AlphaZero: The 10 Top Shots

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

Moderators: hgm, Dann Corbit, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
zullil
Posts: 6442
Joined: Mon Jan 08, 2007 11:31 pm
Location: PA USA
Full name: Louis Zulli

Re: AlphaZero: The 10 Top Shots

Post by zullil » Mon Jan 08, 2018 3:57 pm

IQ wrote:
Also white can earlier play 70. Re1... how does black defend?
[D] 8/4kp1p/4p1pP/P7/3q4/5QP1/3r1PK1/4R3 b - - 10 70
Cfish offers:

Code: Select all

0.00 70... Qa4 71. Rb1 Rxf2+ 72. Qxf2 Qe4+ 73. Kh2 Qxb1 74. a6 Qa1 75. Qc5+ Kd7 76. a7 Qa4 77. Kg2 f6 78. Kf2 Qa2+ 79. Ke3 Qb3+ 80. Kd2 Qb2+ 81. Kd1 Qb1+ 82. Ke2 Qb2+ 83. Kd3 Qb3+ 84. Ke4 Qb1+ 85. Kd4 Qb3 86. Ke4 (depth 69, 1:18:32)

zullil
Posts: 6442
Joined: Mon Jan 08, 2007 11:31 pm
Location: PA USA
Full name: Louis Zulli

Re: AlphaZero: The 10 Top Shots

Post by zullil » Mon Jan 08, 2018 8:55 pm

zullil wrote:
IQ wrote:
Also white can earlier play 70. Re1... how does black defend?
[D] 8/4kp1p/4p1pP/P7/3q4/5QP1/3r1PK1/4R3 b - - 10 70
Cfish offers:

Code: Select all

0.00 70... Qa4 71. Rb1 Rxf2+ 72. Qxf2 Qe4+ 73. Kh2 Qxb1 74. a6 Qa1 75. Qc5+ Kd7 76. a7 Qa4 77. Kg2 f6 78. Kf2 Qa2+ 79. Ke3 Qb3+ 80. Kd2 Qb2+ 81. Kd1 Qb1+ 82. Ke2 Qb2+ 83. Kd3 Qb3+ 84. Ke4 Qb1+ 85. Kd4 Qb3 86. Ke4 (depth 69, 1:18:32)
And after deep search with large hash and back-tracking to the root, Cfish says 70....Qa4 draws. Here are the three "best" lines for White:

Code: Select all

0.00  71. Qb7+ Rd7 72. Qb6 Rd6 73. Qc7+ Qd7 74. Rc1 Qxc7 75. Rxc7+ Kf6 76. Kf3 Rd3+ 77. Ke4 Ra3 78. Rc5 Ra2 79. f4 Ke7 80. g4 f5+ 81. gxf5 gxf5+ 82. Kd4 Ra4+ 83. Ke3 Kd6 84. Rb5 Ra3+ 85. Kd4 Ra4+ 86. Ke3 (depth 70, 0:27:29)

0.00  71. Rc1 Qxa5 72. Rc8 Rd7 73. Rh8 Qd5 74. Qxd5 exd5 75. Rxh7 Rd8 76. Kf3 Ra8 77. Ke3 Kf6 78. Kd4 Ra2 79. Kxd5 Rxf2 80. Ke4 Re2+ 81. Kf3 Re5 82. Rh8 Rf5+ 83. Kg4 Rg5+ 84. Kf4 Rf5+ 85. Kg4 (depth 70, 0:27:29)

0.00  71. Rb1 Rxf2+ 72. Qxf2 Qe4+ 73. Kh2 Qxb1 74. a6 Qa1 75. Qc5+ Kd7 76. a7 Qa4 77. Kg2 f6 78. Kf2 Qa2+ 79. Ke3 Qb3+ 80. Kd2 Qb2+ 81. Kd1 Qb1+ 82. Ke2 Qb2+ 83. Kd3 Qb3+ 84. Ke4 Qb1+ 85. Kd4 Qb3 86. Ke4 (depth 70, 0:27:29)

peter
Posts: 2246
Joined: Sat Feb 16, 2008 6:38 am
Full name: Peter Martan

Re: AlphaZero: The 10 Top Shots

Post by peter » Mon Jan 08, 2018 11:08 pm

zullil wrote:

Code: Select all

0.00  71. Qb7+ Rd7 72. Qb6 Rd6 73. Qc7+ Qd7 74. Rc1 Qxc7 75. Rxc7+ Kf6 76. Kf3 Rd3+ 77. Ke4 Ra3 78. Rc5 Ra2 79. f4 Ke7 80. g4 f5+ 81. gxf5 gxf5+ 82. Kd4 Ra4+ 83. Ke3 Kd6 84. Rb5 Ra3+ 85. Kd4 Ra4+ 86. Ke3 (depth 70, 0:27:29)
Thank you, Louis, first two moves transpose to my corresponding line from here again,
http://www.talkchess.com/forum/posting. ... e&p=746805
, then leading to a very similar rook- ending like mine was after exchange of Queens.
Backward lets keep 0.00 in hash quickly till 73.Qc7+, where your output- line starts to differ from mine,
Peter.

jdart
Posts: 4115
Joined: Fri Mar 10, 2006 4:23 am
Location: http://www.arasanchess.org

Re: AlphaZero: The 10 Top Shots

Post by jdart » Wed Jan 10, 2018 4:20 am

This one (game 1) looks really interesting:

[D] rn3r1k/pn1p1ppq/bpp4p/7P/4N1Q1/6P1/PP3PB1/R1B1R1K1 w - - 0 21

AlphaZero played Bg5. ChessBase indicates Stockfish 171217 got this, with a score of +2.21, at a depth of 58.

--Jon

User avatar
mclane
Posts: 18403
Joined: Thu Mar 09, 2006 5:40 pm
Location: US of Europe, germany
Full name: Thorsten Czub
Contact:

Re: AlphaZero: The 10 Top Shots

Post by mclane » Wed Jan 10, 2018 5:42 am

Cool. 58 plies alpha beta to see something AZ saw in 1‘ ?!
What seems like a fairy tale today may be reality tomorrow.
Here we have a fairy tale of the day after tomorrow....

Milos
Posts: 4055
Joined: Wed Nov 25, 2009 12:47 am

Re: AlphaZero: The 10 Top Shots

Post by Milos » Wed Jan 10, 2018 6:09 am

mclane wrote:Cool. 58 plies alpha beta to see something AZ saw in 1‘ ?!
I admire your ignorance. When ppl who never had a contact with certain field have no clue about it, that can be understood. However, when someone who spent his whole life in one field essentially has no clue about it, that only shows that he's really not a very bright person.
UCT without rollouts is way more selective than alpha-beta, meaning that depth that A0 reached with 80k sims/s was most probably noticeably larger than of SF with 70Mnps.
But what I am writing is hopeless coz you didn't even understand that A0 is using search. Based on what you wrote up there, you are clearly thinking that A0 has only NN eval :shock: :shock:.

User avatar
Ovyron
Posts: 4415
Joined: Tue Jul 03, 2007 2:30 am

Re: AlphaZero: The 10 Top Shots

Post by Ovyron » Wed Jan 10, 2018 6:32 am

Wow, Milos, you have a formidable ability to deduce what other people are able to understand or not, are you a psychic or a medium? :roll:

User avatar
mclane
Posts: 18403
Joined: Thu Mar 09, 2006 5:40 pm
Location: US of Europe, germany
Full name: Thorsten Czub
Contact:

Re: AlphaZero: The 10 Top Shots

Post by mclane » Wed Jan 10, 2018 7:08 am

It's unimportant if az has search, NN or masturbates to find the move.
It can even watch out the stars in the sky .

Stockfish had no idea how to find a way out of the mud, AZ pulled stockfish in.
(I remember a time when Chess System Tal did the same strategy with genius, fritz, Mchess. This was very similar and funny to watch on the autoplayers. )

The whole thing would not have been better if there was Houdini or Komodo as opponent.

These 3 variants are so similar that there is nearly no difference between them.
In 100 games the best stockfish could afford was a draw.

The fact that stockfish alpha beta search needs 58 Plies to find a move shows the
reason it finds no way out.

It's ineffective.
It was effective enough to make 10 Elo more then Komodo, and Komodo 10 Elo better than Houdini, or vice versa. But the incest on top of the rating lists is a bluff.
The bluff is that these 10 Elo mean the program is BETTER when in fact it is better against the other equal programs.
In the moment a foreigner comes into the group, the whole Elo progress idea gets
Ruined.

Now you can continue insulting.

It's a paradigm shift.
What seems like a fairy tale today may be reality tomorrow.
Here we have a fairy tale of the day after tomorrow....

zullil
Posts: 6442
Joined: Mon Jan 08, 2007 11:31 pm
Location: PA USA
Full name: Louis Zulli

Re: AlphaZero: The 10 Top Shots

Post by zullil » Wed Jan 10, 2018 11:41 am

jdart wrote:This one (game 1) looks really interesting:

[D] rn3r1k/pn1p1ppq/bpp4p/7P/4N1Q1/6P1/PP3PB1/R1B1R1K1 w - - 0 21

AlphaZero played Bg5. ChessBase indicates Stockfish 171217 got this, with a score of +2.21, at a depth of 58.

--Jon
Latest Cfish also finds Bg5:

Code: Select all

info depth 45 seldepth 61 multipv 1 score cp 86 lowerbound nodes 45179139148 nps 36870184 hashfull 999 tbhits 0 time 1225357 pv c1g5

Code: Select all

info depth 45 seldepth 79 multipv 1 score cp 130 nodes 61294494274 nps 37460966 hashfull 999 tbhits 0 time 1636223 pv c1g5 f7f5 g4f4 b7c5 f4d6 h8g8 g5e7 c5e4 g2e4 f5e4 e7f8 a6d3 f2f3 b8a6 f3e4 a8f8 d6d3 a6c5 d3c4 c5e6 a1d1 g7g6 e1f1 g6h5 c4c3 f8f1 g1f1 h7f7 f1g2 e6g5 e4e5 f7g7 e5e6 d7d5 c3c6 g7b2 g2h1 b2e2 c6d5 g5e6 d5a8 g8g7 d1d7 g7g6 a8g8 g6f5 d7f7 f5e5 g8b8 e5d5 f7f5 d5c4 b8a8 c4b4 a8b7 h5h4 b7e7 b4c3 e7a3 c3c2 g3h4 c2b1 a3b3 b1a1 b3d5 e6c5 f5f4 e2e1 h1g2 e1e2 f4f2 e2e4 d5e4 c5e4

User avatar
Spacious_Mind
Posts: 317
Joined: Sun Nov 01, 2009 11:05 pm
Location: Alabama
Contact:

Re: AlphaZero: The 10 Top Shots

Post by Spacious_Mind » Wed Jan 10, 2018 12:10 pm

zullil wrote:
jdart wrote:This one (game 1) looks really interesting:

[D] rn3r1k/pn1p1ppq/bpp4p/7P/4N1Q1/6P1/PP3PB1/R1B1R1K1 w - - 0 21

AlphaZero played Bg5. ChessBase indicates Stockfish 171217 got this, with a score of +2.21, at a depth of 58.

--Jon
Latest Cfish also finds Bg5:

Code: Select all

info depth 45 seldepth 61 multipv 1 score cp 86 lowerbound nodes 45179139148 nps 36870184 hashfull 999 tbhits 0 time 1225357 pv c1g5

Code: Select all

info depth 45 seldepth 79 multipv 1 score cp 130 nodes 61294494274 nps 37460966 hashfull 999 tbhits 0 time 1636223 pv c1g5 f7f5 g4f4 b7c5 f4d6 h8g8 g5e7 c5e4 g2e4 f5e4 e7f8 a6d3 f2f3 b8a6 f3e4 a8f8 d6d3 a6c5 d3c4 c5e6 a1d1 g7g6 e1f1 g6h5 c4c3 f8f1 g1f1 h7f7 f1g2 e6g5 e4e5 f7g7 e5e6 d7d5 c3c6 g7b2 g2h1 b2e2 c6d5 g5e6 d5a8 g8g7 d1d7 g7g6 a8g8 g6f5 d7f7 f5e5 g8b8 e5d5 f7f5 d5c4 b8a8 c4b4 a8b7 h5h4 b7e7 b4c3 e7a3 c3c2 g3h4 c2b1 a3b3 b1a1 b3d5 e6c5 f5f4 e2e1 h1g2 e1e2 f4f2 e2e4 d5e4 c5e4
SF8 if you play the next 10 moves of the game forward and follow them you will see SF8 should really have found Bg5 as best move at ply 26 or 27 if I recall when I tried it.

Same with pretty much all the other test positions that I tried. Each one of them should have been found in less than 30 ply.

Since I am not a chess programmer I can only surmise that SF8 and all the other modern programs nowadays cut off other interesting chess move alternatives in their search decisions too soon and therefore don't see the obvious anymore.

Pretty much in my opinion all that A0 has to do is let SF8 play against itself and beat itself every time by sending FENs to each of its 1024 threads and let them search about 18 deep. Keep repeating that about 10 ply deep and you pretty much find every move in 60 seconds and less than 30 ply deep. ie... 10 ply deep repeated fen shots plus 18 ply = 28 ply and pretty much everyone one of these moves are found by SF8 without having to search 59 ply.

To find something in 59 plies that should be found in 28 plies suggests that there are some major gaps in how searches behave today in programs.

If you act like a correspondence player and manually click through the moves you will see how SF finds these moves much quicker than 59 moves.

Best regards

Post Reply