On-line engine blitz tourney February

[mvk]

Gaviota nicely beat Rookie in round 2.

[pgn][Event "HGM's monthly online engine blitz tourney"]
[Site "winboard.nl 5000"]
[Date "2015.02.14"]
[Time "21:21:00"]
[Round "2"]
[White "Gaviota"]
[Black "Rookie"]
[Result "1-0"]
[ECO "C78 Ruy Lopez, Archangelsk variation"]

1. e4 e5 2. Nf3 Nc6 3. Bb5 a6 4. Ba4 Nf6 5. O-O b5 6. Bb3 Bb7 7. Re1 Bc5 8. c3
Bb6 9. d4 O-O 10. Bg5 h6 11. Bh4 d6 12. a4 g5 13. Nxg5 hxg5 14. Bxg5 exd4 15.
cxd4 Nxd4 16. Bd5 c6 17. Ra3 cxd5 18. Rg3 Kh7 19. e5 dxe5 20. Rxe5 Rc8 21. Bd2
Nf5 22. Rxf5 Bxf2+ 23. Rxf2 Qb6 24. Be3 Rc5 25. Rxf6 Rc1 26. Qxc1 Qxf6 27. Qc2+
Qg6 28. Rxg6 fxg6 29. Qc7+ Kg8 30. Bd4 Rf7 31. Qd8+ Rf8 32. Qe7 Rf6 33. Qxf6 g5
34. Qg7# 1-0[/pgn]

Both programs left book after move 16, thinking Black was winning.

Rookie's first move out of book was 17. ... cxd5?? with evaluation +2, but in reality only 17. ... Re8 seems to hold the position so far. (I need to check that more deeply, because it still isn't great, but cxd5 outright loses and it is not hard to find White's moves for that). During the game, one move later evaluation was already -2. A classical horizon effect. That doesn't happen a lot with modern evaluations. Rookie needs 17 ply to avoid immediate disaster:

[D]r2q1rk1/1b3p2/pbpp1n2/1p1B2B1/P2nP3/R7/1P3PPP/1N1QR1K1 b - -

Code: Select all

  time ply  score variation
   0.0   1 +5.950 cxd5
   0.0   2 +5.871 cxd5, Rg3
   0.0   3 +5.561 cxd5, Rd3 Ne6
   0.0   4 +5.366 cxd5, Rg3 Kh7, Rh3+ Kg6
   0.0   5 +5.468 cxd5, Rg3 Kh7, a5 Bc5
   0.0   6 +5.136 cxd5, Rg3 Kh7, a5 Bc5, Rh3+ Kg6
   0.1   7 +3.707 cxd5, Rg3 Kh7, Rh3+ Kg7, Rh6 Qc7, Rxf6 dxe4
   0.1   8 +3.677 cxd5, Rg3 Kh7, Rh3+ Kg7, Rh6 Qd7, Rxf6 dxe4
   0.2   9 +3.387 cxd5, Rg3 Kh7, Rh3+ Kg7, Rh6 Qd7, Rxf6 dxe4, Nc3
   0.3  10 +3.346 cxd5, Rg3 Kh7, Rh3+ Kg7, Rh6 Qd7, Rxf6 dxe4, Nc3 Rae8, axb5 axb5
   0.6  11 +3.272 cxd5, Rg3 Kh7, Rh3+ Kg7, Rh6 Qc7, Rxf6 dxe4, Nc3 d5, Rh6
   3.0  12 +2.461 cxd5, Rg3 Kh7, Rh3+ Kg7, Rh6 dxe4, Bxf6+ Kxh6, Qd2+ Kh7, Qg5 Qxf6, Qxf6 Rad8
   5.6  13 +2.415 cxd5, Rg3 Kh7, Rh3+ Kg7, Rh6 dxe4, Bxf6+ Kxh6, Qd2+ Kh7, Qg5 Nf3+, gxf3 Rg8, Bxd8 Rxg5+, Bxg5 exf3
  10.9  14 +2.305 cxd5, Rg3 Kh7, Rh3+ Kg7, Rh6 dxe4, a5 Bc5, Bxf6+ Kxh6, Qd2+ Kh7, Bxd8 Raxd8, b4 Nb3
  24.3  15 +2.274 cxd5, Rg3 Kh7, Rh3+ Kg7, Rh6 dxe4, a5 Bc5, Bxf6+ Kxh6, Qd2+ Kh7, Bxd8 Raxd8, b4 Nb3, Qc3
 211.5  16 +0.228 cxd5, Rg3 Kh7, e5 dxe5, Rxe5 Rc8, Rh3+ Kg6, Rh6+ Kg7, axb5 Nxb5, Bxf6+ Qxf6, Qg4+ Qg6, Rxg6+ fxg6, Re7+ Rf7, Rxb7 Bxf2+, Kh1
 332.8  17 +0.165 Re8, Rg3 Kf8, Bh6+ Ke7, Bxf7 Rh8, Rg7 Rxh6, Ba2+ Kf8, Rf7+ Ke8, Rxb7 Ra7, Rxa7 Bxa7, e5 dxe5, Rxe5+ Kf8, axb5 axb5

The book endpoints are only verified up to 16 ply, and during the blitz game it didn't get that deep...

I wanted to feedback the game manually into the book learner to see if it would spot the problem and avoid the loss. Normally this feedback mechanism is automatic. But it relies on the server's PGN game mailing function to work, and this is still broken on HGM's server.

This is my wake-up call. Here are the relevant book snippets for some consecutive books. (My books are in csv format)

bookie-2014-10-29.csv (This book was used during PT47)

Code: Select all

r2q1rk1/1b3p2/pbpp1n2/1p1B2B1/P2nP3/R7/1P3PPP/1N1QR1K1 b - -,16,cxd5,0.787,33,34,0.0,1.054
r2q1rk1/1b3p2/pbpp1n2/1p1B2B1/P2nP3/R7/1P3PPP/1N1QR1K1 b - -,0,Re8,0.165,33,34,0.622,1.054
r2q1rk1/1b3p2/pbpp1n2/1p1B2B1/P2nP3/R7/1P3PPP/1N1QR1K1 b - -,0,Ne6,0.039,33,34,0.748,1.054

bookie-2014-12-22.csv (first updated book after PT47)

Code: Select all

r2q1rk1/1b3p2/pbpp1n2/1p1B2B1/P2nP3/R7/1P3PPP/1N1QR1K1 b - -,0,Re8,0.165,33,34,0.329,0.729
r2q1rk1/1b3p2/pbpp1n2/1p1B2B1/P2nP3/R7/1P3PPP/1N1QR1K1 b - -,16,Kg7,-0.587,33,34,1.081,0.729
r2q1rk1/1b3p2/pbpp1n2/1p1B2B1/P2nP3/R7/1P3PPP/1N1QR1K1 b - -,0,Ne6,-0.767,33,34,1.261,0.729
r2q1rk1/1b3p2/pbpp1n2/1p1B2B1/P2nP3/R7/1P3PPP/1N1QR1K1 b - -,0,cxd5,-3.891,33,34,4.385,0.729

Field 4 is the evaluation for the move in field 3. Field 1 is obviously the position. Field 2 is the analysis depth in plies, which is 0 for moves that stay in book. The other fields are ply depths and path errors, which are not so relevant for this post.

During this Saturday's online tournament, I was using a machine that still had the PT47 book on it. And in the background somewhere last year, my book learner, using drop-out expansion, had already studied cxd5 and concluded that it loses. So it had already put an alternative in the book, and this was ready in December...

Lesson learnt: book learning is great, but it doesn't help if you don't update your book on the machine you play tournaments with.

[JVMerlino]

I'd be interested in your results after more analysis. I wonder if this is going to end up a "best move" position or an "avoid move" position.

Myrddin at 4 cores gives up on cxd5 rather quickly, but settles on Ne6 and I didn't bother letting it run for much longer than you did in your post.

Code: Select all

 1   -493      0          180 c6d5 a4b5 d4b5 g5d8 
 2   -489      1          547 c6d5 a3g3 (34 KNPS)
 3   -487      1         5693 c6d5 a3g3 g8h8 g5f6 (355 KNPS)
 4   -471      1        11882 c6d5? a3g3 (742 KNPS)
 4   -455      3        16579 c6d5? a3g3 (534 KNPS)
 4   -423      3        24639 c6d5? a3g3 (794 KNPS)
 4   -453      3        31974 c6d5 a3g3 g8h7 a4a5 d5e4 g5f6 (1031 KNPS)
 5   -437      4        39136 c6d5? a3g3 (832 KNPS)
 5   -421      4        49603 c6d5? a3g3 (1055 KNPS)
 5   -420      4        55802 c6d5 a3g3 (1187 KNPS)
 6   -404      6        65015 c6d5? a3g3 (1031 KNPS)
 6   -410      6        90456 c6d5 a3g3 g8h7 g5f6 (1435 KNPS)
 7   -394      7       101965 c6d5? a3g3 (1307 KNPS)
 7   -378      9       142034 c6d5? a3g3 (1511 KNPS)
 7   -366     10       223707 c6d5 a3g3 g8h7 g3h3 (2052 KNPS)
 8   -350     14       285940 c6d5? a3g3 (2027 KNPS)
 8   -334     25       583360 c6d5? a3g3 (2333 KNPS)
 8   -317     34       854144 c6d5 a3g3 g8h7 g3h3 h7g6 h3g3 a8c8 g5e3 g6h7 e3d4 d5e4 d4b6 d8b6 a4b5 a6b5 (2490 KNPS)
 9   -301     46      1266325 c6d5? a3g3 (2705 KNPS)
 9   -285     60      1746186 c6d5? a3g3 (2867 KNPS)
 9   -253     84      2444514 c6d5? a3g3 (2899 KNPS)
 9   -205     99      2893700 c6d5? a3g3 (2896 KNPS)
 9   -141    127      3665996 c6d5? a3g3 (2866 KNPS)
 9    -61    179      5059926 c6d5? a3g3 (2820 KNPS)
 9    -82    196      5549756 d4e6 d5e6 f7e6 g5f6 f8f6 a3g3 (2822 KNPS)
10    -66    240      6748568 d4e6? d5e6 (2808 KNPS)
10    -65    333      9214766 d4e6 d5e6 f7e6 a3f3 e6e5 b1c3 b6d4 d1b3 d6d5 e1f1 a8c8 a4b5 a6b5 e4d5 c6d5 b3b5 (2759 KNPS)
11    -63    407     11256878 d4e6 d5e6 f7e6 a3f3 b5a4 d1a4 d8e7 b1c3 b6c5 h2h3 d6d5 (2764 KNPS)
12    -54    603     16483988 d4e6 d5e6 f7e6 a3f3 d8e7 b1c3 b5b4 d1b3 b6c5 g5f6 f8f6 f3g3 g8h8 (2730 KNPS)
13    -54    998     26320568 d4e6 d5e6 (2636 KNPS)
14    -39   2098     54567365 d4e6 d5e6 f7e6 a3f3 b5a4 d1a4 d8e7 a4b3 b6d4 h2h3 e7d7 g2g3 d6d5 g5f6 f8f6 (2600 KNPS)
15    -39   4293    109802928 d4e6 d5e6 (2557 KNPS)
16    -23   9662    242596057 d4e6? d5e6 (2510 KNPS)
16    -20  55051   1341807558 d4e6 d5e6 f7e6 a3f3 f8f7 d1d2 a8b8 b1c3 b5b4 c3d1 c6c5 h2h3 e6e5 g5f6 f7f6 d2g5 g8f7 f3f6 (2437 KNPS)

jm

[Joost Buijs]

That position looks very suspicious for black, it just means the engines have no clue.
I don't believe in book learning at all, the best books are tuned manually, the only problem with this is that there are not many people with enough time on their hands to accomplish this.

[mvk]

That position looks very suspicious for black, it just means the engines have no clue.
I don't believe in book learning at all, the best books are tuned manually, the only problem with this is that there are not many people with enough time on their hands to accomplish this.

The refutation of cxd5 also removes the whole line leading upto the position from the repertoire, because the path error for Black becomes too large. Without the path errors, learning is hopeless. With them, the learning saves me time.

[mvk]

I'd be interested in your results after more analysis. I wonder if this is going to end up a "best move" position or an "avoid move" position.

Myrddin at 4 cores gives up on cxd5 rather quickly, but settles on Ne6 and I didn't bother letting it run for much longer than you did in your post.

Thank you. Ne6 is also good.

Out of curiosity, I did a 2,246 node drop-out for this position. Ne6 and Re8 stay close to each other, Ne6 slightly better. Visualisation of the mainlines here. All other moves lose right away.

I must note that the removal of cxd5 already kicks the entire line leading to this position out of book, due to high path errors. I don't have to worry, and I didn't analyze it further than this. I will keep it running for a few days more and post back if there is a change of evaluation.

[JVMerlino]

I'd be interested in your results after more analysis. I wonder if this is going to end up a "best move" position or an "avoid move" position.

Myrddin at 4 cores gives up on cxd5 rather quickly, but settles on Ne6 and I didn't bother letting it run for much longer than you did in your post.

Thank you. Ne6 is also good.

Out of curiosity, I did a 2,246 node drop-out for this position. Ne6 and Re8 stay close to each other, Ne6 slightly better. Visualisation of the mainlines here. All other moves lose right away.

I must note that the removal of cxd5 already kicks the entire line leading to this position out of book, due to high path errors. I don't have to worry, and I didn't analyze it further than this. I will keep it running for a few days more and post back if there is a change of evaluation.

Thanks very much. However, it seems like Ne6 might be lost anyway.

After 1...Ne6 2.Bxe6 fxe6 3.Rg3, there are two reasonable replies: 3...Bxf2+ and 3...Kf7.

If 3...Bxf2+ 4.Kxf2 Nxe4+ 5.Kg1 Nxg5 6.Rxe6 Rf5 7.h4, then White is down a piece but has a strong attack against an unprotected king. I get scores close to +1.00 with this position:

[D]r2q2k1/1b6/p1ppR3/1p3rn1/P6P/6R1/1P4P1/1N1Q2K1 b - h3 0 7

if 3...Kf7 4.Bxf6 Qxf6 5.Rf3 Ke7 6.Rxf6 Rxf6, then these scores are a bit closer to draw. I'll leave somebody else to give more thorough analysis on this position if anybody cares:

[D]r7/1b2k3/pbpppr2/1p6/P3P3/8/1P3PPP/1N1QR1K1 w - - 0 6

jm

[mvk]

The second line is my mainline also. But I don't know if I can trust my program with this unbalance.

[JVMerlino]

The second line is my mainline also. But I don't know if I can trust my program with this unbalance.

I decided to let Stockfish take over. After three hours, at depth 35, it gives a score of +0.53:

[D]r7/1b2k3/pbpppr2/1p6/P3P3/8/1P3PPP/1N1QR1K1 w - - 0 6

info depth 35 seldepth 71 score cp 53 nodes 13701923466 nps 898315 time 15252914
multipv 1 pv g1h1 b6f2 e1e2 a8h8 h2h3 e6e5 a4b5 a6b5 h1h2 h8g8 d1b3 f2g3 h2g1 f6g6 b1d2 g8a8 b3d1 g3f4 d2f1 c6c5 f1e3 b7e4 e3d5 e4d5 d1d5 a8a1 g1f2 f4g3 f2f3 a1f1 f3e3 b5b4 d5b7 e7f6 b7d7 f6g5 d7d8

However, the score didn't increase much more than 20cp from depth 20. So this position is way too deep even for the best engines.

jm

[Joost Buijs]

Indeed, I am quite happy with the attendance, and especially that not all participants are 2600+ engines, but that they populate the entire range from 1800 Elo up. (Of course NEG is about 0 Elo; I just joined it to avoide a bye.)

Will there be a tournament this weekend (21/3)?
If so, I have to make some arrangements for it.

I didn't touch the source of my current engine for at least 5 months but it is always fun to play the monthly blitz.

[hgm]

Yes, it will. I was just going to announce it.