Machines like it concretely

The areas in which human intellect outwits the machines’ calculation power become smaller and smaller. But they are still there. Machines don’t understand fortresses for instance, often misevaluate opposite coloured bishops or have trouble grasping long term concepts in general.

When German U18 champion Roven Vogel crushed Indian wunderkind Karthik Thrish in Bamberg recently, he created an exemplary win that machines are not yet able to reproduce. Neither Stockfish nor other engines find Vogel’s 20.Ne6, cutting off the opponent’s queen from the kingside defense and setting up the decisive attack at the same time. The white concept is too abstract, machines like it concretely.

But they’re getting better. The correspondence game below has been annotated two years ago. While the general remarks and concepts still stand, today’s Stockfish 9 handles some of the crucial spots much better than its predecessor Stockfish 6 that we refer to whenever the annotation says “Stockfish”. For instance at move 11 in 2016 we needed to overrule the machine in order to play 11.Nxb3. Stockfish 9 likes the move.

Werbung

To win the won position at move 23 the abstract concept of “creating a second front” is required. In order to follow it White needs to refrain from winning back his sacrificed material immediately. Stockfish 6 couldn’t do it at all, Stockfish 9 at least smells that there are alternatives to pawngrabbing (which would most likely lead to an endgame a pawn up that is drawn nevertheless).

Hero (2.219) – Villain (2.060)
Correspondence game, Lechenicher Schachserver 2016

1. d4 d5 2. c4 c6 3. Nf3 e6 4. g3

Opening choice and opening analysis are two of the keys to correspondence chess. The chosen opening must reflect the fact that calm positional play won’t work. The classical Catalan for instance, a decent OTB opening, may lead to slightly pleasant positions and some pressure for White. But the black player and his engine will not get tired and collapse like in an OTB game. Instead he will just hold.

If you want to beat a centaur you need to create imbalances and dynamics and understand their implications better than the opponent. The barely explored Catalan setups that include a pawn sacrifice on c4 are a good choice as long as Black accepts the sacrifice. If he doesn’t White will have a hard time getting something substantial out of a regular Catalan.

4… dxc4 5. Bg2 b5 6. O-O Bb7 7. a4 Nf6 8. Ne5 Qc8 9. b3 cxb3

I wanted to point out positions where the engine struggles and needs to be guided, but here it was the other way around: The engine guided and overruled me. This position has been reached a few times in 2500+ GM play, and in each of these games except one White went with the obvious and seemingly natural 10.Qxb3 – which Stockfish hates. I couldn’t see what’s wrong with the GM move Qxb3, but Stockfish prefers Nd2 by almost 0.5 pawns for no clear reason, so what to do?

To make sure I took a timeout in order to study some annotated GM games where the players outlined concepts and principles White should follow in these kind of positions. But Stockfish hated many of the lines given by grandmasters as well, and in no case could I prove it wrong. In the end I came to the conclusion that the engine understands this structure better than humans and went with Nd2.

10. Nd2

In fact the great Alexander Khalifman ist the only human who has ever played this instead of 10.Qxb3. Not the worst guy to follow. White points out that the c5 square and Black’s backward pawn on c6 are part of his immediate plans.

(Akobian, Varuzhan – Schneider, Dmitry, USA-ch 2004 (1-0, 33) went 10. Qxb3 b4 11. Bb2 Nbd7 12. Nxd7 Qxd7 13. Rc1 and now the machine claims that after 13… a5 the white compensation is not sufficient.)

10… Nfd7 11. Nxb3

This isn’t among the top choices of the engine which doesn’t think that the structure change after …Nxe5 dxe5 favours White. Stockfish even prefers 11.Nf3 over 11.Nxb3, while a human with some OTB Catalan experience likes the e5 pawn cramping the black position, the e3-a7 diagonal for his bishop pressuring the queenside and potentially the open d file and the d6 square. The machine was probably right with insisting on 10.Nd2, but now it is time to overrule it.

11… Nxe5 12. dxe5 b4 13. Na5 Qc7 14. Nxb7 Qxb7 15. a5 Be7 16. Qa4 O-O 17. Be3 Qa6 18. Ra2 Rc8 19. Rc1 Qb7 20. a6 Qc7 21. f4 Qd8 22. Rd2 Qc7 23. Bf3 g6

Black is paralyzed, can’t develop his queenside, White is winning. But despite its 3.300 Elo an engine on its own would not be able to find a winning plan. No matter how long you let it calculate, the engine will go for the b4 and c6 pawn, free Black’s pieces in the process and end up with an unwinnable endgame 5 vs 4 pawns on one wing.

For a human the winning idea is easy to come up with: Open a second front. With Black being tied down on the queenside opening the kingside should be decisive. And White has all the time in the world to execute this. Black can’t do anything.

24. Kg2 Kg7? 

This even invites g3-g4 followed by f4-f5 since …Qxe5 defenses are now out of the picture due to Bd4 pinning the queen. However, Stockfish at depth 30 refuses to go for the winning blow and insists on playing 25.Rb2?.

25. g4! Rd8 26. Rxd8 Qxd8 27. f5

Black resigned.

A rook and a knight down effectively he is now facing an attack that will be crushing. Being led to this point the engine has no trouble to quickly realize that Black is lost.

1-0