This position appears to be very hard for engines to resolve

[zullil]

2.5 hours into the analysis and Stockfish came up with this for me:
36 +0.12 2:25:27 89001M a4 b4 Ne2 Nbd7 O-O c5 c3 d5 Bxf6 Nxf6 e5 bxc3 Nxc3 Nd7 Nxd5 Bxf3 Rxf3 Nb6 Nc3 Qxd4+ Qf2 Rad8 a5 Qxf2+ Kxf2 Nd5 Be2 Nxc3 bxc3 f6 exf6 Bxf6 Rb1 Rb8 Rxb8 Rxb8 g4 e5 Bc4+ Kh8 Ke3 exf4+ Kxf4 Rb2

Not really sure what you are looking for---an evaluation score that is more favorable for White than what you are seeing? As for the move, there seem to be many reasonable choices, and also opportunities for transpositions. In any case, here's my last offering:

Code: Select all

info depth 43 seldepth 64 multipv 1 score cp 20 nodes 816855450547 nps 22339537 hashfull 999 tbhits 0 time 36565460 pv f4f5 b8d7 e1g1 a8c8 g1h1 b5b4 c3e2 c6c5 c2c3 b4c3 b2c3 d8a5 h2h3 g4f3 g2f3 c5c4 d3c2 c8b8 h3h4 b8b2 h4h5 h7h6 g5f6 d7f6 h5g6 f7g6 d2c1 b2b6 f1g1 g6g5 a1b1 b6b1 c1b1 a5c7 b1c1 f8b8 h1g2 c7b7 a2a4 b7b2 g2f2 d6d5 e4e5 b2c1 g1c1 f6d7

[Eelco de Groot]

rn1q1rk1/p3ppbp/2pp1np1/1p4B1/3PPPb1/2NB1N2/PPPQ2PP/R3K2R w KQ -

Engine: Rainbow Serpent Sf20150403 MOD MP (512 MB)
by Tord Romstad, Marco Costalba and Joona Kiiski

31 60:08 +0.26 1.O-O-O Nbd7 2.h3 Bxf3 3.gxf3 Qb6
4.Ne2 b4 5.Rhg1 a5 6.f5 a4 7.Kb1 Rfb8
8.Be3 b3 9.a3 bxc2+ 10.Qxc2 Qb3
11.Qxb3 Rxb3 12.Nc1 Rb6 13.Rg2 Rab8
14.fxg6 (8.772.889.594) 2431

31 60:08 +0.26 1.e5 Nh5 2.Ne2 dxe5 3.fxe5 f6 4.exf6 exf6
5.Bh4 Bxf3 6.gxf3 Re8 7.O-O-O Qd5
8.Nc3 Qxd4 9.Rhe1 Nd7 10.Bf2 Qf4
11.Bxg6 hxg6 12.Qxf4 Nxf4 13.Rxd7 a5
14.Rd6 (8.772.889.594) 2431

31 60:08 +0.25 1.Ne2 Nbd7 2.O-O Qb6 3.c3 c5 4.Bh4 Bxf3
5.gxf3 Rfb8 6.Kg2 a5 7.e5 c4 8.exf6 cxd3
9.fxg7 dxe2 10.Qxe2 e6 11.d5 exd5
12.f5 Kxg7 13.Qe7 Qc6 14.Rfe1 (8.772.889.594) 2431

I think that your engine probably has a better grasp of this position than the bleeding edge stockfish versions do.

Hello Dann, I'm afraid that would be much too much credit here, I think it is more likely blind luck it found these moves, unfortunately on the next iteration it are different moves

[d]rn1q1rk1/p3ppbp/2pp1np1/1p4B1/3PPPb1/2NB1N2/PPPQ2PP/R3K2R w KQ -

Engine: Sf20150403 MOD MP (512 MB)
by Tord Romstad, Marco Costalba and Joona

32 163:58 +0.24 1.Ne2 Nbd7 2.O-O Qb6 3.c3 c5 4.Bh4 Rfe8
5.a4 bxa4 6.Rxa4 a5 7.h3 d5 8.e5 Bxf3
9.Rxf3 Ne4 10.Qc2 cxd4 11.cxd4 Rec8
12.Nc3 e6 13.Re3 Rab8 14.Bxe4 (24.122.822.354) 2451

32 163:58 +0.23 1.f5 Nbd7 2.O-O b4 3.Ne2 c5 4.c3 Qb6
5.Kh1 Rab8 6.h3 Bxf3 7.Rxf3 Rfc8
8.Qe3 Qa5 9.a4 c4 10.Bc2 bxc3
11.bxc3 Rb2 12.Qc1 Rcb8 13.Bd2 Kh8
14.Kg1 (24.122.822.354) 2451

32 163:58 +0.12 1.Qf2 Na6 2.h3 Bxf3 3.Qxf3 Nb4 4.a3 Nxd3+
5.Qxd3 Qb6 6.Rd1 a5 7.O-O b4 8.Ne2 Qc7
9.f5 c5 10.Nf4 Qb7 11.Bxf6 Bxf6
12.dxc5 dxc5 13.c3 bxa3 14.bxa3 (24.122.822.354) 2451
___________________________________________________

33 265:46 +0.20 1.Ne2 Nbd7 2.O-O c5 3.c3 Qb6 4.Bh4 e6
5.Rae1 Rae8 6.Bb1 Rb8 7.h3 Bxf3
8.Rxf3 b4 9.e5 cxd4 10.exf6 dxc3+
11.Qd4 Qxd4+ 12.Nxd4 Bxf6 13.Bxf6 Nxf6
14.bxc3 (38.960.757.499) 2443

33 265:46 +0.17 1.f5 Nbd7 2.O-O b4 3.Ne2 Qb6 4.Kh1 c5
5.c3 Rab8 6.Rae1 Bxf3 7.Rxf3 Rfc8
8.Rff1 Qa5 9.Rc1 bxc3 10.bxc3 c4
11.Bb1 Rb7 12.a4 Qxa4 13.Bc2 Qc6
14.Qf4 (38.960.757.499) 2443

33 265:46 +0.12 1.Qf2 Na6 2.a3 Qb6 3.e5 Rae8 4.h3 Bxf3
5.Qxf3 dxe5 6.dxe5 Nd5 7.Nxd5 cxd5
8.Bh4 Nc5 9.Bf2 Qc6 10.O-O a6
11.Bxc5 Qxc5+ 12.Qf2 Qc6 13.c3 e6
14.Bc2 (38.960.757.499) 2443

Anyway, this particular version of Stockfish with Rainbow Serpent changes is probably the last one of the very basic version; the new one will try to do singular extensions again in PV nodes. The old one did not do that.

I just managed to compile a 32 bit version using Stockfish' latest development version codebase, with my super de luxe C++11 pthreads compiler for 32 bits that I found with the help of several people on Fishtest/Fishcooking ! Thanks to Vince and Lyudmil Antonov especially! It compiles and the new Stockfish Serpent runs!

It will not make any difference here because this seems to be an eval issue, related maybe to the search but I think Stockfish is searching for tactics here and tactics are not telling it anything useful. It is possible that White is really good here, I don't know but advantage seems to be way over the search horizon.

Two of the three choices are among those played (and the score from real game outcomes appears to be worth more than a pawn on average).

Ne2 is also interesting, since the other engines I tried did not list that one among the top three moves.

I will probably try a few strong engines on multi-pv run over at least 8 hours.

[peter]

Hi!

As for corr.- games 0-0 has by far the best results statistically, I think. In chessbase- corr.cbh I found 15 wins for White, 3 for Black und 7 remis.

Looking at possible transformations, I'd say the question is mainly, if long or short rochade is better for White.

0-0-0 has only about equal statistics for White and Black but there are even less games, same for e5.

I didn't find corr.- examples for Se2 at once.

What's your point, Dann? Have you found an improvement for Se2?

[peter]

I didn't find corr.- examples for Se2 at once.

What's your point, Dann? Have you found an improvement for Se2?

P.S. After 9.0-0 Qb6 10.Ne2 is played oftenly, but there a bunch of black answers have been tried then, so I think, we'd have a lot of work to do to get a clear answer to the question of the one and only best line.

[Dann Corbit]

Hi!

As for corr.- games 0-0 has by far the best results statistically, I think. In chessbase- corr.cbh I found 15 wins for White, 3 for Black und 7 remis.

0-0-0 has about equal statistics for White and Black but there are even less games.

Looking at possible transformations, I'd say the question is mainly, if long or short rochade is better for White.

I didn't find corr.- examples for Se2 at once.

What's your point, Dan? Have you found an improvement for Se2 or e5?

Most of the time, a 1/2 hour search will give me analysis that agrees with the best move from actual play as for move choice and the centipawn evaluation will agree approximately with the centipawn estimate based on wins, losses, and draws.

This is one of those odd positions where the engines seem to miss the plan. Probably a long idea search or an all day search by several strong engines will be needed to verify it.

In my database, I have categorized all the positions that are played a lot (5 times or more).
Most positions have the most frequently played move equal to the computer analyzed best move. OK, we know what to do and end of story for that position.

A smaller subset (maybe 5%) has one of the alternate positions as superior on the computer analysis. Sometimes that is quite valuable, especially when the search is deep and the analysis give a really high score to the alternative move. But when the move has a bad score and the games played indicated that the frequently played moves do well, then we have a problem on our hands. Very often, a simple half hour search will resolve it. But this position (for instance) stubbornly refuses to yield a sensible answer.

I also have an even smaller subset I call novelties. This set has computer generated moves that were not played by any of the human or computer participants that played that position in the past. When the computer generated score is better than the heuristic output for the position (based on wins, losses, and draws) then we have a real gem. But when the computer score is bad and the outcomes were good, we have another problem like above.

Here is a good novelty:
[d]2r1kbnr/1bqp1pp1/p3p3/1p5p/3QPP2/P1N1BB2/1PP3PP/R3K2R w KQk - acd 37; acs 7157; bm Qd3; cce -48; ce 24; pm O-O {13} h3 {9} O-O-O {5} Qd2 {5} Qa7 {4} Rd1 {1} Rf1 {1}; pv Qd3 Nf6 h3 h4 O-O Be7 Bd4 Nh5 Bxh5 Rxh5 Bxg7 a5 Rf2 b4 Qe2 Rh7 Nb5 Qc4 Be5 Qxe4 Qg4 Qg6 Qxg6 fxg6 axb4 axb4 Nd6+ Bxd6 Bxd6 Be4 Ra4 Bxc2 Rxb4 Kd8 Be5 Rf7 Rd2 d5; Eichborn ; white_wins 11; black_wins 15; draws 11;

Here is a bad one:
[d]r2q1rk1/pbpp1ppp/1pn1pn2/8/1bPP4/2N2NP1/PP2PPBP/R1BQ1RK1 w - - acd 35; acs 7005; bm Qc2; cce 162; ce 28; pm Ne5 {29} Qd3 {12} Bf4 {4} Bg5 {2} Nh4 {2} Qa4 {2} Be3 {1}; pv Qc2 h6 b3 Bxc3 Qxc3 Ne7 Bb2 d6 Qd2 Be4 d5 e5 Nh4 Bxg2 Nxg2 Ng6 f3 a5 e4 Qd7 Ne3 h5 Bc3 Rfb8 Qc2 h4 Nf5 hxg3 hxg3 Ne7 Ne3 a4 Rfd1; Eichborn ; white_wins 32; black_wins 13; draws 7;

[peter]

Thanks for sharing, Dann.

As for the thread- position, I think something like the backward-solving of a short corr.- game, I picked out of the ones I looked at, useable as a test- line.
Here it is with SF's .pgn of automatic backward with 32G hash, 24 threads and 10 min./ply. The game was 28 moves only after a blunder at 18th by Black. From there I started the engine's backward-solving.

Code: Select all

[Event "ICCF 2002"]
[Site "?"]
[Date "????.??.??"]
[Round "?"]
[White "Viberg,J-Knudsen,J , J.."]
[Black "Knudsen,J , J."]
[Result "1-0"]
[ECO "B07"]
[Annotator "SF"]
[PlyCount "36"]
[EventDate "1979.??.??"]
[EventType "tourn (corr)"]
[EventRounds "9"]
[EventCountry "?"]
[Source "ChessBase"]
[SourceDate "2000.04.19"]

1. e4 d6 2. d4 Nf6 3. Nc3 g6 4. Bg5 Bg7 5. Qd2 c6 6. f4 O-O 7. Nf3 b5 8. Bd3 Bg4 9. O-O Qb6 10. Ne2 Nbd7 11. Kh1 d5 12. Bxf6 {0.19/36 letzter Buchzug} Nxf6 {0.40/41} ({Stockfish 120415 64 POPCNT:} 12... exf6 13. exd5 Bxf3 14. Rxf3 c5 15. f5 cxd4 16. Rg3 Rfe8 17. Rg4 Ne5 18. Rxd4 Bf8 19. c3 Bc5 20. Re4 Rad8 21. b4 Bf2 22. Nf4 Ng4 23. Rxe8+ Rxe8 24. g3 g5 25. Nh3 h5 26. Bf1 a6 27. Bg2 Rc8 28. a4 Be3 29. Qe1 Qd6 30. axb5 axb5 31. Bf3 Qe5 {0.19/36}) 13. e5 {0.43/40} Nd7 {0.39/39} 14. h3 {0.50/38} Bxf3 {0.45/38} 15. Rxf3 {0.42/37} e6 {0.54/37} 16. h4 {0.63/37} c5 {0.47/37} 17. c3 {0.43/34} f6 {0.85/40} ({Stockfish 120415 64 POPCNT:} 17... b4 18. h5 Rab8 19. Rh3 Qc6 20. g4 c4 21. Bc2 bxc3 22. bxc3 Rb2 23. Ng1 Nc5 24. dxc5 d4+ 25. Rf3 d3 26. Ba4 Rxd2 27. Bxc6 Rc8 28. Bd7 Rxc5 29. Rff1 Rc7 30. Bb5 Rc2 31. hxg6 hxg6 32. Rac1 Rxa2 33. Nf3 d2 34. Rcd1 Rb2 { 0.43/34}) 18. h5 {0.69/40} gxh5 {1.85/37 1-0 (28) Viberg,J-Knudsen,J ICCF email 2002} ({Stockfish 120415 64 POPCNT:} 18... fxe5 19. fxe5 Rxf3 20. gxf3 Nxe5 21. dxe5 c4 22. Nf4 cxd3 23. hxg6 h6 24. Kg2 Rf8 25. Nxd3 Qd8 26. Rh1 d4 27. cxd4 Qxd4 28. Qe2 Qd5 29. Rc1 Qxa2 30. Qe4 Qa5 31. Rc3 Qd8 32. f4 Qe8 33. Kf2 Qe7 34. Ke2 Qh4 35. Kd1 a5 36. Kc2 Qh2+ 37. Kb1 Qd2 38. Rc7 Qd1+ 39. Ka2 Qa4+ 40. Qxa4 bxa4 {0.69/40}) 1-0

As always it's only the relation between the evals over the course of the plies that counts as for my pov