AlphaZero: The 10 Top Shots

[Milos]

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?

It's called a bullshit detector and in your case it was glowing red in most of your "technical" posts .

[jdart]

To clarify: it means the move was selected as of depth 58, it doesn't mean it was not selected at earlier depths. This is from the ChessBase online book, which only shows the last result, not the whole search history.

--Jon

[IQ]

And after deep search with large hash and back-tracking to the root, Cfish says 70....Qa4 draws.

Yes. In view that both lines especially the a6 lines fail - I am changing my assessment of the position. Now I believe that the position after Kf8 is very likely a draw, even if the defense is hanging by a thin thread.

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. Ra3 Qb2 71. a6 Rxf2+ 72. Qxf2 Qxa3 73.Qa7+ Kf6 74. Qb6 Qd3!

From afar the position after 74. Qb6 looked promising but Qd3! kills it - its not completely clear, but i see no way to make progress in the various resulting queen endings (with white having a free pawn on h7).

So finally i also consider Rf8 a blunder and not a mere mistake. I still suspect that there must be a way for white to improve in the Bd5 line or that one can refine the move order and play bd5 later - but i cannot make it work. Maybe somebody here has another try which might give white winning chances? Louis? Peter? Maybe you found some problematic line which poses more problems for black even if you ultimately thought it to be a draw - maybe these will provide some new ideas.

[IQ]

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

This logic is flawed as there are much more drawing lines than winning lines in chess. The draw margin is very high. Also (3) and (4) are contradictory - if you assume that a position that cannot be proven to be a win or loss is a draw (see 3) then you can shorten the path considerably and stop right there. While if you claim a position is winning you have to present the complete line from that position to a known win (or sufficent high eval). Also (2) is generally not true as there might not be a true winning line. And (3) is quite biased isn't it?!

Practically i see it like this:

a) the draw margin is very high
b) many positions that look promising can be held by exact defense
c) engines with (lomosov) tablebases and some human assitance in fortress positions are extremely good at finding exact defenses
d) if a position yields a positiv eval but no progress can be seen in the main line and/or the eval plateus over a large number of iterations the burden of proof is on the one who clames a win/loss. This is analogus to your third assumption, but makes it harder to prove a win/loss - and much easier claim a draw.

[pilgrimdan]

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.

Thorsten...

I normally don't like what you post...

but this time you hit the nail on the head...

+1

[zullil]

To clarify: it means the move was selected as of depth 58, it doesn't mean it was not selected at earlier depths. This is from the ChessBase online book, which only shows the last result, not the whole search history.

--Jon

OK. So Cfish says:

Code: Select all

info depth 60 seldepth 94 multipv 1 score cp 265 nodes 1539816502751 nps 42005352 hashfull 999 tbhits 0 time 36657626 pv c1g5 f7f5 g4f4 b7c5 g5e7 c5e6 f4d6 f5e4 e7f8 a6d3 a1d1 h7f5 f8e7 a7a5 e1e3 h8g8 d1e1 e6c5 f2f3 f5e6 f3e4 d3c4 d6e6 c4e6 e7c5 b6c5 e3a3 g8f8 e4e5 a8a7 b2b3 f8e7 a3a4 c5c4 b3c4 e6g4 c4c5 g4h5 g1f2 h5f7 e1b1 a7a8 f2e3 f7d5 g2e4 d5f7 e3d4 f7e6 d4c3 e6f7 b1b2 f7e6 c3d4 e7f7 b2f2 f7e8 e4g6 e8e7 d4e3 a8a7 f2b2 a7a8 b2b6 e6g8 b6b7 g8d5 e3d4 e7e6 b7b2 e6e7 g6e4

Ending search there.

[Spacious_Mind]

To clarify: it means the move was selected as of depth 58, it doesn't mean it was not selected at earlier depths. This is from the ChessBase online book, which only shows the last result, not the whole search history.

--Jon

OK. So Cfish says:

Code: Select all

info depth 60 seldepth 94 multipv 1 score cp 265 nodes 1539816502751 nps 42005352 hashfull 999 tbhits 0 time 36657626 pv c1g5 f7f5 g4f4 b7c5 g5e7 c5e6 f4d6 f5e4 e7f8 a6d3 a1d1 h7f5 f8e7 a7a5 e1e3 h8g8 d1e1 e6c5 f2f3 f5e6 f3e4 d3c4 d6e6 c4e6 e7c5 b6c5 e3a3 g8f8 e4e5 a8a7 b2b3 f8e7 a3a4 c5c4 b3c4 e6g4 c4c5 g4h5 g1f2 h5f7 e1b1 a7a8 f2e3 f7d5 g2e4 d5f7 e3d4 f7e6 d4c3 e6f7 b1b2 f7e6 c3d4 e7f7 b2f2 f7e8 e4g6 e8e7 d4e3 a8a7 f2b2 a7a8 b2b6 e6g8 b6b7 g8d5 e3d4 e7e6 b7b2 e6e7 g6e4

Ending search there.

I have previously run this with full MPV and gave up after about 4 hours and about 50 depth. SF8 does not see Bg5 as best move. From someone else's test I know it only sees it as best move at 59 ply which you are confirming.

If you were to do full MPV in Arena, search the first move for 18 ply, then play Bg5, do the Black move for 18 ply, and then follow that with the next AO white move for 18 ply and so on for 5 full moves (10 half moves) you will then slowly follow how Bg5 starts to move closer to the top of the search tree.

Finally after you have manually played the first 5 moves for white and 5 Black responses you will see in your search tree that SF8 at around 16 Ply switches and makes Bg5 the best move. From that point onward it keeps it as best move.

So if you add up the plies you have 10 manual half moves that you made plus a search depth of 16 ply = 26 ply. SF8 itself sees Bg5 as best move in 26 ply!! Now on my I7 using full MPV in Arena, I can search manually 18 ply deep 10 half moves and it takes 1 minute and 19 seconds in total by adding up the seconds to find Bg5!

I am pretty sure that I dont have the fastest computer around. Therefore with 1024 threads and decent speed it would be a sinch to find the best moves to beat SF8 every time in 60 seconds.

As I stated before SF8 just has to play against itself and it will beat itself every time.

However if you let SF8 search from the start position then it takes 59 plies and what 5-8 hours of searching?

Do the same with all the test positions and the same results will repeat themselves every time.

It seems to me:

More knowledge width + less knowledge depth = 27 Wins (=more human in thought processing)
More knowledge depth + less knowledge width = 0 Wins.

Depth + small width is no longer the solution as A0 has shown.

Regards

[zullil]

SF8 does not see Bg5 as best move. From someone else's test I know it only sees it as best move at 59 ply which you are confirming.

I have not tested Stockfish 8 on this position. As I showed, the latest Cfish found Bg5 at depth 45, after about 20 minutes. Of course, since it used 20 threads, that search was highly non-deterministic. I have no idea how many nodes a single-threaded Cfish would need to find this move.

[Henk]

Are they going to discuss next few years about these scarce alpha zero games that were published. Alpha zero does not play anymore or not ? So it is a dead engine.

[Spacious_Mind]

Are they going to discuss next few years about these scarce alpha zero games that were published. Alpha zero does not play anymore or not. So it is a dead engine.

Henk,

As far as I am concerned Alpha can post a million games and the same pattern will repeat itself. The loss moves will be found in less than 30 ply.

All that is probably happening is that you have 1000 independent search threads and a master chess program shooting fens to the threads multiple times. The individual threads could be receiving multiple search instructions in 60 seconds. Master program then picks move. It doesn't sound that hard to me.

Regarding intelligence... save the winning FEN's and have a search program search and match against the 1000 cores in play. If it is already in your db then you can play it. It's like you searching something in Google and it comes up instantly.

The approach could be as simple as this. Just like we are manually shooting FENs to a program and it begins searching it instantly as soon as you have provided the input. You just have to automate this.

I don't think there is anything mysterious about any of this you just have a bunch creative new eyes approaching an old puzzle from different new perspectives.

At the end of the day you still have to search and make a decision, just like you are doing today.

As I said SF8 could be playing SF8 with this approach and lose against itself every time.

Regards