AlphaZero: The 10 Top Shots

[TommyTC]

[D] 5q2/3rkp1p/Q3p1pP/P7/5R2/6P1/5PK1/8 b - - 0 65

Cfish and 6-man tables is not seeing a win---yet.

Code: Select all

info depth 65 seldepth 128 multipv 1 score cp -27 nodes 1840738919005 nps 53839642 hashfull 997 tbhits 4861049579 time 34189286 pv f8b8 f4a4 d7d2 a6c6 b8a7 c6f3 a7c5 a4a1 c5d4 a1a3 d4b2 a3a4 b2b5 a4a1 d2d3 f3e4 d3d2 e4f4 b5b7 g2g1 b7b2 a1e1 b2d4 f4c7 e7f6 e1e3 d4d5 c7f4 f6e7 f4b4 d5d6 b4a3 d6a3 e3a3 d2d7 a5a6 d7a7 g1f1 e7d6 f1e2 f7f6 g3g4 d6c5 g4g5 f6f5

 e2e3 c5b4 a3a1 e6e5 a1d1 b4c5 d1d8 e5e4 f2f3 e4f3 e3f3 a7e7 d8b8 c5c6 f3g3 e7a7 b8f8 c6b6 f8f6 b6b5 g3f3 a7c7 f3f4 b5a5 f4e3 c7a7 e3f3 a5b5 f3f4 a7c7 f6e6 c7a7 f4f3 a7f7 f3e2 f7a7 e2f2 a7c7 f2e2

After 87...f5 [f6f5], 88. Kd3 (instead of 88. Ke3 [e2e3]) is winning (SF +5.22 d38).

[peter]

That's the problem with output- lines. You can't trust them for more then the first few plies only always
☺️

[zullil]

That's the problem with output- lines. You can't trust them for more then the first few plies only always
☺️

Yes. How about 86...g5 instead of Kc5?

[IQ]

but because I've tried enough Forward-Backward after that, and the output given is started with already full hash from all these to me relevant lines.

Fantastic, that you joined to do some real analysis!

Even from 65.Qxa6+ onwards there aren't any forced lines to draw, such Queen- and Rook- endgames are simply too long for that, White needs not trade off to the 7some positions, so to be really sure, you have to go to 50 moves boundary.

I am not that pessemistic - I think we can resolve this position to everybodys satisfaction. Using lomosov, finalgen and some known endings from the endgame manual we can go a long way. Proving a win is always 1000 times harder than proving a draw. But there are not that many key positions. And if there are some positions with a positive eval but no-one can come up with a plan to make progress, i am happy to concede that these positons might be drawish too. For example in the rook endings near the end of Loius line after Ra7 a6 black just plays f6 and this fortress cannot be penetrated and its a draw.

If I can be sure, a single game would be enough for you as a decision to your remaining questions, and if we stick to, let's say 1 move per day

I think we can resolve it here, as Louis alread found one of the key defenses with Rd2! which is much better than playing first Qa7 (as from your computer-computer games). You seem to acknowledge this because you now switched to this line too.

As I already mentioned my intial claim rested playing a6 in Loius line:

50. Bd5 Rc8 51. Qf4 Qg8 52. Bb7 Rce8 53. Qd6 c4
54. a4 c3 55. Bd5 Rd8 56. Qc7 Ke8 57. Qxc3 Qf8 58. Rf4 Re6
59. Bxe6 dxe6 60. a5 Rd7 61. Qc8+ Rd8 62. Qc6+ Rd7 63. Qc3 a6 64. Qc8+ Ke7 65. Qxa6
65...Qb8 66.Ra4 Rd2 67.Qc6 Qa7 68.Qf3 Qc5 69.Ra1 Qd4 70.Ra3 Qb2
and now
71. a6!?
when i was first analysing this I marked this as winning, but now i am not so sure anymore. What is your analysis here?

Also white can earlier play 70. Re1... how does black defend?

If these two lines crumble - than i have to dig deep and the position might after all be a draw and then i would also call Rf8 a blunder. Let me know what you find.

[Paloma]

As Peter wrote: You can't trust them for more then the first few plies only

Your 86...g5 instead of Kc5? is far behind your initial 65...f8b8, almost over 40 plies, and therefore not to trust.
You can also say - wrong.

[peter]

Proving a win is always 1000 times harder than proving a draw.

Pardon me, Roberto, but that's nonsense.
For both you need just one forced line.
Even more than that, lines to mate are much more often forced than such to draw, in both cases it just depends on the position only of course.

50. Bd5 Rc8 51. Qf4 Qg8 52. Bb7 Rce8 53. Qd6 c4
54. a4 c3 55. Bd5 Rd8 56. Qc7 Ke8 57. Qxc3 Qf8 58. Rf4 Re6
59. Bxe6 dxe6 60. a5 Rd7 61. Qc8+ Rd8 62. Qc6+ Rd7 63. Qc3 a6 64. Qc8+ Ke7 65. Qxa6
65...Qb8 66.Ra4 Rd2 67.Qc6 Qa7 68.Qf3 Qc5 69.Ra1 Qd4 70.Ra3 Qb2
and now
71. a6!?
when i was first analysing this I marked this as winning, but now i am not so sure anymore. What is your analysis here?

50. Bd5 Rc8 51. Qf4 Qg8 52. Bb7 Rce8 53. Qd6 c4
54. a4 c3 55. Bd5 Rd8 56. Qc7 Ke8 57. Qxc3 Qf8 58. Rf4 Re6 59. Bxe6 dxe6 60. a5
Rd7 61. Qc8+ Rd8 62. Qc6+ Rd7 63. Qc3 a6 64. Qc8+ Ke7 65. Qxa6

65...Qb8 66. Ra4 Rd2
67. Qc6 Qa7 68. Qf3 Qc5 69. Ra1 Qd4 70. Ra3 Qb2 71. a6

71... Rxf2+ 72. Qxf2 Qxa3 73.
Qa7+ Kf6 74. Qb6 Qd3 75. Qf2+ Ke7 76. Qc5+ Kd7 77. a7 Qa6 78. Kf2 f6 79. Ke3
Qb7 80. Qa3 Qa8 81. Kf2 Kc7 82. g4 Kb6 83. Qe3+ Kb7 84. Qf3+ Kxa7 85. Qe3+ Kb8
86. Qxe6 Qb7 87. Qxf6 Kc7 88. Qf4+ Kc6 89. Qe4+ Kc7 90. Qxb7+ Kxb7
drawn 6MOB

Also white can earlier play 70. Re1... how does black defend?

That's why I wanted to play that out move per move, not to let you come along with several moves for the white side at once questioning me which one to chose.


50. Bd5 Rc8 51. Qf4 Qg8 52. Bb7 Rce8 53. Qd6 c4
54. a4 c3 55. Bd5 Rd8 56. Qc7 Ke8 57. Qxc3 Qf8 58. Rf4 Re6 59. Bxe6 dxe6 60. a5
Rd7 61. Qc8+ Rd8 62. Qc6+ Rd7 63. Qc3 a6 64. Qc8+ Ke7 65. Qxa6 Qb8 66. Ra4 Rd2
67. Qc6 Qa7 68. Qf3 Qc5 69. Ra1 Qd4 70. Re1


70...Qa4 71. Qb7+ Rd7 72. Qb6 Rd6 73.
Qb7+ Rd7 74. Qa8 Qb4 75. Ra1 Qb2 76. Re1 Qb4 77. Rc1 Rd5 78. Qc8 Rd8 79. Qa6
Rd5 80. Qc8 Rd8 81. Qa6 Rd5 82. Qb6 Qxb6 83. axb6 Rd7 84. Rb1 Rb7 85. Kf3 Kd7
86. g4 g5 87. Ke4 f6 88. f3 Kc6

Yet you still can go on playing from an "ending" like the one of the second line, as long as you want, but if you make a Backward of such lines, you get an output like this, even just after our 69...Qd4, the last move we are both sure of, aren't we?

[d]8/4kp1p/4p1pP/P7/3q4/5QP1/3r1PK1/R7 w - - 0 1

Analysis by Houdini 6.03 Pro x64-popc:

70.Ra3 Qb2 71.Ra4 Qb5 72.Ra1 Rd3 73.Qe4 Rd2 74.Qf4 Qb7+ 75.Kg1 Qb2 76.Rf1 Qd4 77.Qg5+ Kf8 78.Qb5 Ra2 79.Qb1 Ra3 80.Qb8+ Ke7 81.Rb1 Rd3 82.Qb7+ Kf6 83.Kg2 Qd5+ 84.Qxd5 Rxd5 85.Ra1 Kg5 86.Ra4 f5 87.f4+ Kxh6 88.a6 Rd2+ 89.Kf3 Rd8 90.a7 Ra8 91.Ra6 Kh5 92.Ra3 g5 93.Ra6 h6 94.Ra4 Kg6 95.Ra2 Kf7 96.Ke3 Kg6 97.Ra6 Kf6 98.Ra5 Kg6 99.Ra1
= (0.15) Depth: 37/111 00:16:13 34651MN, tb=57019134

Of course you have to do Backward from more than one line to get such low evals in such a short computing- time, but you as an old analyst do know that anyhow. The 32G hash were already full when coming back to the position of the dia for the last time.
And in this case it was somewhat easier, cause I had 32G hash already stored at 65.Qa6+, where you saw the output of Houdini's too in the posting here:

http://www.talkchess.com/forum/viewtopi ... 11&t=66184

Now, that you started to look at single output- lines more than at reasonable games obviously, I'm sure you are an educated reader of such output- lines and don't take them for reliable for more then the first few plies only. If you want to take some moves from me, stick to one of the two lines given, don't take the output- line, even if I checked it up to 80th move and it seemed ok so far too.

And for our further analysis please decide for one of the two lines to hop in with an alternative white move whereever you want. From there onward of course you can give several moves again, if I'm free to change any of your moves for Black's side then again.
And stick to one more try, as if we'd have a real game, not a Lyudmil Deja Vu.

And If we reach 100 moves without any progress for White, you'll have to go on on your own alone again, Roberto

[Guenther]

Proving a win is always 1000 times harder than proving a draw.

Pardon me, Roberto, but that's nonsense.
For both you need just one forced line.
Even more than that, lines to mate are much more often forced than such to draw, in both cases it just depends on the position only of course.

Simple logic tells you are right.

1. We always start at a position which has an unknown result, otherwise it would be useless
2. On the way of the analyis we find win/loss lines
3. As long as we cannot find/prove win/loss lines we still have to assume a draw
4. This means we have to go a much longer path for proving a draw

[zullil]

As I already mentioned my intial claim rested playing a6 in Louis's line:

50. Bd5 Rc8 51. Qf4 Qg8 52. Bb7 Rce8 53. Qd6 c4
54. a4 c3 55. Bd5 Rd8 56. Qc7 Ke8 57. Qxc3 Qf8 58. Rf4 Re6
59. Bxe6 dxe6 60. a5 Rd7 61. Qc8+ Rd8 62. Qc6+ Rd7 63. Qc3 a6 64. Qc8+ Ke7 65. Qxa6
65...Qb8 66.Ra4 Rd2 67.Qc6 Qa7 68.Qf3 Qc5 69.Ra1 Qd4 70.Ra3 Qb2
and now
71. a6!?
when i was first analysing this I marked this as winning, but now i am not so sure anymore. What is your analysis here?

[D] 8/4kp1p/P3p1pP/8/8/R4QP1/1q1r1PK1/8 b - - 0 71
Asmfish doesn't see 71. a6 as winning:

Code: Select all

0.00  71... Rxf2+ 72. Qxf2 Qxa3 73. Qa7+ Kf6 74. Qb6 Qd3 75. Qf2+ Ke7 76. Qc5+ Kd7 77. a7 Qa6 78. Kf2 f6 79. Ke3 Qb7 80. Qa3 Qa8 81. Kf2 Kc7 82. Qe7+ Kb6 83. Qxh7 Qd5 84. a8=N+ Qxa8 85. Qxg6 Qa2+ 86. Kg1 Qb3 87. Kh2 Qf3 88. Qb1+ Ka6 89. Qa1+ Kb5 90. Qb2+ Ka6 91. Qd2 e5 92. Qd6+ Kb5 93. Qd7+ Kc4 94. Qf7+ Kd4 95. Qa7+ Kc4 96. Qf7+ (depth 51, 0:06:18)

[zullil]

As Peter wrote: You can't trust them for more then the first few plies only

Your 86...g5 instead of Kc5? is far behind your initial 65...f8b8, almost over 40 plies, and therefore not to trust.
You can also say - wrong.

My point was to see if folks agreed that 86...g5 draws. If so, back-tracking can be done from there to see if the earlier part of the line is sound.

[peter]

[[D] 8/4kp1p/P3p1pP/8/8/R4QP1/1q1r1PK1/8 b - - 0 71
Asmfish doesn't see 71. a6 as winning:

Code: Select all

0.00  71... Rxf2+ 72. Qxf2 Qxa3 73. Qa7+ Kf6 74. Qb6 Qd3 75. Qf2+ Ke7 76. Qc5+ Kd7 77. a7 Qa6 78. Kf2 f6 79. Ke3 Qb7 80. Qa3 Qa8 81. Kf2 Kc7 82. Qe7+ Kb6 83. Qxh7 Qd5 84. a8=N+ Qxa8 85. Qxg6 Qa2+ 86. Kg1 Qb3 87. Kh2 Qf3 88. Qb1+ Ka6 89. Qa1+ Kb5 90. Qb2+ Ka6 91. Qd2 e5 92. Qd6+ Kb5 93. Qd7+ Kc4 94. Qf7+ Kd4 95. Qa7+ Kc4 96. Qf7+ (depth 51, 0:06:18)

That output- line is identical to the line I gave here:

http://www.talkchess.com/forum/posting. ... e&p=746805

50. Bd5 Rc8 51. Qf4 Qg8 52. Bb7 Rce8 53. Qd6 c4
54. a4 c3 55. Bd5 Rd8 56. Qc7 Ke8 57. Qxc3 Qf8 58. Rf4 Re6 59. Bxe6 dxe6 60. a5
Rd7 61. Qc8+ Rd8 62. Qc6+ Rd7 63. Qc3 a6 64. Qc8+ Ke7 65. Qxa6

65...Qb8 66. Ra4 Rd2
67. Qc6 Qa7 68. Qf3 Qc5 69. Ra1 Qd4 70. Ra3 Qb2 71. a6

71... Rxf2+ 72. Qxf2 Qxa3 73.
Qa7+ Kf6 74. Qb6 Qd3 75. Qf2+ Ke7 76. Qc5+ Kd7 77. a7 Qa6 78. Kf2 f6 79. Ke3
Qb7 80. Qa3 Qa8 81. Kf2 Kc7 82. g4 Kb6 83. Qe3+ Kb7 84. Qf3+ Kxa7 85. Qe3+ Kb8
86. Qxe6 Qb7 87. Qxf6 Kc7 88. Qf4+ Kc6 89. Qe4+ Kc7 90. Qxb7+ Kxb7
drawn 6MOB

up to move 81, than changing to another line in my database up to move nr. 86, where with 87.Qxf6 a drawn 7- some is reached again, which to me is proving enough again.

But the rest of the output- line seems to be ok too, as far as I looked at it quickly only from 87th onwards,