LCZero blunder

[zullil]

[d]6k1/2q2pp1/p4n1p/1prp4/2nN1P1P/3B1P2/PP3QP1/3R2K1 b - - 4 33

LCZero played 33...Qxf4 and handed the game to Stockfish.

[pgn][Event "TCEC Season 14 - Division P"] [Site "http://tcec.chessdom.com"] [Date "2019.01.09"] [Round "28.4"] [White "Stockfish 181224"] [Black "LCZero v20rc2-32194"] [Result "1-0"] [BlackElo "3297"] [ECO "D55"] [GameDuration "03:08:43"] [GameEndTime "2019-01-09T13:57:32.976 W. Europe Standard Time"] [GameStartTime "2019-01-09T10:48:49.513 W. Europe Standard Time"] [Opening "QGD"] [PlyCount "146"] [Termination "adjudication"] [TerminationDetails "TCEC win rule"] [TimeControl "5400+10"] [Variation "6.Nf3"] [WhiteElo "3612"] {WhiteEngineOptions: Protocol=uci; Hash=16384; Threads=43; SyzygyPath=C:/Syzygy/; Move Overhead=1000; OwnBook=false; Ponder=false;, BlackEngineOptions: Protocol=uci; MoveOverheadMs=2000; WeightsFile=c:/Users/TCEC/Engines/lc0/32194.pb.gz; Threads=4; NNCacheSize=10000000; MinibatchSize=256; MaxCollisionEvents=256; CPuct=3.1; Slowmover=1.3; MaxPrefetch=32; LogFile=11248-debug.log; SyzygyPath=C:/tb; CommandLineOptions=--backend=multiplexing '--backend-opts=(backend=cudnn-fp16,gpu=0),(backend=cudnn-fp16,gpu=1)';} 1. d4 Nf6 2. c4 e6 3. Nf3 d5 4. Nc3 Be7 5. Bg5 O-O 6. e3 b6 7. cxd5 exd5 8. Bd3 Nbd7 9. O-O c5 10. Bf5 Bb7 11. Qa4 h6 12. Bf4 a6 13. Qc2 Re8 14. Rfd1 Bf8 15. h4 cxd4 16. Rxd4 Rc8 17. Rad1 Bc5 18. R4d2 b5 19. Qb1 Qe7 20. Ne2 Ne4 21. Rc2 Bb6 22. Rxc8 Rxc8 23. Ned4 Rc5 24. Bh3 Qd8 25. Nf5 Bc7 26. N3d4 Bxf4 27. exf4 Ndf6 28. f3 Nd6 29. Qd3 Bc8 30. Qe3 Bxf5 31. Bxf5 Qc7 32. Bd3 Nc4 33. Qf2 Qxf4 34. Ne6 fxe6 35. Qxc5 Nxb2 36. Qc8+ Kf7 37. Qb7+ Ke8 38. Bxb5+ Kf8 39. Qa8+ Ke7 40. Qa7+ Kd8 41. Qb6+ Ke7 42. Qa7+ Kd8 43. Qb6+ Ke7 44. Qc5+ Kf7 45. Rb1 axb5 46. Rxb2 Qxh4 47. Qa7+ Kg6 48. Qf2 Qf4 49. Qc2+ Kf7 50. Rxb5 e5 51. Rb7+ Kg8 52. Qc3 Kh8 53. a4 e4 54. Qc7 Qg5 55. f4 Qg3 56. Qd8+ Ng8 57. Rb8 Qe3+ 58. Kf1 Qxf4+ 59. Ke2 Qg4+ 60. Ke3 d4+ 61. Kxd4 Qd1+ 62. Ke5 Qh5+ 63. Kf4 Qh2+ 64. Kxe4 Qxg2+ 65. Kd3 Qf1+ 66. Kc2 Qf5+ 67. Kb2 Qe5+ 68. Kb3 Qg3+ 69. Kc4 Qg4+ 70. Kb5 Qe2+ 71. Ka5 Qe5+ 72. Ka6 Qe2+ 73. Ka7 Qf2+ 1-0[/pgn]

[M ANSARI]

This is a perfect example of how LC0 can fall to tactical shots that are obvious for AB engines. This is what makes LC0 so interesting, it really plays like a human and just like a human, it seems to have a much better understanding of positional chess, but can still fall for what is a simple tactic for today's engines. I did a quick test with SF to see how obvious Qxf4? is and I had to open 14 splits to finally see Qxf4 as a choice. This is why if LC0 wants to gain some quick ELO points a sanity check of moves with a hybrid AB engine in the background might prevent such oversights. But to be honest, LC0 has only been out a very short time and I think it is only a matter of time before these things get cleared up.

New game
[d]6k1/2q2pp1/p4n1p/1prp4/2nN1P1P/3B1P2/PP3QP1/3R2K1 b - - 0 1

Analysis by Stockfish 060119 64 POPCNT:

1. = (0.00): 1...a5 2.Re1 b4 3.g3 Nd6 4.Qe3 h5 5.b3 Qd7 6.Kg2 Qc7 7.Kg1
2. = (0.28): 1...Na5 2.Rf1 Rc1 3.Ne2 Rxf1+ 4.Kxf1 Qc8 5.g3 Qe6 6.Nd4 Qh3+ 7.Kg1 Qc8 8.Kg2 Qc5 9.Nf5 Nd7 10.Qd4 Qxd4 11.Nxd4 Nc5 12.Bf5 Nc4 13.b4 Ne3+ 14.Kh3
3. = (0.28): 1...Nd6 2.Ne2 Nc4 3.Qd4 Rc6 4.b3 Na3 5.g4 Nc2 6.Bxc2 Rxc2 7.Rd2 Rxd2 8.Qxd2 Qc5+ 9.Kf1 b4 10.Qd4 Qd6 11.Kg2 Qe7 12.Kf2 Ne8 13.g5 Nc7 14.Qe5 Qxe5 15.fxe5 hxg5
4. +/= (0.43): 1...b4 2.b3 Nd6 3.Qe3 Ra5 4.Nc2 Rxa2 5.Nxb4 Rb2 6.Qd4 Rxb3 7.Nxd5 Nxd5 8.Qxd5 Rb6 9.Qa8+ Rb8 10.Qxa6 Rd8 11.Bb1 Ne8 12.Rxd8 Qxd8 13.Qa4 g6 14.g3 Nf6 15.Ba2 Qd3 16.Qa7
5. +/= (0.45): 1...Nb6 2.Nb3 Rc6 3.g3 Na4 4.Bb1 Qd6 5.Kg2 Rc7 6.Re1 Qb4 7.Re2 Re7 8.Re5 Rxe5 9.fxe5 Nd7 10.f4 Ndc5 11.Qc2 Qe1 12.Nxc5 Nxc5 13.Qxc5 Qxb1 14.Qxd5 Qxb2+ 15.Kh3 Qc3 16.Qa8+ Kh7 17.Qxa6
6. +/= (0.33): 1...Qd7 2.Bf5
7. +/= (0.47): 1...Qc8 2.Nf5 Qc7 3.Qd4 Rc6 4.Re1 Kf8 5.g4 Qb6 6.Qxb6 Nxb6 7.Nd4 Rc8 8.Kf2 Na4 9.Rb1 Nc5 10.Ke3 Re8+ 11.Kd2 Ne6 12.Nxe6+ Rxe6 13.Rc1 d4 14.g5 hxg5 15.fxg5 Nd5
8. +/= (0.49): 1...Qb6 2.b3 Nd6 3.Ne2 h5 4.b4 Rc6 5.Qxb6 Rxb6 6.Rc1 Nc4 7.Kf2 Rb7 8.Nd4 Re7 9.g3 Re8 10.Re1 Kf8 11.Rxe8+ Nxe8 12.g4 Ncd6 13.Ke3 Nc4+ 14.Bxc4 dxc4 15.gxh5
9. +/= (0.50): 1...Qb8 2.Nf5 Qc7 3.Qd4 Rc6 4.Rc1 Qd7 5.Re1 Re6 6.Rxe6 Qxe6 7.Kf2 Nd6 8.Ne3 Nc4 9.Nc2 h5 10.b3 Nb6 11.a4 bxa4 12.bxa4 a5 13.Ne3 Qc6 14.Bb5 Qe6 15.f5 Qd6 16.g3
10. +/= (0.58): 1...Qd6 2.b3 Nb6 3.Qe3 Nbd7 4.g4 Rc7 5.Re1 Nf8 6.Nf5 Qb4 7.Ne7+ Rxe7 8.Qxe7 Qxf4 9.Kg2 Ne6 10.Rxe6 fxe6 11.Qxe6+ Kh8 12.Qc8+ Ng8 13.Qf5 Qxf5 14.Bxf5 Ne7 15.Bd3
11. +/= (0.63): 1...Qb7 2.g4 Rc8 3.b3 Nd6 4.Ne2 Qe7 5.Bb1 b4 6.g5 Nh5 7.Rxd5 Nxf4 8.Nxf4 Rc1+ 9.Kh2 hxg5 10.Rxg5 Rxb1 11.Nd5 Qb7 12.Qd4 Ne8 13.Qxb4 Rb2+ 14.Kg3
12. +/- (0.91): 1...Qe7 2.Re1 Qf8 3.Nf5 d4 4.a4 Rc7 5.axb5 axb5 6.g4 Ne3 7.Nxd4 Ned5 8.f5 b4 9.g5 hxg5 10.hxg5 Nh5 11.Bf1 Qd8 12.f6 Ndf4 13.Qe3 gxf6
13. +/- (0.96): 1...Qa7 2.b3 Nb6 3.g4 Nbd7 4.Qe3 Rc8 5.Bf5 g6 6.Bxd7 Qxd7 7.f5 Qd6 8.Kg2 h5 9.g5 Ne8 10.fxg6 fxg6 11.Re1 Ng7 12.Qe5 Qxe5 13.Rxe5 Kf7 14.Rxd5 Ne6 15.Kg3 Rc1 16.Nxe6
14. +/- (1.23): 1...Qxf4 2.Ne6 fxe6 3.Qxc5 Nxb2 4.Qc8+ Kf7 5.Qb7+ Ke8 6.Bxb5+ Kf8 7.Qa8+ Kf7 8.Be8+ Nxe8 9.Qb7+ Nc7 10.Qxb2 Qxh4 11.Rc1 Nb5 12.Qe5 Qf6 13.Qxf6+ Kxf6 14.Rc6 a5 15.Ra6 d4 16.Rxa5 Nd6 17.Ra6 Nf5 18.Kf2 Ne3 19.g3 Ke7
15. +- (2.54): 1...Qd8 2.b3 Qb6 3.bxc4 dxc4 4.Be4 Qd8 5.Bc2 Rd5 6.g4 Qe7 7.Nf5 Qd8 8.Rxd5 Nxd5 9.Qd4 Qb6 10.Qxb6 Nxb6 11.g5 Nd5 12.Kf2 h5 13.Be4 Nc3 14.Ke3 Nxa2 15.Kd4 Nc1

(Spearfish, 09.01.2019)

[zullil]

But to be honest, LC0 has only been out a very short time and I think it is only a matter of time before these things get cleared up.

That LCZero is already perhaps the second strongest engine is utterly remarkable. An incredibly exciting time for us.

[Uri Blass]

I think that Qxf4 is a positional mistake and not a tactical mistakes.
A-B Engines that are positionally weak can make this mistake.

top A-B engines are simply positional monsters and not only tactical monsters so I guess that they do not do this mistake even after one second of search but if you give old engines 10 minutes then they can still do this mistake even when they search to depth 19 or 20 and get a lot of time to compensate for their tactical weakness relative to top A-B engines.

Here is analysis by Anmon5.75 that is a GM level engine(2537 CCRL chess rating) that cannot find a different move even at depth 20 and keep a draw evaluation for Qxf4

Note that it can see the game line 1...Qxf4 2.Ne6 fxe6 3.Qxc5 Nxb2 4.Qc8+ Kf7 5.Qb7+ Ke8 6.Bxb5+ Kf8 7.Qa8+ inside the pv at depths 16-17 but evaluate it as a draw.

[d]6k1/2q2pp1/p4n1p/1prp4/2nN1P1P/3B1P2/PP3QP1/3R2K1 b - - 0 1

AnMon 5.75:
1+ 00:00 1 3 +0.71 1. ... Qxf4
1 00:00 5 16 +0.78 1. ... Qxf4
2 00:00 111 375 +0.76 1. ... Qxf4 2.b4
3 00:00 1k 4k +0.06 1. ... Qxf4 2.a4
4 00:00 3k 10k +0.54 1. ... Qxf4 2.Ne6 Nh7 3.Nxf4
5 00:00 8k 525k +0.13 1. ... Qxf4 2.Ne6 fxe6 3.Qxc5 Nxb2 4.Qc8+ Kf7
6 00:00 26k 852k +0.13 1. ... Qxf4 2.Ne6 fxe6 3.Qxc5 Nxb2 4.Qc8+ Kf7
7 00:00 52k 837k +0.13 1. ... Qxf4 2.Ne6 fxe6 3.Qxc5 Nxb2 4.Qc8+ Kf7 5.Qb7+ Kf8 6.Qc8+ Ne8
8 00:00 114k 1,049k +0.03 1. ... Qxf4 2.Nf5 Rc7 3.b3 Ne5 4.Nxg7
9 00:00 295k 1,048k +0.13 1. ... Qxf4 2.Ne6 fxe6 3.Qxc5 Kf7 4.Bxc4 bxc4 5.Qc8 Qxh4 6.Qb7+ Ke8 7.Qc8+ Kf7 8.Qb7+ Ke8 9.Qc8+ Kf7 10.Qb7+ Ke8 11.Qc8+ Kf7 12.Qb7+ Ke8 13.Qc8+ Kf7 14.Qb7+ Ke8 15.Qc8+ Kf7 16.Qb7+ Ke8 17.Qc8+
10 00:00 497k 1,098k -0.03 1. ... Qxf4 2.Bxc4 Rxc4 3.Qd2 Qxh4 4.b3 Rxd4 5.Qxd4 Qxd4+ 6.Rxd4
11 00:01 1,340k 1,144k 0.00 1. ... Qxf4 2.Ne6 fxe6 3.Qxc5 Nxb2 4.Qc8+ Kf7 5.Qb7+ Kg8 6.Qc8+ Kf7 7.Qb7+ Kg8 8.Qc8+ Kf7 9.Qb7+ Kg8 10.Qc8+ Kf7 11.Qb7+ Kg8 12.Qc8+ Kf7 13.Qb7+ Kg8 14.Qc8+ Kf7 15.Qb7+ Kg8 16.Qc8+ Kf7 17.Qb7+
12 00:01 2,349k 1,193k -0.04 1. ... Qxf4 2.Ne6 fxe6 3.Qxc5 Nxb2 4.Qc8+ Kf7 5.Qb7+ Ke8 6.Bg6+ Kd8 7.Qb6+ Qc7 8.Qxc7+ Kxc7 9.Re1 Kd7 10.g3
13+ 00:04 5,082k 1,209k -0.03 1. ... Qxf4 2.Ne6 fxe6 3.Qxc5 Nxb2 4.Qc8+ Kf7 5.Qb7+ Ke8 6.Bg6+ Kd8 7.Qb6+ Qc7 8.Be8
13 00:04 5,673k 1,223k 0.00 1. ... Qxf4 2.Ne6 fxe6 3.Qxc5 Nxb2 4.Qc8+ Kf7 5.Qb7+ Ke8 6.Bxb5+ Kf8 7.Qc8+ Kf7 8.Qb7+ Kf8 9.Qc8+ Kf7 10.Qb7+ Kf8 11.Qc8+ Kf7 12.Qb7+ Kf8 13.Qc8+ Kf7 14.Qb7+ Kf8 15.Qc8+ Kf7 16.Qb7+ Kf8 17.Qc8+
14- 00:06 7,388k 1,219k -0.01 1. ... Qxf4 2.Ne6 fxe6 3.Qxc5 Nxb2 4.Qc8+ Kf7 5.Qb7+ Ke8 6.Bxb5+ Kf8 7.Qc8+ Kf7 8.Be8+ Nxe8 9.Qb7+ Kg8 10.Qxb2 Qxh4 11.Qb6 Kf8
14 00:07 9,510k 1,240k -0.06 1. ... Qxf4 2.Ne6 fxe6 3.Qxc5 Nxb2 4.Qc8+ Kf7 5.Qb7+ Ke8 6.Bxb5+ Kf8 7.Qa8+ Kf7 8.Ra1 axb5 9.Qa3 Qxh4 10.Qa7+ Kf8 11.Qa3+ Kf7 12.Qa7+ Kf8 13.Qa3+ Kf7 14.Qa7+ Kf8 15.Qa3+ Kf7 16.Qa7+ Kf8 17.Qa3+
15 00:19 24,054k 1,226k -0.06 1. ... Qxf4 2.Ne6 fxe6 3.Qxc5 Nxb2 4.Qc8+ Kf7 5.Qb7+ Ke8 6.Bg6+ Kd8 7.Qb6+ Qc7 8.Qxc7+ Kxc7 9.Re1 Kd6 10.Bf7 e5 11.Rc1
16+ 00:35 43,423k 1,237k -0.05 1. ... Qxf4 2.Ne6 fxe6 3.Qxc5 Nxb2 4.Qc8+ Kf7 5.Qb7+ Ke8 6.Bg6+ Kd8 7.Qb6+ Qc7 8.Qxc7+ Kxc7 9.Re1 Kd6 10.Bf7 e5 11.g3 Nc4 12.f4
16+ 00:36 45,837k 1,247k -0.03 1. ... Qxf4 2.Ne6 fxe6 3.Qxc5 Nxb2 4.Qc8+ Kf7 5.Qb7+ Ke8 6.Bg6+ Kd8 7.Qb6+ Qc7 8.Qxc7+ Kxc7 9.Re1 Kd6 10.h5 Nc4 11.f4 Ng4 12.Be8
16+ 00:37 47,468k 1,255k -0.01 1. ... Qxf4 2.Ne6 fxe6 3.Qxc5 Nxb2 4.Qc8+ Kf7 5.Qb7+ Ke8 6.Bg6+ Kd8 7.Qb6+ Qc7 8.Qxc7+ Kxc7 9.Re1 Kd6 10.Bh7
16 00:39 49,192k 1,259k 0.00 1. ... Qxf4 2.Ne6 fxe6 3.Qxc5 Nxb2 4.Qc8+ Kf7 5.Qb7+ Ke8 6.Bxb5+ Kf8 7.Qa8+ Kf7 8.Qb7+ Kf8 9.Qa8+ Kf7 10.Qb7+ Kf8 11.Qa8+ Kf7 12.Qb7+ Kf8 13.Qa8+ Kf7 14.Qb7+ Kf8 15.Qa8+ Kf7 16.Qb7+ Kf8 17.Qa8+
17 01:38 120,692k 1,220k 0.00 1. ... Qxf4 2.Ne6 fxe6 3.Qxc5 Nxb2 4.Qc8+ Kf7 5.Qb7+ Ke8 6.Bxb5+ Kf8 7.Qa8+ Kf7 8.Qb7+ Kf8 9.Qa8+ Kf7 10.Qb7+ Kf8 11.Qa8+ Kf7 12.Qb7+ Kf8 13.Qa8+ Kf7 14.Qb7+ Kf8 15.Qa8+ Kf7 16.Qb7+ Kf8 17.Qa8+
18 03:16 237,860k 1,211k 0.00 1. ... Qxf4 2.Ne6 fxe6 3.Qxc5 Nxb2 4.Qc8+ Kf7 5.Qb7+ Ke8
19 07:46 568,500k 1,218k 0.00 1. ... Qxf4 2.Ne6 fxe6 3.Qxc5 Nxb2 4.Qc8+ Kf7 5.Qb7+ Ke8
20 23:42 1,690,100k 1,188k 0.00 1. ... Qxf4 2.Ne6 fxe6


The positional point that only strong positional monsters like stockfish can see is that white is clearly better in the following position

[d]Q4k2/6p1/p3pn1p/1B1p4/5q1P/5P2/Pn4P1/3R2K1 b - - 2 7

Here white does not have a material advantage afrer 7...Kf7 8.Be8+ Nxe8 9.Qb7+ Nc7 10.Qxb2 Qxh4(a rook is not more than a knight and 2 pawns)

[d]8/2n2kp1/p3p2p/3p4/7q/5P2/PQ4P1/3R2K1 w - - 0 11

Only a positional monster like stockfish (and maybe other top programs) see it immediately as a big advantage for white.

[CMCanavessi]

Test 10 net 11248 doesn't blunder like that, interesting