. Ne2 {
0.86/15 (Ne4) 114} Nd7 {0.00/34 74} 18. Nd4 {0.90/19 19} g6 {0.06/38 160} 19.
a5 {1.11/14 (Be2) 115} bxa5 {0.09/35 248} 20. Rb1 {1.22/18 (Ba1) 172} Bb4 {
0.10/33 124} 21. Ba1 {1.43/24 (c5) 82} Ne5 {0.27/37 210} 22. c5 {1.39/36 0} Ba6
{0.09/33 39} 23. Rfd1 {1.37/28 5} Kf8 {0.00/37 82} 24. Be4 {2.46/16 111} Bb5 {
0.00/40 0} 25. Nxb5 {2.45/16 0 (Bc3)} cxb5 {0.00/40 57} 26. c6 {2.01/19 0 (Bd5)
2.1/19 0} Nc4 {0.00/38 79} 27. Bd5 {2.61/17 107} Rxd5 {0.20/41 117} 28. Rxd5 {
2.91/21 12} Qxc6 {0.09/38 66} 29. Rbd1 {3.15/19 37} a6 {0.00/41 65} 30. e4 {
3.16/23 29 (Qc1)} f6 {0.10/41 69} 31. Qc1 {2.94/25 109} Kg8 {0.09/44 0} 32. Qf4
{2.97/28 59} Be7 {0.09/45 9} 33. h4 {2.94/28 28 (e5) 2.94/ 28} Qe6 {0.09/38 278
} 34. h5 {2.88/31 0} g5 {0.73/39 202} 35. Qc7 {3.08/34 0} a4 {0.22/41 251} 36.
e5 {3.34/33 0} fxe5 {1.13/44 419} 37. Bxe5 {3.11/30 0} Rc8 {1.03/37 57} 38. Qb7
{3.07/39 0} Nxe5 {0.65/38 34} 39. Rxe5 {3.05/30 0} Qxe5 {1.24/43 190} 40. Qxc8+
{2.91/26 1} Kf7 {1.46/44 244} 41. Qc2 {3.83/22 0 (Qxa6)} Kg7 {1.10/37 78} 42.
Qc6 {4.05/25 0} Kf7 {0.92/37 67} 43. Rd3 {5.09/19 (Qxa6) 107} g4 {0.89/33 57}
44. Re3 {5.39/21 58} Qa1+ {1.53/36 3} 45. Kh2 {5.31/19 20} Qf6 {1.43/35 28} 46.
Qe4 {5.43/19 15} Kf8 {3.04/36 169} 47. Kg1 {5.81/20 35} a3 {3.19/34 11} 48. Qc2
{5.92/18 1} Kg7 {3.25/33 37} 49. g3 {7.60/13 113} h6 {3.39/40 15} 50. Re4 {
9.19/15 107} Kh8 {5.84/40 70} 51. Rxg4 {10.81/16 55} Qf8 {6.59/40 Illumina,
Brainfish 091218 64 abbandona (Lag: Av=0.22s, max=0.4s) 71} 1-0

[Event "16m+5s, rated"]
[Site "Engine Room"]
[Date "2018.12.29"]
[Round "?"]
[White "Erdo, Lc0 v0.20.0-rc2 Deu"]
[Black "Dambord, Brainfish LP5"]
[Result "1-0"]
[ECO "E54"]
[WhiteElo "2485"]
[BlackElo "2517"]
[Annotator "?ev?;0.00"]
[PlyCount "115"]
[EventDate "2018.12.29"]
[EventType "rapid"]
[TimeControl "960+5"]

1. d4 {39} Nf6 {B 0} 2. c4 {20} e6 {B 0} 3. Nc3 {23} Bb4 {B 0} 4. e3 {66} O-O {
B 0} 5. Bd3 {2} d5 {B 0} 6. Nf3 {36} dxc4 {B 0} 7. Bxc4 {10} c5 {B 0} 8. O-O {
53} cxd4 {B 0} 9. exd4 {15} b6 {B 0} 10. Bg5 {17} Bb7 {B 0} 11. Qe2 {7} Bxc3 {
B 0} 12. bxc3 {23} Nbd7 {B 0} 13. Bb5 {49} Qc7 {0.00/35 80} 14. Rfc1 {1} Rac8 {
-0.14/30 9} 15. a4 {31} Ne4 {0.00/37 50} 16. Bh4 {41 (Be7)} Ndf6 {0.00/37 27}
17. Qe1 {23 (Qe3)} Nd6 {0.00/34 30} 18. Ne5 {20} Nd5 {0.00/37 0} 19. Bg3 {
26 (f3)} Nf5 {0.55/36 83} 20. c4 {17} Nde7 {0.48/34 10} 21. c5 {0} Nxg3 {
0.40/35 25} 22. hxg3 {0} bxc5 {0.13/36 17} 23. dxc5 {0} Nc6 {0.13/37 25} 24.
Nc4 {25} Ba8 {0.57/36 164} 25. Nd6 {28} Rcd8 {0.65/32 38} 26. Rd1 {1 (Ra3)} Ne7
{0.39/31 22} 27. Rac1 {24} Bd5 {0.54/34 30} 28. Qc3 {24 (Bc4)} Rb8 {0.66/32 49}
29. Rd4 {29} Rbd8 {0.85/34 67} 30. Rg4 {39 (Bf1)} Ng6 {0.85/30 14} 31. Bd3 {
39 (Rd4)} Qc6 {1.11/34 40} 32. Bf1 {17 (Bb5)} Qc7 {1.06/33 45} 33. Nb5 {
28 (Rb4)} Qe7 {0.74/32 35} 34. Rb4 {28 (Rd4)} Rc8 {0.87/32 36} 35. a5 {26 (Nd6)
} h5 {1.53/34 107} 36. a6 {9 (Qe3)} h4 {1.69/29 10} 37. gxh4 {5} Nxh4 {1.52/28
2} 38. Rg4 {20 (Qe5)} f5 {1.62/28 7} 39. Rg3 {10} f4 {1.66/31 1} 40. Rg4 {4}
Nf5 {1.62/32 17} 41. Rxf4 {17} Rfd8 {2.67/34 49} 42. Rg4 {3} e5 {2.68/31 12}
43. Bd3 {13} e4 {2.83/32 13} 44. Bb1 {3} Rc6 {2.96/33 13} 45. Nd4 {18 (Rf4)}
Nxd4 {2.42/25 6} 46. Qxd4 {0} Qf6 {2.39/30 9} 47. Qe3 {26} Rxa6 {2.93/33 0} 48.
Rf4 {36 (Bxe4)} Qb2 {2.00/29 7} 49. Bxe4 {17} Bxe4 {2.76/32 6} 50. Rxe4 {4} Rf6
{3.86/33 11} 51. Rf4 {9} Rxf4 {4.16/34 2} 52. Qxf4 {17} Rd4 {4.71/34 16} 53.
Qe3 {15} Re4 {5.56/34 13} 54. Qg5 {14} Qe5 {6.26/30 10} 55. Qd8+ {18} Qe8 {
6.94/30 3} 56. Qa5 {13} Qc6 {7.60/31 8} 57. Qxa7 {17} Rg4 {8.68/30 4} 58. Qb8+
{12 Dambord,Brainfish LP5 resigns (Lag: Av=0.67s, max=2.0s)} 1-0

[jp]

I found in my kibitzing database 2 interesting games, played on playchess.com:

Okay, but who is White exactly? What do "alp" and "deu" mean?

[AdminX]

I found in my kibitzing database 2 interesting games, played on playchess.com:

Okay, but who is White exactly? What do "alp" and "deu" mean?

Erdo is a tester of Deus X, Alp is for Alpha and Deu is Deus X. The text strings are cut short, but it appears that Erdo was tesing an Alpha version of Deus X in one of the games.

[jp]

Erdo is a tester of Deus X, Alp is for Alpha and Deu is Deus X. The text strings are cut short, but it appears that Erdo was tesing an Alpha version of Deus X in one of the games.

I see. Makes sense. Thanks, Ted.