DeepMinds AlphaZero Analysis Of World Champs 2018!

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Rebel, chrisw

supersharp77
Posts: 1242
Joined: Sat Jul 05, 2014 7:54 am
Location: Southwest USA

DeepMinds AlphaZero Analysis Of World Champs 2018!

Post by supersharp77 »

Amazing Development as DeepMinds AlphaZero (aka Matthew Sadler) emerges with Thoughts on Caruana & Carlsens handling of positions arising from the move 3. Bb5 variation of the Sicilian! Fascinating and critical analysis from the Acclaimed AlphaZero ..winning positions for white found? Will Leela0 weigh in Next with analysis? :D :wink:


https://www.youtube.com/watch?v=VeqbDMzi_IY
Javier Ros
Posts: 200
Joined: Fri Oct 12, 2012 12:48 pm
Location: Seville (SPAIN)
Full name: Javier Ros

Re: DeepMinds AlphaZero Analysis Of World Champs 2018!

Post by Javier Ros »

Very interesting to know that AlphaZero is again in action!!

But more interesting for us is that the evaluations and analysis of AlphaZero are very similar to the Leela Chess Zero one's in game 1.
Here are the evaluation by lc0_11261 of the first game, following the analysis of Sadler and AlphaZero.
As you can see, lc0 plays exactly the same moves as A0.


[pgn] [Event "Carlsen - Caruana World Championship Mat"] [Site "0:09:33-0:16:33"] [Date "2018.11.09"] [Round "1"] [White "Fabiano Caruana"] [Black "Magnus Carlsen"] [Result "1/2-1/2"] [ECO "B31"] [Annotator "R,J"] [PlyCount "43"] [EventDate "2018.11.09"] 1. e4 c5 2. Nf3 Nc6 3. Bb5 g6 4. Bxc6 dxc6 5. d3 Bg7 6. h3 Nf6 7. Nc3 Nd7 8. Be3 e5 9. O-O b6 10. Nh2 Nf8 11. f4 exf4 12. Rxf4 Be6 13. Rf2 h6 14. a4 ({Lc0 _11261: 1)} 14. a4 g5 15. a5 Ng6 16. Nf1 Be5 17. Nh2 O-O 18. Qh5 Kg7 19. Ng4 Bxg4 20. Qxg4 Qc8 21. Qh5 Qd7 22. Raf1 f6 23. Nd1 c4 24. dxc4 bxa5 25. Bc5 Bd4 26. Bxf8+ Rxf8 27. Qf3 Ne5 28. Qe2 {[%eval 56,12]}) ({Lc0 _11261: 2)} 14. Nf3 Qc7 15. d4 O-O-O 16. Rd2 cxd4 17. Bxd4 Bxd4+ 18. Nxd4 Nd7 19. Qe2 Ne5 20. Rad1 Kb7 21. b3 Qe7 22. Qe3 {[%eval 50,12]}) ({Lc0 _11261: 3)} 14. Qd2 g5 15. Ne2 Qc7 16. Raf1 Ng6 17. b4 cxb4 18. Qxb4 c5 19. Qa4+ Qd7 20. Qxd7+ Kxd7 21. a4 {[%eval 40,12]}) (14. Qd2 g5 15. Raf1 Qd6 16. Ng4 O-O-O 17. Nf6 Nd7 18. Nh5 Be5 19. g4 f6 20. b3 Bf7 21. Nd1 Nf8 22. Nxf6 Ne6 23. Nh5 Bxh5 24. gxh5 Nf4 25. Bxf4 gxf4 26. Rg2 Rhg8 27. Qe2 Rxg2+ 28. Qxg2 Qe6 29. Nf2 Rg8 30. Ng4 Qe8 31. Qf3 Qxh5 32. Kf2 Bc7 33. Ke2 Qg5 34. Nh2 h5 35. Rf2 Qg1 36. Nf1 h4 37. Kd2 Kb7 38. c3 Be5 39. Kc2 Qg7 40. Nh2 Bxc3 41. Qxf4 Bd4 42. Qf7+ Ka6 43. Qxg7 Rxg7 44. Re2 Rg3 45. Ng4 Rxh3 46. e5 Rf3 47. e6 Rf8 48. e7 Re8 49. Nh6 h3 50. Nf5 Bf6 51. a3 b5 52. b4 cxb4 53. axb4 Bxe7 54. Nxe7 h2 55. Rxh2 Rxe7 56. Rh6 Kb6 57. Kc3 Rd7 58. Rg6 Kc7 59. Rh6 Rd6 60. Rh8 Rg6 61. Ra8 Kb7 62. Rh8 Rg5 63. Rh7+ Kb6 64. Rh6 Rg1 65. Kc2 Rf1 66. Rg6 Rh1 67. Rf6 Rh8 68. Kc3 Ra8 69. d4 Rd8 70. Rh6 Rd7 71. Rg6 Kc7 72. Rg5 Rd6 73. Rg8 Rh6 74. Ra8 Rh3+ 75. Kc2 Ra3 76. Kb2 Ra4 77. Kc3 a6 78. Rh8 Ra3+ 79. Kb2 Rg3 80. Kc2 Rg5 81. Rh6 Rd5 82. Kc3 Rd6 83. Rh8 Rg6 84. Kc2 Kb7 85. Kc3 Rg3+ 86. Kc2 Rg1 87. Rh5 Rg2+ 88. Kc3 Rg3+ 89. Kc2 Rg4 90. Kc3 Kb6 91. Rh6 Rg5 92. Rf6 Rh5 93. Rg6 Rh3+ 94. Kc2 Rh5 95. Kc3 Rd5 96. Rh6 Kc7 97. Rh7+ Rd7 98. Rh5 Rd6 99. Rh8 Rg6 100. Rf8 Rg3+ 101. Kc2 Ra3 102. Rf7+ Kd6 103. Ra7 Kd5 104. Kb2 Rd3 105. Rxa6 Rxd4 106. Kb3 Re4 107. Kc3 Rc4+ 108. Kb3 Kd4 109. Rb6 Kd3 110. Ra6 Rc2 111. Rb6 Rc3+ 112. Kb2 Rc4 113. Kb3 Kd4 114. Ra6 Kd5 115. Ra8) 14... a5 (14... g5 15. a5 ({Lc0 _11261: 1)} 15. a5 Ng6 16. Nf1 Be5 17. Nh2 O-O 18. Qh5 Kg7 19. Ng4 Bxg4 20. Qxg4 Qc8 21. Qh5 f6 22. Nd1 Qe6 23. Bd2 Ne7 24. Qf3 b5 25. Bc3 f5 26. Qe3 Bxc3 27. Nxc3 Qe5 28. Qf3 f4 29. Raf1 {[%eval 55,16]}) ({Lc0 _11261: 2)} 15. Qh5 Qc7 16. a5 O-O-O 17. Qe2 Kb7 18. Nf3 Ng6 19. Na4 g4 20. hxg4 Bxg4 21. axb6 axb6 {[%eval 50,16]}) ({Lc0 _11261: 3)} 15. Nf1 Qc7 16. a5 Rd8 17. Qc1 Ng6 18. Ne2 O-O 19. Nfg3 Be5 20. Nf5 Bxf5 21. Rxf5 f6 22. axb6 {[%eval 25,16]}) 15... Ng6 16. Qh5 ({Lc0 _11261: 1)} 16. Qh5 c4 17. Rd1 cxd3 18. cxd3 Bd4 19. Qe2 Nf4 20. Qd2 Bb3 21. Re1 Bxe3 22. Rxe3 bxa5 23. Ne2 Nxe2+ 24. Qxe2 a4 25. e5 Qe7 26. Ng4 O-O-O 27. Rf6 h5 28. Nh6 Be6 29. Qf2 Rh7 30. Nf5 Qc5 31. Nd6+ {[%eval 55,21]}) ({Lc0 _11261: 2)} 16. Nf1 Be5 17. Nh2 O-O 18. Qh5 Kg7 19. Ng4 Bxg4 20. Qxg4 Qc8 21. Qh5 f6 22. Nd1 Qe6 23. Bd2 Bd4 24. Ne3 Bxe3 25. Bxe3 c4 26. Qe2 Qe5 27. Ra4 cxd3 28. cxd3 c5 29. b4 cxb4 30. Rxb4 Qxa5 31. Rc4 Rf7 32. Bd4 Raf8 33. Rf5 Qa3 34. Qc2 {[%eval 54,21]}) ({Lc0 _11261: 3)} 16. Qf3 O-O 17. Ng4 Bxg4 18. Qxg4 c4 19. dxc4 bxa5 20. Bc5 Re8 21. Rd1 Qb8 22. Qf5 Ne5 {[%eval 41,21]})) ( 14... Nd7 15. a5 ({Lc0 _11261: 1)} 15. a5 O-O 16. Qd2 g5 17. Ne2 Qc7 18. Raf1 Rae8 19. Nf3 f6 20. b3 Rf7 21. Qc1 f5 22. exf5 Rxf5 23. Bd2 Bd5 24. axb6 axb6 25. c4 Bxf3 26. gxf3 {[%eval 43,12]}) ({Lc0 _11261: 2)} 15. Nf3 Qc7 16. a5 O-O 17. Qd2 g5 18. Raf1 Rae8 19. Ne2 f6 20. b3 Rf7 21. Qc1 f5 22. exf5 Rxf5 23. Bd2 Bd5 24. axb6 axb6 25. c4 Bxf3 {[%eval 43,12]}) ({Lc0 _11261: 3)} 15. Qd2 g5 16. Nf3 Qc7 17. a5 b5 18. d4 cxd4 19. Bxd4 Ne5 20. Nxe5 Bxe5 21. Qe3 O-O 22. Ne2 Rfd8 23. b4 Bc4 24. Bxe5 Qxe5 25. Re1 Bxe2 {[%eval 41,12]})) 15. Nf3 ({Lc0 _11261: 1)} 15. Nf3 Qc7 16. d4 O-O-O 17. Qd3 Qb7 18. Rd1 Nd7 19. Qf1 Rhe8 20. Rfd2 g5 21. e5 Bf8 22. dxc5 Bxc5 23. Bxc5 Nxc5 24. Nd4 Qc7 25. Nxe6 fxe6 26. Rd6 Nb7 27. Ne4 Kb8 28. Rxd8+ Rxd8 29. Rxd8+ Qxd8 30. Qd3 Qd5 31. Qxd5 {[%eval 86,17]}) ({Lc0 _11261: 2)} 15. Qf3 Nd7 16. Qg3 g5 17. Nf3 Qb8 18. Ne2 Qxg3 19. Nxg3 Bxb2 20. Rb1 Bc3 {[%eval 50,17]}) ({Lc0 _11261: 3)} 15. Qd2 Qc7 16. Bf4 Qe7 17. Nf3 g5 18. Bg3 Ng6 19. e5 O-O-O 20. Ne4 Kb8 21. Re1 Ka7 {[%eval 42,17]}) 15... Qc7 16. d4 ({Lc0 _11261: 1)} 16. d4 Rd8 17. Rd2 cxd4 18. Bxd4 Bxd4+ 19. Rxd4 g5 20. Rxd8+ Qxd8 21. Nd4 Qc7 22. Qf3 Ng6 23. Rf1 O-O 24. Qf6 Re8 25. Rf2 Kh7 26. Nce2 c5 27. Nxe6 fxe6 28. Qf7+ Qxf7 29. Rxf7+ Kg8 30. Rb7 Rd8 31. Rxb6 Rd2 32. Kf1 Kf7 {[%eval 81,18]}) ({Lc0 _11261: 2) } 16. Ne2 O-O-O 17. c3 Bc4 18. d4 Ne6 19. Qc2 Rhe8 20. Re1 Kb7 21. b3 Bxe2 22. Rexe2 {[%eval 54,18]}) ({Lc0 _11261: 3)} 16. Qd2 g5 17. d4 O-O-O 18. d5 Ng6 19. Qe2 cxd5 20. exd5 Bxd5 21. Nxd5 Rxd5 22. Nd2 Kb7 23. Nc4 Rhd8 24. Raf1 R8d7 25. c3 {[%eval 18,18]}) 16... O-O-O 17. Qd3 ({Lc0 _11261: 1)} 17. Qd3 cxd4 18. Bxd4 Bxd4 19. Nxd4 Nd7 20. Nxe6 fxe6 21. e5 g5 22. Re1 Kb8 23. Qe4 Rdf8 24. Rxf8+ Rxf8 25. Qg6 Nc5 26. Qxh6 Rf5 27. Qh5 Qd8 28. Rd1 Qf8 29. Qe2 Kc7 30. Qe3 Qg7 31. Re1 g4 32. Qg3 {[%eval 85,18]}) ({Lc0 _11261: 2)} 17. Rd2 cxd4 18. Nxd4 Bc4 19. Qf3 Ne6 20. Rad1 Nxd4 21. Bxd4 Rxd4 22. Rxd4 Bxd4+ 23. Rxd4 Be6 24. Qd3 Kb7 25. Rd6 Re8 26. e5 Bf5 {[%eval 39,18]}) ({Lc0 v0.19. 0-dev: 3)} 17. d5 Bxc3 18. bxc3 cxd5 19. Qf1 dxe4 20. Ne5 f5 21. Rb1 Kb7 22. Bf4 g5 23. Bh2 f4 24. Nc4 {[%eval -37,18]}) 17... cxd4 18. Bxd4 ({Lc0 v0.19. 0-dev: 1)} 18. Bxd4 Bxd4 19. Nxd4 Nd7 20. Nxe6 fxe6 21. e5 g5 22. Re1 Nc5 23. Qf3 Rd4 24. Qe3 Rhd8 25. Ref1 Kb7 26. b3 R8d7 27. Rf6 Rd2 28. R1f2 Rxf2 29. Rxf2 Qd8 30. Kh2 h5 31. Rf6 g4 32. Ne4 {[%eval 83,20]}) ({Lc0 _11261: 2)} 18. Nxd4 Kb7 19. Rb1 Qe5 20. Nce2 c5 21. Bf4 Qh5 22. Nxe6 Nxe6 23. Bd6 Rd7 24. e5 Bxe5 25. Qe4+ Ka7 26. Bxe5 Re8 {[%eval 39,20]}) ({Lc0 _11261: 3)} 18. Qa6+ Qb7 19. Qxb7+ Kxb7 20. Rd2 c5 21. Nb5 g5 22. c3 d3 23. Rad1 Bb3 24. Ra1 Ng6 25. Rf2 Rhe8 {[%eval -282,20]}) 18... Bxd4 19. Nxd4 Nd7 20. Nxe6 ({Lc0 v0. 19.0-dev: 1)} 20. Nxe6 fxe6 21. e5 g5 22. Re1 Kb8 23. Qe4 Rdf8 24. Rxf8+ Rxf8 25. Qg6 Nc5 26. Qxh6 Rf5 27. Qh5 Qd8 28. Rd1 Qf8 29. Ne2 Rxe5 30. Nd4 Rd5 31. Nxc6+ Kc7 32. Nd4 Qf4 {[%eval 90,14]}) ({Lc0 _11261: 2)} 20. Qe3 Ne5 21. Nxe6 fxe6 22. Ne2 Kb7 23. Nd4 Qd6 24. c3 g5 25. Rf6 Rde8 {[%eval 35,14]}) ({ Lc0 _11261: 3)} 20. Qa6+ Kb8 21. Nxe6 fxe6 22. Rf7 Rhf8 23. Raf1 Rxf7 24. Rxf7 Qd6 25. Kh1 Rf8 26. Rxf8+ Qxf8 {[%eval 20,14]}) 20... fxe6 21. e5 ({Lc0 _11261: 1)} 21. e5 g5 22. Re1 Nc5 23. Qf3 Rhg8 24. Re3 Kb8 25. Ne4 Nxe4 26. Rxe4 Qe7 27. Qb3 Qc5 28. Qe3 Qxe3 29. Rxe3 g4 30. Kh2 gxh3 31. g3 Rd4 32. Rf6 Re8 33. b3 Rd2+ 34. Kxh3 Rxc2 35. Rxh6 b5 36. g4 b4 37. g5 c5 38. g6 { [%eval 78,17]}) ({Lc0 _11261: 2)} 21. Rf7 Rhf8 22. Rg7 Rg8 23. Rh7 Rh8 24. Rg7 Rhg8 25. Rf7 Rgf8 26. Rh7 Rh8 27. Re7 Rhe8 28. Rxe8 Rxe8 29. Rf1 Ne5 30. Qe3 g5 31. b3 Kb7 32. Nd1 g4 {[%eval 77,17]}) ({Lc0 _11261: 3)} 21. Qc4 Ne5 22. Qxe6+ Kb7 23. Rd1 Rhe8 24. Qf6 Rxd1+ 25. Nxd1 Qd7 26. Nc3 Qd4 27. Qg7+ Ka6 28. Qxh6 Qb4 29. Qc1 Nc4 {[%eval 66,17]}) 21... Nxe5 22. Qe3 ({Lc0 v0.19. 0-dev: 1)} 22. Qe3 Rhf8 23. Re2 Nf7 24. Qxe6+ Kb7 25. Ne4 Qe5 26. Qxg6 Ng5 27. Kh1 Rd7 28. Rae1 Nxe4 29. Rxe4 Qxb2 30. Re7 Qd4 31. Qxh6 Rf7 32. Rxf7 Rxf7 33. Qe6 Rc7 34. Qe4 Qf2 35. Rd1 Rg7 {[%eval 95,13]}) ({Lc0 _11261: 2)} 22. Qe2 Rhf8 23. Re1 Rxf2 24. Qxf2 Kb7 25. Qf6 Nc4 26. b3 Nd2 27. Rxe6 Rd7 28. Qe5 Qxe5 29. Rxe5 c5 30. Re6 c4 31. bxc4 Nxc4 32. Rxg6 Rd2 33. Rxh6 Rxc2 34. Nb5 Ra2 35. Rh7+ {[%eval 88,13]}) ({Lc0 _11261: 3)} 22. Qa6+ Kb8 23. Re1 Nd7 24. Rf7 Rhf8 25. Rg7 Rg8 26. Re7 Rge8 27. R1xe6 Rxe7 28. Rxe7 Qd6 {[%eval 48,13]}) 1/2-1/2 [/pgn]

https://sites.google.com/site/unoallavo ... ects=0&d=1
Kanizsa
Posts: 51
Joined: Mon Feb 20, 2017 8:29 am
Location: Rialto, Venice

Re: DeepMinds AlphaZero Analysis Of World Champs 2018!

Post by Kanizsa »

Sadler says that Nc3-d5 is a sideline in the Sveshnikov (main line is Bc1-g5), but Alpha Zero's first line.
And what about LCO preference on this topic ?
Kanizsa
Posts: 51
Joined: Mon Feb 20, 2017 8:29 am
Location: Rialto, Venice

Re: DeepMinds AlphaZero Analysis Of World Champs 2018!

Post by Kanizsa »

In game n. 8, on move 14

r1bq1rk1/1p1nb1pp/p2p4/P2Ppp2/8/N7/1PPBBPPP/R2Q1RK1 b - - 0 14

Carlsen played 14: ..e4?.
According to Alpha Zero he had to play 14: ..f4 ! with slight White advantage.
Kanizsa
Posts: 51
Joined: Mon Feb 20, 2017 8:29 am
Location: Rialto, Venice

Re: DeepMinds AlphaZero Analysis Of World Champs 2018!

Post by Kanizsa »

Following this game,

1rbq1rk1/1p2b2p/pN1p4/P2Pn1p1/2P2p2/2B2B2/1P4PP/R2Q1RK1 b - - 0 20

20. … Bc8-f5 was a mistake according to AlphaZero. Every chessplayer in the world had played this aggressive move !

According to Alpha Zero better was 20. ..Dd8-c7, giving up every idea to attack the King
jp
Posts: 1470
Joined: Mon Apr 23, 2018 7:54 am

Re: DeepMinds AlphaZero Analysis Of World Champs 2018!

Post by jp »

Javier Ros wrote: Sun Nov 25, 2018 11:26 am But more interesting for us is that the evaluations and analysis of AlphaZero are very similar to the Leela Chess Zero one's in game 1.
Here are the evaluation by lc0_11261 of the first game, following the analysis of Sadler and AlphaZero.
As you can see, lc0 plays exactly the same moves as A0.
Thanks, Javier. What do you mean by exactly the same moves? Do you mean for every A0 move Sadler talks about (which is how many moves?)? What time/nodes did you give lc0_11261 for each move?
Javier Ros
Posts: 200
Joined: Fri Oct 12, 2012 12:48 pm
Location: Seville (SPAIN)
Full name: Javier Ros

Re: DeepMinds AlphaZero Analysis Of World Champs 2018!

Post by Javier Ros »

I have analyzed the movements with lc0 11261 thinking in multivariate mode using from 30 seconds to 1 minute and recording the analysis to the game under Chessbase gui with kibitzer Mode, that unfortunately doesn't add times.
Javier Ros
Posts: 200
Joined: Fri Oct 12, 2012 12:48 pm
Location: Seville (SPAIN)
Full name: Javier Ros

Re: DeepMinds AlphaZero Analysis Of World Champs 2018!

Post by Javier Ros »

I forgot to mention the GPU was a gtx 1070.

Here the link with all analysis and pgn download by Sadler.

https://chess24.com/en/read/news/alphaz ... -games-1-8
Kanizsa
Posts: 51
Joined: Mon Feb 20, 2017 8:29 am
Location: Rialto, Venice

Re: DeepMinds AlphaZero Analysis Of World Champs 2018!

Post by Kanizsa »

[d]r1bqk2r/p2n1pbp/1pp3p1/2p1p3/4P3/2NPBN1P/PPP2PP1/R2Q1RK1 w kq - 0 10

Caruana played 10. Nf3-h2 ?
StockFish 021118 prefers 10. Qc1 Nf8 at depth 50

Alpha Zero suggests 10. a4
Kanizsa
Posts: 51
Joined: Mon Feb 20, 2017 8:29 am
Location: Rialto, Venice

Re: DeepMinds AlphaZero Analysis Of World Champs 2018!

Post by Kanizsa »

[d]2kr1n1r/2q2pb1/1pp1b1pp/p7/P2pP3/2NQBN1P/1PP2RP1/R5K1 w - - 0 18

Which recapture is better?
18: B:d4 or 18: N:d4
??

Sadler:
AlphaZero's preference was for 18 B:d4, which I think is related to its love of creating and exploiting weak square complexes within the opponent's position. The plan of exchanging dark-squared bishops and then the surprising idea of following up with a quick e4-e5 fits this imperative very well. It was less keen on the idea that attracted me the most - 18.Nxd4 - but did its best to keep me happy! After N:d4 the pin on the d4-knight is quite scary, but by keeping the dark-squared bishop on the board, pointing at the b6-pawn, White emphasises the weakening of Black's queenside resulting from White throwing in a4 ...a5. Moreover, Black's pawn on a5 is now within reach of the b2-b4 break from White which opens up some tempting sacrificial possibilities. The following analysis is not definitive, but gives an idea of the danger that Black faces in this position as well as his active resources.

Stockfish 201118 at depth 34 plays B:d4 (+0,71), but Stockfish 9 plays N:d4