Historic Milestone: AlphaZero

[melajara]

Well, this might be settled by having AlphaZero as a special guest on TCEC 11.

Meanwhile, Google has enough time to optimize then scale down a custom version compatible with TCEC hardware

[kranium]

[pgn][Event "AlphaZero-SF"]
[Site "https://lichess.org/zbCOTavB"]
[Date "2017.12.06"]
[Round "-"]
[White "AlphaZero"]
[Black "Stockfish8"]
[Result "1-0"]
[UTCDate "2017.12.06"]
[UTCTime "09:29:15"]
[WhiteElo "?"]
[BlackElo "?"]
[Variant "Standard"]
[TimeControl "-"]
[ECO "E17"]
[Opening "Queen's Indian Defense: Classical Variation, Polugaevsky Gambit"]
[Termination "Normal"]
[Annotator "lichess.org"]

1. Nf3 Nf6 2. d4 e6 3. c4 b6 4. g3 Bb7 5. Bg2 Be7 6. O-O O-O 7. d5 exd5 8. Nh4 { E17 Queen's Indian Defense: Classical Variation, Polugaevsky Gambit } c6 9. cxd5 Nxd5 10. Nf5 Nc7 11. e4 d5 12. exd5 Nxd5 13. Nc3 Nxc3 14. Qg4 g6 15. Nh6+ Kg7 16. bxc3 Bc8 17. Qf4 Qd6 18. Qa4 g5 19. Re1 Kxh6 20. h4 f6 21. Be3?! { (-0.64 → -1.48) Inaccuracy. Best move was Rxe7. } (21. Rxe7 Qxe7 22. Ba3 c5 23. Bxa8 Bf5 24. Bg2 Rd8 25. hxg5+ fxg5 26. Bc1 Kg7 27. Be3 h6) 21... Bf5? { (-1.48 → -0.25) Mistake. Best move was Kg7. } (21... Kg7 22. Rad1 Qe6 23. Bxg5 Qf7 24. Be3 b5 25. Qe4 a5 26. a4 bxa4 27. Qxa4 Re8 28. Bf1) 22. Rad1 Qa3?! { (-0.61 → 0.00) Inaccuracy. Best move was Qc7. } (22... Qc7 23. Be4) 23. Qc4 b5 24. hxg5+ fxg5 25. Qh4+ Kg6 26. Qh1 Kg7 27. Be4 Bg6 28. Bxg6 hxg6 29. Qh3 Bf6 30. Kg2 Qxa2 31. Rh1 Qg8 32. c4 Re8 33. Bd4? { (0.00 → -1.07) Mistake. Best move was Rd6. } (33. Rd6 bxc4 34. Qh6+ Kf7 35. Rxf6+ Kxf6 36. Qxg5+ Ke6 37. Rd1 Qf7 38. Qg4+ Qf5 39. Qxc4+ Ke7) 33... Bxd4? { (-1.07 → 0.00) Mistake. Best move was Rd8. } (33... Rd8 34. Bxf6+ Kxf6 35. cxb5 Rd5 36. Qg4 Nd7 37. Qf3+ Kg7 38. bxc6 Nf6 39. Rb1 Rf8 40. Rb7+) 34. Rxd4 Rd8? { (-0.24 → 1.59) Mistake. Best move was Re5. } (34... Re5 35. cxb5 cxb5 36. Rd3 Re6 37. Qh6+ Kf7 38. Qxg5 Nc6 39. Rf3+ Ke8 40. Qxb5 Rb8 41. Qc5) 35. Rxd8 Qxd8 36. Qe6 Nd7 37. Rd1 Nc5 38. Rxd8 Nxe6 39. Rxa8 Kf6 40. cxb5 cxb5 41. Kf3 Nd4+?! { (1.29 → 2.10) Inaccuracy. Best move was Ke5. } (41... Ke5 42. Ke3 Ng7 43. Rxa7 Nf5+ 44. Kd3 Kd5 45. Rb7 Nd6 46. Rb8 g4 47. Rg8 Nc4 48. Kc3) 42. Ke4 Nc6 43. Rc8 Ne7? { (1.60 → 3.92) Mistake. Best move was Nb4. } (43... Nb4 44. Rc5) 44. Rb8 Nf5?! { (3.50 → 4.33) Inaccuracy. Best move was b4. } (44... b4 45. Rxb4) 45. g4 Nh6? { (3.92 → 4.98) Mistake. Best move was Nd6+. } (45... Nd6+ 46. Kd5 Nf7 47. Rxb5 Ne5 48. Rb4 Nf3 49. Ra4 Nh2 50. Ra6+ Kg7 51. Rxa7+ Kf6 52. Ra4) 46. f3 Nf7 47. Ra8 Nd6+ 48. Kd5 Nc4 49. Rxa7 Ne3+ 50. Ke4 Nc4 51. Ra6+ Kg7 52. Rc6 Kf7?! { (4.23 → 4.78) Inaccuracy. Best move was Nd2+. } (52... Nd2+ 53. Ke3 Nb1 54. Kd3 b4 55. Rb6 Nc3 56. Rxb4 Nd5 57. Rb5 Nf4+ 58. Ke3 Nh3 59. Ke4) 53. Rc5 Ke6?! { (4.01 → 4.87) Inaccuracy. Best move was Kf6. } (53... Kf6 54. Rd5) 54. Rxg5 Kf6 55. Rc5 g5 56. Kd4 { Black resigns. } 1-0
[/pgn]

I tried with the 'old' SF8 here and even with 1 cpu setting it would never play 21. Be3??
Also it would always go for a dead draw by repetition with 16. Nf5+ (=) instead of the weaker 16. bxc3?

This raises some questions. Did they perhaps add extreme contempt to SF? Of course the whole setup could be erroneous anyways.

Thanks to Vincent for the lichess analysis pgn.

Stockfish greatly prefers Be3 on my i7 4930k w/12 threads

FEN: rnb2r2/p3b2p/1ppq1p1k/6p1/Q6P/2P3P1/P4PB1/R1B1R1K1 w - - 0 21

Stockfish 8 64 POPCNT Arena multiPV=2 after 20. h4 f6 :
---------------------------------------------------------------------------
39/78 42:31 31,322,562k 12,274k -0.54 Re1xe7 Qd6xe7 Bc1a3 c6c5 Bg2xa8 Bc8f5 Ba8f3 Qe7d7 Qa4b3 Qd7e6 Ra1d1 g5xh4 Qb3a4 h4xg3 Ba3c1+ Kh6g7 f2xg3 Qe6e7 Bc1f4 Rf8e8 Kg1h2 a7a6 Rd1d5 Bf5d7 Qa4d1 Bd7c6 Rd5d2 Bc6xf3 Qd1xf3 Qe7e6 Qf3b7+ Qe6f7 Qb7xb6 Qf7h5+ Kh2g2 Re8e2+ Rd2xe2 Qh5xe2+ Kg2h3 Nb8d7 Qb6d6 Qe2h5+ Kh3g2 Nd7e5 Qd6e7+ Qh5f7 Qe7xf7+ Kg7xf7 Bf4e3 c5c4 Kg2f2 Ne5g4+ Kf2e2 Ng4xe3 Ke2xe3

39/78 42:31 31,322,562k 12,274k 0.00 Bc1e3 Bc8f5 Ra1d1 Qd6a3 h4xg5+ f6xg5 Qa4h4+ Kh6g6 Qh4h1 Kg6g7 Be3c1 Qa3c5 Bc1e3 Qc5a3

[Guenther]

White is AlphaZero in this game :-) and white won :-)

oops my bad.

This means the excessive SF blunder here is 34...Rd8???
No time to analyse now.

Guenther

[Lyudmil Tsvetkov]

First example game, Stockfish-Alphazero 0-1
[pgn]
1. e4 e5 2. Nf3 Nc6 3. Bb5 Nf6 4. d3 Bc5 5. Bxc6 dxc6 6. O-O Nd7 7. Nbd2 O-O 8. Qe1 f6 9. Nc4 Rf7
10. a4 Bf8 11. Kh1 Nc5 12. a5 Ne6 13. Ncxe5 fxe5 14. Nxe5 Rf6 15. Ng4 Rf7 16. Ne5 Re7 17. a6 c5
18. f4 Qe8 19. axb7 Bxb7 20. Qa5 Nd4 21. Qc3 Re6 22. Be3 Rb6 23. Nc4 Rb4 24. b3 a5 25. Rxa5
Rxa5 26. Nxa5 Ba6 27. Bxd4 Rxd4 28. Nc4 Rd8 29. g3 h6 30. Qa5 Bc8 31. Qxc7 Bh3 32. Rg1 Rd7
33. Qe5 Qxe5 34. Nxe5 Ra7 35. Nc4 g5 36. Rc1 Bg7 37. Ne5 Ra8 38. Nf3 Bb2 39. Rb1 Bc3 40. Ng1
Bd7 41. Ne2 Bd2 42. Rd1 Be3 43. Kg2 Bg4 44. Re1 Bd2 45. Rf1 Ra2 46. h3 Bxe2 47. Rf2 Bxf4 48.
Rxe2 Be5 49. Rf2 Kg7 50. g4 Bd4 51. Re2 Kf6 52. e5+ Bxe5 53. Kf3 Ra1 54. Rf2 Re1 55. Kg2+ Bf4
56. c3 Rc1 57. d4 Rxc3 58. dxc5 Rxc5 59. b4 Rc3 60. h4 Ke5 61. hxg5 hxg5 62. Re2+ Kf6 63. Kf2
Be5 64. Ra2 Rc4 65. Ra6+ Ke7 66. Ra5 Ke6 67. Ra6+ Bd6 0-1
[/pgn]

The games are very high quality, there is no doubt about that.

[Lyudmil Tsvetkov]

Sounds like april fool. But isn't !. I just read, that they are not interested in chess? But I read somewhere: So basically they played 400 cores engine vs 60 cores engine. Hmm..

Well, THAT EXPLAINS EVERYTHING.

I did not know that the hardware advantage was this huge one!

No, I don't believe in the Alpha project, on equal hardware, they might reach SF level in about couple of years at best.
Chess is much more difficult than go.

Too much hype for nothing.

[Vinvin]

White is AlphaZero in this game and white won

oops my bad.

This means the excessive SF blunder here is 34...Rd8???
...

It's probably too late for black anyway, white has too many threats !
R+N undeveloped (and blocked) and the king is under attack.
Equal score in the eval is probably due to bad pruning/horizon effect/null-moves.

Code: Select all

McBrain_2017_v30_TCEC-x86-64-bmi2:
FEN: rn2r1q1/p5k1/2p3p1/1p4p1/2PR4/6PQ/5PK1/7R b - - 0 34

 32/51	00:47	 385.267k	8.172k	+2,10	34. ... a6 35.Td3 Ta7 36.Tf3 Dh8 37.Dg4 Dxh1+ 38.Rxh1 Td7 39.Dxg5 Td1+ 40.Rg2 Cd7 41.Ta3 bxc4 42.Txa6 Ce5 43.De3 Td3 44.De2 Rf7 45.Ta7+ Td7 46.Txd7+ Cxd7 47.Df3+ Cf6 48.Dxc6 g5 49.Dxc4+ Rg6 50.f3 Td8 51.De6 Td2+ 52.Rf1 Td5 53.Re2 Rg7 54.Dc6 Rf7 55.Dc4 Rg6 56.Dc2+ Rf7 57.Rf2 Rg7 58.Rg2 Rf7 59.Dc7+ Rg6
 32/61	00:47	 385.267k	8.172k	+1,96	34. ... Td8 35.Txd8 Dxd8 36.De6 Cd7 37.Td1 Cf8 38.Txd8 Cxe6 39.Txa8 bxc4 40.Txa7+ Rf6 41.Ta4 c3 42.Tc4 c5 43.Txc3 Re5 44.Tc4 Rd5 45.Tg4 Re5 46.Rf1 Rf6 47.Re2 Cd4+ 48.Rd3 Cf3 49.Ta4 Re5 50.Ta8 g4 51.Re3 Rd5 52.Rf4 c4 53.Td8+ Rc5 54.Rxg4
 32/46	00:47	 385.267k	8.172k	+1,96	34. ... Te6 35.Td8 Dxd8 36.Dxe6 Cd7 37.Td1 Cf8 38.Txd8 Cxe6 39.Txa8 bxc4 40.Txa7+ Rf6 41.Ta4 c3 42.Tc4 c5 43.Txc3 Re5 44.Tc4 Rd5 45.Tg4 Re5 46.Rf1 Rf6 47.Re2 Cd4+ 48.Rd3 Cf3 49.Ta4 Re5 50.Ta8 g4 51.Tg8 Rd5 52.Td8+ Rc6 53.Re3 c4 54.Rf4 Ce1 55.Rxg4 Rb6 56.Tb8+ Rc5 57.Tb2

[sovaz1997]

Not blunder:

Code: Select all

info depth 37 seldepth 64 multipv 1 score cp -171 nodes 1327544451 nps 8797977 hashfull 999 tbhits 344836 time 150892 pv e8d8 d4d8 g8d8 h3e6 b8d7 h1d1 d7f8 d1d8...

[Lyudmil Tsvetkov]

The 10 games with graph eval on Lichess :
https://lichess.org/zbCOTavB
https://lichess.org/CXpk8N5X
https://lichess.org/zDJ4IyBP
https://lichess.org/tnbEgxLG
https://lichess.org/ZMuecDNu
https://lichess.org/Y6tIfcqu
https://lichess.org/KRYcCvtd
https://lichess.org/tJzewmVJ
https://lichess.org/SfaOcHXM
https://lichess.org/B6ebUyzC

What is the usual time or depth lichess-SF8 uses for those annotations?

At a first quick glance it seems impossible that SF on good hardware
is even able to make some of the bad moves that are 'colported' from blank pgn.
From the annotations it looks more likely that in reality SF8 was AlphaChess and vice versa

I don't remember exactly but from my memory the analyse uses around 3 sec per move on a 2 GHz CPU.
The moves in the game has been played with Stockfish 8 on 64 CPUs 1 min/per move (quote from the paper : "In this section we include 10 example games played by AlphaZero against Stockfish during the 100 game match using 1 minute per move." )
70 Mnodes/sec for SF8 vs 80Knodes/sec for AlphaZero, see tables S4 in the paper : https://arxiv.org/pdf/1712.01815.pdf

I suppose it is Knodes for both?
Nps pretty much mean nothing here, we don't know what Alpha's search and evaluation is and even how it counts nodes.
Is it true that Alpha used 7 times stronger hardware?

[Vinvin]

White is AlphaZero in this game and white won

oops my bad.

This means the excessive SF blunder here is 34...Rd8???
...

There's a nice test position to get there BTW :
33.Bd4!!
[d]rn2r1q1/p5k1/2p2bp1/1p4p1/2P5/4B1PQ/5PK1/3R3R w - - 1 33

Code: Select all

McBrain_2017_v30_TCEC-x86-64-bmi2:
 29/13	00:17	 139.613k	8.115k	 0,00	33.Dh6+ Rf7 34.Fxg5 Dg7 35.Dh4 Te5 36.Fxf6 Dxf6 37.Dh7+ Dg7 38.Dh4 Df6
 30/20	00:17	 139.613k	8.115k	 0,00	33.Td6 bxc4 34.Dh6+ Rf7 35.Txf6+ Rxf6 36.Dxg5+ Re6 37.Td1 Df7 38.Dg4+ Df5 39.Dxc4+ Re7 40.Fc5+ Rf6 41.Fd4+ Re7 42.Fc5+
 30/47	00:17	 139.613k	8.115k	+1,54	33.Fd4 Fxd4 34.Txd4 Td8 35.Txd8 Dxd8 36.De6 Cd7 37.Td1 Cc5 38.Txd8 Cxe6 39.Txa8 Cd4 40.Txa7+ Rf6 41.cxb5 cxb5 42.Rf1 Re5 43.Re1 Rd5 44.Rd2 b4 45.Re3 Cb5 46.Ta8 Cd6 47.Td8 Rc5 48.Tg8 g4 49.Rd3 Cf7 50.Rc2 Ce5 51.Tc8+ Rd5 52.Tf8 g5 53.Tg8 Rc4 54.Tc8+ Rd4 55.Rb3 Rd5 56.Tc2

[Steve Maughan]

I'm gobsmacked.

There's a bit of me that's skeptical, but since they've cracked Go I guess it's legitimate.

When will the UCI version be available? I need to play with it.