A0 has same "tactical blindness bug" as Lc0

[M ANSARI]

I have recently downloaded the A0 games with SF and was interested in how similar the play of A0 was compared to Lc0. It is quite remarkable that Lc0 wants to play almost the same exact moves as A0. Of course A0 has much stronger hardware and on some moves it will take some time for Lc0 to choose A0 move ... but otherwise they play almost identical moves. On another post I mentioned that Lc0 suffered from a "tactical blindess" bug where in one move it can totally throw away a game and that this was the main cause of many of its losses. I wanted to see if A0 also suffered from this. In 100 games against SF, A0 only lost 3 games and here is one of the games. Albert Silver had mentioned on that post that A0 also suffers from tactical blindness on some moves so would be interesting to see. I downloaded the 100 games of A0 vs SF games and looked through one of the losses. Here is the game

[Event "Computer Match"]
[Site "London, UK"]
[Date "2018.01.18"]
[Round "255"]
[White "AlphaZero"]
[Black "Stockfish 8"]
[Result "0-1"]
[PlyCount "224"]
[EventDate "2018.??.??"]

1. d4 Nf6 2. Nf3 e6 3. c4 d5 4. Nc3 c6 5. Bg5 Be7 6. e3 O-O 7. cxd5 exd5 8. Bd3
Nbd7 9. h3 Re8 10. O-O Ne4 11. Bf4 Ndf6 12. Ne5 Bd6 13. Rc1 Qe7 14. Re1 g6 15.
Qc2 Kg7 16. Na4 Nh5 17. Bh2 Kg8 18. a3 Ng7 19. Bxe4 dxe4 20. Nc3 f5 21. Nc4
Bxh2+ 22. Kxh2 Be6 23. Ne2 c5 24. Red1 cxd4 25. Rxd4 Red8 26. Rxd8+ Qxd8 27.
Nd2 Rc8 28. Qd1 Rxc1 29. Qxc1 Bd5 30. g3 Ne6 31. b4 Qd7 32. Kg2 Kf7 33. Qc3 Qb5
34. Nf4 Nxf4+ 35. gxf4 Qc6 36. Qh8 h5 37. Qh7+ Kf8 38. Nb1 a6 39. Qh8+ Kf7 40.
Nc3 Bc4 41. Qd4 b5 42. Qa7+ Kf8 43. Qd4 Kf7 44. Qd8 Qe6 45. Kh2 Qf6 46. Qxf6+
Kxf6 47. h4 Ke6 48. Kh1 Bf1 49. Kg1 Bc4 50. Kg2 Bd5 51. Kf1 Bc4+ 52. Ke1 Bd3
53. Kd1 Kd6 54. Kd2 Bc4 55. Kc2 Ke6 56. Kd1 Bd3 57. Kd2 Bc4 58. Kc1 Bd3 59. Kb2
Kd6 60. Ka2 Ke6 61. Kb3 Kd7 62. a4 Bc4+ 63. Ka3 Kc6 64. a5 Bd3 65. Kb2 Kd6 66.
Kc1 Kc6 67. Kd1 Kd6 68. Kd2 Bc4 69. Kc2 Bd3+ 70. Kb3 Bc4+ 71. Ka3 Bd3 72. Kb2
Ke6 73. Kc1 Kd6 74. Kd1 Kc6 75. Kd2 Bc4 76. Kc2 Kd6 77. Kb1 Kc6 78. Kb2 Bd3 79.
Na2 Kd5 80. Nc1 Kc4 81. Ka3 Bf1 82. Nb3 Kd3 83. Nc5+ Ke2 84. Nxa6 Kxf2 85. Nc7
Kxe3 86. a6 Kf2 87. a7 e3 88. a8=Q e2 89. Qa7+ Kf3 90. Qa8+ Kg3 91. Qe8 Kf2 92.
Qxb5 e1=Q 93. Qe5 Qc1+ 94. Ka4 Kg3 95. Nd5 Kxh4 96. Nc3 Kg3 97. Ne2+ Bxe2 98.
Qxe2 h4 99. Qd3+ Kg4 100. Qe2+ Kxf4 101. b5 Qa1+ 102. Kb4 Qd4+ 103. Ka5 Qa7+
104. Kb4 h3 105. Qf1+ Kg3 106. Qd3+ Kh4 107. Qc4+ Kg5 108. Qc3 Qe7+ 109. Ka5
Kg4 110. Qd4+ Qe4 111. Qg1+ Kh4 112. b6 Qd5+ 0-1

Looking at the position it looked like A0 had an advantage. Let's look at the position at move 75

[d]8/8/p1k3p1/Pp3p1p/1P2pP1P/2NbP3/3K1P2/8 b - - 0 75

Although the position is probably drawn, the white Knight seems much stronger than the white bishop as if white can place his Knight on c5 or e5 the pawns on a6 and g6 can be attacked and the black bishop cannot protect them. Thus the black King has to protect those pawns and if he does, then maybe the white King can penetrate via c3 and d4. Now black has no chances of winning because all white has to do is keep the black King from entering via c4 and that seems very easy to do. At least that is how I see the position. A few moves later we get this position

[d]8/8/p5p1/Pp1k1p1p/1P2pP1P/3bP3/NK3P2/8 w - - 0 80

Here black tries to enter via c4 ... but even an 800 ELO player would know that all white has to do is play Kc3 and stop the black King from entering. But incredibly A0 ... the engine that shocked the world with some alien like chess ... plays Nc1 ????? and loses in one move. Here is SF analysis on this position at 3 PV mode. SF obviously sees the Kc3 is the only playable move to avoid losing instantly.

38: AlphaZero - Stockfish 8, Computer Match 2018
8/8/p5p1/Pp1k1p1p/1P2pP1P/3bP3/NK3P2/8 w - - 0 1

Analysis by Stockfish 060119 64 POPCNT:

1. = (0.00): 80.Kc3 Bc4 81.Nc1 Ke6 82.Kc2 Kd6 83.Kd2 Kc6 84.Kc2
2. -+ (-3.38): 80.Ka3 Kc4 81.Nc1 Bf1 82.Nb3 Kd3 83.Nc5+ Ke2 84.Nxa6 Kxf2 85.Nc7 Kxe3 86.a6 Kf2 87.a7 e3 88.a8Q e2 89.Qa7+ Kf3 90.Qa8+ Kg3 91.Qe8 Kf2 92.Nd5 e1Q 93.Qxe1+ Kxe1 94.Kb2 Kf2 95.Ne7 Bc4 96.Nxg6 Kg3 97.Kc3 Bf7 98.Ne7 Kxf4 99.Nc8 Be8 100.Nd6 Bc6 101.Kd4 Kg4 102.Kc5 Bd7 103.Nf7 Kxh4 104.Ne5 Be8 105.Kd6 Kg3 106.Ke7 h4 107.Kxe8 h3
3. -+ (-3.56): 80.Nc1 Kc4 81.Ka3 Bf1 82.Nb3 Kd3 83.Kb2 Ke2 84.Kc3 Kxf2 85.Kd4 Kg3 86.Nc5 Kxh4 87.Nxa6 Kg3 88.Nc7 h4 89.a6 h3 90.a7 h2 91.a8Q h1Q 92.Qg8 Kf2 93.Kc5 Qg1 94.Qh8 Bc4 95.Nxb5 Bxb5 96.Kxb5 Kxe3 97.Kc6 Kxf4 98.b5 e3 99.b6 e2 100.b7 Qc1+ 101.Kd5 Qd2+ 102.Qd4+ Qxd4+ 103.Kxd4 e1Q 104.b8Q+ Kg5 105.Qb7 Qd2+ 106.Kc4 Kg4 107.Qb6 Qe2+ 108.Kc3 g5 109.Qg1+ Kh5 110.Kd4 Qe4+ 111.Kc3

(Spearfish, 02.02.2019)


Now I had seen similar tactical blindness with Lc0 and decided to see what it thought in this position

38: AlphaZero - Stockfish 8, Computer Match 2018
8/8/p5p1/Pp1k1p1p/1P2pP1P/3bP3/NK3P2/8 w - - 0 1

Analysis by Lc0 v0.20.1:

1. = (0.07): 80.Kc3 Be2 81.Nc1 Bd1 82.Kd2 Ba4 83.Kc3 Bd1 84.Na2 Be2 85.Nc1 Bd1
2. = (0.02): 80.Ka3 Bc2 81.Kb2 Bd3 82.Kc3 Be2 83.Nc1 Bd1 84.Nb3 Bxb3 85.Kxb3 Ke6 86.Kc2 Kd6 87.Kd2 Ke6 88.Kc3 Ke7 89.Kd4 Kd6
3. = (-0.07): 80.Nc1 Kc4 81.Ka3 Bc2 82.Ne2 Bd1 83.Nd4 Bf3 84.Nc6 Kd3 85.Nd4 Kc4 86.Nc6 Kd3 87.Nd4 Kc4

(Spearfish, 02.02.2019)


While Lc0 would have played Kc3 here, maybe with more time it would have chosen Nc1. After all, A0 plays with much more powerful hardware. But more worrying is that it could easily have chosen Ka3?? also losing !!! There is very little difference in evaluation between the moves even though it is painfully obvious that allowing the black King to c4 loses. So maybe this is something that NN suffer from and that needs to be fixed.