2018-09-03

Leela Chess Zero - difficult positions. The new world of NN engines!



 Traditional Chess engines with classic Alpha-Beta search, sometimes have huge problems with some positions that require deep planning and understanding of the position. This is because they try to search all possible moves in a position and then their replies and then all the replies of the replies, etc. They discard most moves of course with clever algorithms that prune almost all moves but a few that they focus on. With a cost of course to discard a move, not good looking at first sight, but that after many moves ahead it proves to be a very good one. Traditional engines also have a handcrafted evaluation function. That means in order to judge a position they have to rely on human rules for it, e.g bonus for Rook on open file, bonus if a Pawn is on 6th rank etc, etc.

 Neural net engines with latest advances Deepmind brought with deep learning techniques used to train convolutional neutral networks, like Leela, on the other hand use a much more powerful evaluation function. One that acquired Chess knowledge about every single aspect of Chess by playing against itself millions of games. No need to manually put hundreds instructions like "if it's a closed position then Knight value increases and Rook's decreases" etc. You let them play against itself and they create by themselves millions of such "rules". Leela also uses a MCTS(Monte Carlo Tree Search) type of search that is completely different than traditional AlphaBeta search.



So it's no wonder that Leela plays very frequently now, moves that other top engines like Stockfish, Komodo, Houdini etc, can't find.
In some cases like these that would be shown here, it even plays moves that not a single one traditional engine can find, or that they need very long searching to do!
Of course that doesn't mean Leela finds most deep testpositions because this is not true. Leela is still not that good anyway in testpositions that require tactics. But when long strategical thinking is required without so much immediate tactics, then Leela usually shines!

Test position 1.

White to play.

White has created a nice concentration of its pieces in the Kingside but has many problems to solve, especially with Bxa3 pending that will destroy white's position. Defensive resources of black are big and King's position seems very solid. The c2 Knight is very annoying too for white. Most traditional engines give here black slightly better or even a lot better or even winning!
And want to play moves like Rh6, Bf2, Nd2. Moves that lead nowhere for white and probably are losing.







 But if you let Leela analyze the position(since Leela Chess Zero is just the project name, by saying LCZero, we're referring to the body of her the Lc0 binary that takes the neural net and provides the search, plus the brain of her, and that is the nets/weights, and here i tried about 9 nets and all find the same move) you will see her after only something less than 1 second of thought proposing Qf6+ sacrificing the Queen with a big plus score for white!!
And not only this but she is finding the correct continuation of Ng5 later on that white has to play!  h5 Bxh5 that she is playing are the correct move for black also(the toughest defense) and the best reply by white.


Analysis by Lc0v17 11089:

 4/8    00:00     222    801    +0,49    Be1-d2 Rf8-h8 Rh3-h6 Nc2xa3 h2-h4 Na3-c2
 4/9    00:00     247    826    +1,55    Qf4-f6+ Be7xf6 g5xf6+ Kg7-g8 Nf3-g5 Nc2xd4 c3xd4 h7-h5 Be2xh5
 5/10    00:00     299    920    +1,70    Qf4-f6+ Be7xf6 g5xf6+ Kg7-g8 Nf3-g5 Nc2xd4 c3xd4 c4-c3 Rh3xh7 Bc6-b5
 5/11    00:00     353    1,002    +1,72    Qf4-f6+ Be7xf6 g5xf6+ Kg7-g8 Nf3-g5 Nc2xd4 c3xd4 c4-c3 Rh3xh7 Bc6-b5
 6/12    00:00     431    1,110    +1,42    Qf4-f6+ Be7xf6 g5xf6+ Kg7-g8 Nf3-g5 Nc2xd4 c3xd4 c4-c3 Rh3xh7 Bc6-b5
 9/16    00:01     2,794    2,457    +1,96    Qf4-f6+ Be7xf6 g5xf6+ Kg7-g8 Nf3-g5 Nc2-e3 Rh3xh7 Rb8-b7 Be1-d2 Ne3xd1
 9/17    00:01     5,697    2,880    +2,09    Qf4-f6+ Be7xf6 g5xf6+ Kg7-g8 Nf3-g5 h7-h5 Be2xh5 Nc2xa3 Bh5-g4 Na3-c2
 10/18    00:02     5,881    2,885    +2,08    Qf4-f6+ Be7xf6 g5xf6+ Kg7-g8 Nf3-g5 h7-h5 Be2xh5 Nc2xa3 Bh5-g4 Na3-c2
 10/19    00:02     7,656    3,098    +2,02    Qf4-f6+ Be7xf6 g5xf6+ Kg7-g8 Nf3-g5 h7-h6 Rh3xh6 Nc2-e3 Rh6-h7 Rb8-b7
 10/20    00:02     9,007    3,236    +2,04    Qf4-f6+ Be7xf6 g5xf6+ Kg7-g8 Nf3-g5 h7-h6 Rh3xh6 Nc2-e3 Rh6-h7 Rb8-b7
 11/20    00:03     10,230    3,333    +2,02    Qf4-f6+ Be7xf6 g5xf6+ Kg7-g8 Nf3-g5 h7-h6 Rh3xh6 Nc2-e3 Rh6-h7 Rb8-b7
 13/28    00:12     48,019    3,804    +1,52    Qf4-f6+ Be7xf6 g5xf6+ Kg7-g8 Nf3-g5 h7-h5 Be2xh5 Nc2xa3 Bh5-g4 Na3-c2
 13/29    00:12     48,392    3,809    +1,52    Qf4-f6+ Be7xf6 g5xf6+ Kg7-g8 Nf3-g5 h7-h5 Be2xh5 Nc2xa3 Bh5-g4 Na3-c2
 14/29    00:13     51,523    3,806    +1,52    Qf4-f6+ Be7xf6 g5xf6+ Kg7-g8 Nf3-g5 h7-h5 Be2xh5 Nc2xa3 Bh5-g4 Na3-c2
 14/30    00:16     63,490    3,810    +1,56    Qf4-f6+ Be7xf6 g5xf6+ Kg7-g8 Nf3-g5 h7-h5 Be2xh5 Nc2xa3 Bh5-g4 Na3-c2
 14/31    00:17     65,141    3,809    +1,54    Qf4-f6+ Be7xf6 g5xf6+ Kg7-g8 Nf3-g5 h7-h5 Be2xh5 Nc2xa3 Bh5-g4 Na3-c2
 14/31    00:18     68,414    3,794    +1,55    Qf4-f6+ Be7xf6 g5xf6+ Kg7-g8 Nf3-g5 h7-h5 Be2xh5 Nc2xa3 Bh5-g4 Na3-c2


 Note that the analysis of this position and for all positions by Leela should be done by not providing the simple FEN or EPD and let it analyze, but by providing a history of moves that ended in that position. And this is because of the way Leela thinks, she thinks of position sequences from the start of the game. But practically it is proven that you do not need to provide Leela with a full history from starting position to the desired one, but only 2-3 moves. By making her analyze from a FEN, you are IN A HUGE DANGER of obtaining an analysis that is completely meaningless.
So for example to obtain the above analysis for the position, we gave her the PGN:

[Event "?"]
[Site "?"]
[Date "????.??.??"]
[Round "?"]
[White "New game"]
[Black "?"]
[Result "*"]
[SetUp "1"]
[FEN "1r3r2/4bpkp/1qb1p1p1/3pP1P1/p1pP4/PpP2NQR/1P2B2P/n2RB2K w - - 0 1"]
[PlyCount "2"]

1. Qf4 Nc2 *

And this is the correct move in this position, white to give its Queen! And it wins the game!

[Event "?"] [Site "?"] [Date "2018.??.??"] [Round "?"] [White "?"] [Black "?"] [Result "1-0"] [SetUp "1"] [FEN "1r3r2/4bpkp/1qb1p1p1/3pP1P1/p1pP1Q2/PpP2N1R/1Pn1B2P/3RB2K w - - 0 22"] [PlyCount "5"] [EventDate "2018.??.??"] [EventType "rapid"] 22. Qf6+ $3 {White wins with this.} Bxf6 23. gxf6+ Kh8 24. Ng5 $1 {And this is 1-0 but there are countless variations with many difficult key moves for white to find.} 1-0


Test position 2.

White to play.

This is a position that results in one of the variations of the previous position 1. White is a Queen down. Yet it is winning and has a forced checkmate with Rg1!!.
Stockfish the best engine as of now, make about 1.6 billion nodes in order to find the winning move. Most other traditional engines have also big problems finding this as they require to calculate many many nodes.

Leela again finds it immediately with a winning score!








Test position 3.

White to play.

Here we have a classic anticomputer position where traditional engines have many problems to see that the only way for white to win is by playing Qxb5!
A human can see it easily. Engines have huge problems in this and want to save their Queen by retreating.
But not Leela, as it plays Qxb5 instantly!










E.g analysis with Lc0v17 11089 net:
 4/7    00:00     1,777    2,307    +16,62    Qa4-a2 Ke7-d8 Qa2-a1 Nf6-h7 Rb2-a2 Nh7-f8
 4/7    00:00     2,289    2,580    +19,83    Qa4xb5 a6xb5 a5-a6 Ke7-d8 a6-a7 Ne8-c7
 4/9    00:01     5,875    3,615    +19,54    Qa4xb5 a6xb5 a5-a6 Ke7-d8 a6-a7 Ne8-c7
 5/9    00:02     9,718    4,197    +19,48    Qa4xb5 a6xb5 a5-a6 Ke7-d7 a6-a7 Ne8-c7 Rb2-a2
 5/10    00:04     23,863    5,387    +19,32    Qa4xb5 a6xb5 a5-a6 Ke7-d8 a6-a7 Ne8-c7 Rb2-a2 Nf6-d7 Ra2-a6
 5/11    00:08     47,496    5,781    +19,26    Qa4xb5 a6xb5 a5-a6 Ne8-c7 Rb2-a2 Ke7-d7 a6-a7 Kd7-e8 Ra2-a6
 5/12    00:09     53,412    5,836    +19,27    Qa4xb5 a6xb5 a5-a6 Ne8-c7 a6-a7 Ke7-d7 Rb2-a2 Kd7-e8 Ra2-a6
 6/12    00:10     59,317    5,878    +19,22    Qa4xb5 a6xb5 a5-a6 Ne8-c7 Rb2-a2 Ke7-d7 a6-a7 Kd7-e8 Ra2-a6
 6/12    00:20     126,846    6,306    +18,67    Qa4xb5 a6xb5 a5-a6 Ne8-c7 Rb2-a2 Nf6-d7 a6-a7 Nd7-b6 Ra2-a5 Nb6-a8
 6/12    00:25     163,418    6,499    +18,67    Qa4xb5 a6xb5 a5-a6 Ne8-c7 a6-a7 Ke7-d7 Rb2-a2 Kd7-e8 Ra2-a6 Nf6-d7
 6/13    00:25     164,964    6,504    +18,66    Qa4xb5 a6xb5 a5-a6 Ne8-c7 Rb2-a2 Nf6-d7 a6-a7 Nd7-b6 Ra2-a5 Nb6-a8
 6/13    00:30     202,289    6,655    +18,68    Qa4xb5 a6xb5 a5-a6 Ne8-c7 Rb2-a2 Ke7-d7 a6-a7 Kd7-e8 Ra2-a6 Nf6-d7 Ra6-h6
 6/13    00:35     241,472    6,817    +18,67    Qa4xb5 a6xb5 a5-a6 Ne8-c7 Rb2-a2 Ke7-d8 a6-a7 Nf6-d7 Ra2-a6 Nc7xa6 a7-a8Q+
 7/13    00:52     388,231    7,353    +18,62    Qa4xb5 a6xb5 a5-a6 Ne8-c7 Rb2-a2 Ke7-d7 a6-a7 Kd7-e8 Ra2-a6 Nf6-d7 Ra6-h6
 7/13    01:17     622,539    7,995    +18,40    Qa4xb5 a6xb5 a5-a6 Ne8-c7 Rb2-a2 Ke7-d7 a6-a7 Kd7-e8 Ra2-a6 Nf6-d7 Ra6-h6


Test position 4.


Black to play.




Here is a position from a real game where Leela as black has played g5!! A very deep positional move, trying to trap the Bishop.
Most engines don't seem to find it in logical time.








[Event "CCCC 1: Rapid Rumble (15|5) Stage 1"] [Site "Chess.com"] [Date "2018.09.02"] [Round "?"] [White "Vajolet 2.6"] [Black "Lc0 17.11089"] [Result "0-1"] [ECO "B92"] [WhiteElo "3101"] [BlackElo "3226"] [SetUp "1"] [FEN "rn1qr1k1/1p2bp1p/3p2pB/p2Ppb2/P7/1N6/1PPQBPPP/R2R2K1 b - - 0 15"] [PlyCount "1"] [EventDate "2018.??.??"] [EventType "rapid"] [TimeControl "900+5"] 15... g5 $3 {[%eval -8,35] [%emt 0:00:00][%csl Rh6] Leela played this very fast! Creating pressure on opponent's Kingside but mainly trying to entrap the Bishop. Finally it succeeded in the game and won.} 0-1



Test position 5.

White to play.



In this again anticomputer position white has to immediately play Bxb5!! in order to win.
Most engines need a very long search in order to find it if at all since many are stuck in capturing the Knight that only draws after Rxe2! by black.
Not Leela, since after just some seconds finds Bxb5 with a winning score.







Analysis by Lc0v17 11089:
 4/9    00:00     3,892    3,967    +16,89    Rf1xe1 Rd2xe2 Re1xe2 Kd8-e8 Rf3-c3 Ke8-d8
 4/9    00:01     4,148    3,996    +17,21    Bf2xe1 Rd2xe2 Rf3-g3 Re2xe5 Rg3-g2 Re5-e3+ Kh3-g4
 4/9    00:01     4,404    4,021    +16,64    Rf1xe1 Rd2xe2 Re1xe2 Kd8-e8 Rf3-c3 Ke8-d8
 5/9    00:01     4,660    4,045    +16,78    Rf1xe1 Rd2xe2 Re1xe2 Kd8-e8 Rf3-c3 Ke8-d8
 5/10    00:01     5,428    4,130    +16,28    Rf1xe1 Rd2xe2 Re1xe2 Kd8-e8 Rf3-c3 Ke8-d8
 5/10    00:01     7,988    4,327    +15,89    Bf2xe1 Rd2xe2 Rf3-g3 Re2xe5 Rg3-g2 Re5-e3+ Kh3-g4 Re3-e5
 5/10    00:01     8,244    4,341    +15,39    Rf1xe1 Rd2xe2 Re1xe2 Kd8-c8 Kh3-g4 Kc8-d8 Kg4-f4 Kd8-e8
 5/10    00:02     8,756    4,375    +15,50    Bf2xe1 Rd2xe2 Rf3-g3 Re2xe5 Rg3-g2 Re5-e3+ Kh3-g4 Re3-e5
 5/10    00:02     9,012    4,387    +15,56    Rf1xe1 Rd2xe2 Re1xe2 Kd8-c8 Kh3-g4 Kc8-d8 Kg4-f4 Kd8-e8 Re2-e3
 5/10    00:02     10,804    4,473    +14,98    Bf2xe1 Rd2xe2 Rf3-g3 Re2xe5 Rg3-g2 Re5-e3+ Kh3-g4 Re3-e5
 5/11    00:03     15,413    4,591    +12,90    Bf2xe1 Rd2xe2 Rf3-d3 Re2xe5 Kh3-g4 Re5-e2 Kg4-f4 Re2-h2
 5/11    00:03     17,100    4,604    +13,68    Be2xb5 Ne1xf3 Bb5xa6 Rd2-b2 Bf2-e3 Nf3-d2 Be3xd2 Rb2-a2 Ba6-c4
 5/12    00:04     19,921    4,623    +13,23    Be2xb5 Ne1xf3 Bb5xa6 Nf3-h2 Kh3xh2 Rd2-b2 Ba6-d3 Rb2xb4 a5-a6
 5/13    00:04     22,483    4,657    +13,02    Be2xb5 Ne1xf3 Bb5xa6 Rd2-b2 Bf2-e3 Nf3xe5 Be3-d4 Rb2-b3+ Kh3-g2
 6/13    00:04     22,998    4,658    +13,00    Be2xb5 Ne1xf3 Bb5xa6 Rd2-b2 Bf2-e3 Nf3xe5 Be3-d4 Rb2-b3+ Kh3-g2
 6/13    00:09     47,568    4,781    +14,10    Be2xb5 Ne1xf3 Bb5xa6 Rd2-a2 Bf2-e3 Nf3-d4 Be3xd4 Ra2-a4 Ba6-d3
 6/14    00:10     49,873    4,792    +14,14    Be2xb5 Ne1xf3 Bb5xa6 Rd2-a2 Bf2-e3 Nf3-d4 Be3xd4 Ra2-a4 Ba6-d3
 7/14    00:11     55,507    4,825    +14,27    Be2xb5 Ne1xf3 Bb5xa6 Rd2-b2 Bf2-e3 Nf3-d2 Be3xd2 Rb2xd2 Ba6-c4
 7/15    00:16     80,453    4,912    +13,35    Be2xb5 a6xb5 Rf1xe1 Rd2-d3 Rf3xd3 Kd8-c8 a5-a6 Kc8-b8 a6-a7+ Kb8xa7
 7/16    00:17     87,629    4,972    +13,25    Be2xb5 a6xb5 Rf1xe1 Rd2-d3 Rf3xd3 Kd8-c8 a5-a6 Kc8-b8 a6-a7+ Kb8xa7
 8/16    00:32     179,923    5,544    +13,23    Be2xb5 a6xb5 Rf1xe1 Rd2-d3 Rf3xd3 Kd8-c8 a5-a6 Kc8-b8 a6-a7+
 9/17    00:42     239,377    5,571    +13,04    Be2xb5 a6xb5 Rf1xe1 Rd2-d3 Rf3xd3 Kd8-c8 a5-a6 Kc8-b8 a6-a7+
 9/18    00:52     298,068    5,699    +13,26    Be2xb5 a6xb5 Rf1xe1 Rd2-d3 Rf3xd3 Kd8-c8 Rd3-d5 Kc8-b7 a5-a6+
 9/18    00:55     318,816    5,757    +13,18    Be2xb5 a6xb5 Rf1xe1 Rd2xf2 Rf3xf2 Kd8-c8 a5-a6 Kc8-b8 a6-a7+


 Yet, in the most difficult types of positions for computers in Chess, the fortress positions, Leela doesn't fare well at all generally despite the fact that in the aforementioned 2 positions it did perfectly. But generally especially in draw fortress situations she does not have too much success. Which is strange for a MCTS type search since one would expect an engine with random rollouts like Monte Carlo does, to be able to see through its statistics that one move in a fortress position holds for example the draw. But this does not occur here since Leela does not do random rollouts but uses the neural net to guide the search.
So in order to be able to solve fortress positions also, she just need more training games.







16 comments:

  1. Leela is a chess revolution. I am really impressed with games of this program. Leela already plays like super human. The way as she plays Kings Indian and other closed position is really impresive for me. Keep like that and make Leela no.1 chess engine in the world - great job!

    ReplyDelete
  2. I put the test position 1 into arena with Leela (ID 11260) and SF to evaluate, but did not enter the last move (Na3c2). I had Leela play white and SF play black. Leela got 5 minutes per move, and SF waited until a depth of 25. This was on a i7 8700k and a Vega 56 GPU, with 16gb ram, just for clarification. GAME: 1. Qf4 Nc2 2. Qf6+ Bxf6 3. gxf6+ Kg8 4. Ng5 h5 5. Bxh5 Nxa3 6. Bg4 Rb7 7.Rh7 Nb1 8. Bf2 Rfb8 9. Rxb1 a3 10. Be3 axb2 11. Rg7+ Kf8 12. Rh7 Ke8 13.Bxe6 fxe6 14. f7+ Ke7 15. f8=Q+ Kxf8 16. Rh8+ Kg7 17. Rh7+ Kg8 18. Rh6 Qa5
    19. Rxg6+ Rg7 (Here Leela gave a +20 eval.) 20. Rxg7+ Kxg7 21. Rg1 Qa1 22. Ne4+ Qxg1+ 23. Kxg1 b1=Q+ 24.
    Kf2 Qxe4 25. Bg5 b2 26. Be7 b1=Q 27. Kg3 Qee1+ 28. Kg4 Qf5# *
    Either way, it was an incredible find by Leela, and very well done. I look forward to seeing her improve. Thank you for all your hard work.

    P.S. This was not meant to take away from the genius of Qf6+!, but to show, that black was still better, and she over evaluated the position. I must suggest when looking at finds like this, that we check with other engines, to make sure that they do not have some refutation to the line - that perhaps was the reason they do not give that line in their evaluation. Otherwise, keep up the great work!

    ReplyDelete
    Replies
    1. Leela played correctly till move 6. Since 6.Bg4 draws(so the loss is because Leela misplayed some other move) loses, while 6.Bd2!! wins.

      Delete
    2. I tried your variation, but my leela already sees the position after 12 ... Ke8 as -1.90,
      at move 19, Leela thinks black is winning. (11198)

      Interstingly, it looks like the key is 5. ... Nxa3.
      SF found it after d25, Leela included it as its main variation when it was considering Bxh5 as early as d5.

      On another note, the g5 move against Vajolet is amazing!
      I tried it with SF and Leela. Leela on a weak laptop found it very quickly.

      Delete
  3. With the same specs as before, and the move 6.Bd2 subbed in, with Leela as white and SF as black - here is the game: 1. Qf4 Nc2 2. Qf6+ Bxf6 3. gxf6+ Kg8 4. Ng5 h5 5. Bxh5 Nxa3 6. Bd2 Nb5 7. Rg1 Qa7 8. Nxf7 Qxf7 9. Rxg6+ Qxg6 10.Bxg6 Rb7 11. Rg3 Rh7 12. Bc1 Kh8 13. Bxh7 Kxh7 14. Rh3+ Kg6 15. Rg3+ Kf716. Rg7+ Ke8 17. h4 a3 18. bxa3 Rf7 19. Rg8+ Kd7 20. Bb2 Nc7 21. h5 Ne8 22.h6 Nxf6 23. exf6 Rxf6 24. Rg7+ Kd6 25. h7 Rh6+ 26. Kg2 Ba4 27. Rf7 Rh5 28.Kg3 Rh1 29. Kg2 Rh6 30. Bc1 Rh5 31. Rg7 b2 32. Bxb2 Bc2 33. a4 Bxa4 34.Ba3+ Kc6 35. Re7 Bc2 36. Rxe6+ Kb5 37. Kg3 Ka4 38. Bb4 Bxh7 39. Rb6 Kb3 40.Kg4 Rh1 41. Kg5 Bb1 42. Kf4 Be4 43. Ke5 Rh3 44. Rb5 Rxc3 45. Ba5+ Kc2 46.Rxd5 Re3 47. Rc5 Kd3 48. d5 Bg2+ 49. Kd6 Re8 50. Kd7 Re5 51. Kc6 Rh5 52.Be1 Rg5 53. Bb4 Kd4 54. Kb5 Bxd5 55. Rc7 Rf5 56. Ba5 Bc6+ 57. Kb6 Rb5+ 58.Ka6 Kc5 59. Rg7 Rb3 60. Rg5+ Bd5 61. Bd8 Ra3+ 62. Ba5 c3 63. Rg3 c2 64.Rxa3 c1=Q 65. Rc3+ Bc4+ 66. Kb7 Qf4 67. Bc7 Qe4+ 68. Ka7 Kc6 69. Kb8 Qb1+ 70. Kc8 Qb7+ 71. Kd8 Qxc7+ 72. Ke8 Qd7+ 73. Kf8 Qf7# While the move 6.Bd2 is a draw (even though you claimed it was winning, which it wasn't) - Leela lost, because of a blunder - 62. Ba5? This move lost her the game. However, that is not the point, both of claimed Qf6+ to be winning, when she could not win against SF following that move. 6.Bd2 was draw, not a win as you claimed it to be.
    All I am saying is, when we look at these incredible lines that Leela shows, that we also check with other engines that there isn't something that she may have missed. I believe there has to be a reason that Komodo, Houdini and SF all didn't like Qf6+. Never the less, it was still a fantastic find, and keep up the good work in developing Leela

    ReplyDelete
    Replies
    1. This comment has been removed by the author.

      Delete
    2. >While the move 6.Bd2 is a draw (even though you claimed it was winning, which it wasn't

      No. Leela didn't play perfectly and i don't know why you assume she plays perfectly.
      Instead of 8.Nxf7 try 8.Rg4 and you will see the light :)

      Delete
    3. This comment has been removed by the author.

      Delete
    4. You are missing the point. Leela still could not convert what was claimed by the OP to be winning, to a win. She also could not convert your line to a win. I do not expect perfect play, but I also expect that she be able to convert the line that she claims is winning. As for Rg4, you are right, it leads to mate, but she still was unable to find the mate with it. The point of my comments have been to say, that all though Leela has found impeccable ideas, she is still unable to convert and therefore we should still use one of the top 3 engines to check her lines. SF found mate in 13 after Rg4, however, Leela played poorly, missing many chances. You are giving moves she did not find, so her seeing Qf6+, was non-reliant on the genius, winning line you gave. I don't expect even SF to play perfectly, however I do believe it is important that we clarify whether or not the line -she gave-! was winning, and not a line that may follow but that she did not see. That is all I am trying to say. Even if she finds these moves, she is still not converting them. Otherwise you have made very good points, and Leela has developed phenomenally! Looking forward to the future, have a good day!

      Delete
    5. Well... it would be interesting to see if Leela had performed better against SF with a NVidea Card like 1060 TI because it is optimized for those cards. A Radeon Vega may not be the best choice against SF in this case.

      Delete
    6. The point was not Leela to win from these positions or from the first one specifically. The point is that Leela solves it! It's brain(neural net) is that clever that somehow sees the Qf6+ as a viable move worth trying, while other engines cannot see.
      Of course you can't expect that Leela or any other engine will be able to predict and find all the correct continuations in the sea of variations in this incredibly complex position.

      Furthermore about your "there has to be a reason that Komodo, Houdini and SF all didn't like Qf6+.", the reason is not that the position is not correct and a white win as you assumed, because it is, white wins, but the reason is that Komodo, Houdini and SF are not adequate enough to see this brilliant move!

      Delete
  4. This comment has been removed by the author.

    ReplyDelete
  5. Given the comment re not using FENS for analysis, can you please provide a list of PGNs for these positions, with the 2 or 3 preceding moves. I'd love to give a few of them a try with LC0, McBrain, Komodo and others! Also what is considered to be the best/most useful net from before the recent reset? Thanks!

    ReplyDelete
  6. How do you get the second position from the first one?

    ReplyDelete
  7. This comment has been removed by the author.

    ReplyDelete
  8. I did some analysis with help of an engine on the test position.

    The first few moves are pretty straightforward even for a human.
    22. Qf6+ Bxf6 23. gxf6+ Kg8 24. Ng5 h5 25. Bxh5 Nxa3

    First critical position
    26. Bf2 and Bd2 make most sense, because the rook needs space in any case.
    Alternative: 26. Bg4 Nb5 27. Rh7 (27. Nxe6 Nc7) 27. .. a3 draws


    26. Bf2 Nb5 27. Rg1 Qc7
    Qc7 seems to me the most critical defense. It gives black some nice defensive resources.

    28. Rg4 Nd6 29. exd6 (29. Nxf7 Qxf7) 29. .. Qd8 30. Nxf7 Qxf6
    28. Bf3 Nd6 29. exd6 Qd7 30. Rg7 (30. Rg4 e5) 29. .. Qd8
    28. Bg4 a3 29. Rh7 Qxe5

    These lines are drawing. The only way to win seems to be:
    28. Bd1 Nd6 29. exd6 Qd8 30. Bg3 Qxf6
    (30. .. a3 31. Be5 axb2 32. Bh5 wins)
    31. Be5 Qxe5 32. dxe5 a3 33. Bf3.
    Now I think a2 is black's best attempt and winning this is far from obvious.
    White can play Nh7 and trade knight for rook or Rh4, but Ra8 forces Ra1 in any case.

    26. Bd2 Nb5 27. Rg1 Qc7 28. Rg4 Nd6 29. Nxf7 Qxf7 30. Bxg6 Qxg6 31. Rxg6+ Kf7 32. Rg7+ Ke8 33. exd6
    Rb7 34. Re7+ Kd8 35. Bh6 a3 36. Rg3 axb2 37. Rg1 b1=Q 38. Rxb1 Rf7 39. Bg5
    Kc8 40. Kg2 Rfxe7 41. dxe7 Be8 isn't a clear win either.

    So I am not convinced that the test position is winning for white.
    Maybe all engines don't understand the position.

    ReplyDelete