12

Basically, I would like to ask people to share their knowledge on famous and/or particularly instructive examples of chess positions, which chess engines do not understand well (i.e. do not evaluate correctly).

A good example of the kind of positions I am interested in is the Vančura Position-Wiki (cf. also Vančura Position-Chessbase), which occurred very recently (at the Sinquefield Cup 2014) in the game Carlsen-Aronian. Although the engine was saying white is completely winning, the game was a theoretical draw after black's 48th move.

Thanks in advance!

P.S. I should stress that I am looking for realistic examples, i.e. those which are likely to happen in a real game. Positions of the following type,

[fen "4k3/4p3/3pPp2/2pP1Pp1/1pP3Pp/pP1K3P/P2RRR2/8 w - - 0 1"]

and similar artificial positions is not the kind of thing I am looking for.

I should also say that endgame tablebase examples are OK. The only requirement is that engines (run on an average computer) are not able to evaluate the position correctly after given, say, 10-15 minutes time to "think".

ETD
  • 23,452
  • 4
  • 96
  • 169
Zvonimir
  • 1,315
  • 1
  • 11
  • 23
  • 2
    Maybe "engine" is a bit too vague. I keep wondering why "machines" are often believed to have one unique chess style. – Pierre Arlaud Sep 13 '14 at 11:40
  • 1
    @Pierre Arlaud: Specifically, when I used the word "engine", I had for example Stockfish or Houdini in mind (the chess engines most often used in live broadcasts of chess events, and considered to be the the best widely used chess engines). The purpose of my question was to pinpoint the most important situations, in which these, most commonly used and trusted, tools fail to provide a satisfactory or correct answer. And also, because the positions which engines do not evaluate well are often very instructive and interesting to analyze (when they are realistic). – Zvonimir Sep 13 '14 at 12:22
  • Are you mainly looking for positions other than fortresses? –  Sep 13 '14 at 12:36
  • @Ricky Demer: Not necessarily. The examples can include fortresses, but since there are many types of fortresses and most of them are a problem for engines to evaluate correctly, I am searching for particularly aesthetic and instructive examples of fortresses; the ones that really catch your eye when you see them. Also, an example of a famous game featuring some nice fortress ideas would also be appreciated. The stress is on aesthetics and theoretical or historical importance of the examples. – Zvonimir Sep 13 '14 at 13:21
  • 1
    This game has what I think is a particularly instructive example. –  Sep 13 '14 at 14:58
  • @Ricky Demer: Brilliant, thank you! Btw, I think I am going to remove the word "list" from the title of the question, since it seems to discourage people from answering the question :) – Zvonimir Sep 13 '14 at 16:14
  • 1
    Regarding your PS: What about this game? –  Sep 13 '14 at 17:01
  • The position is illegal: White has 3 rooks so at least one must be promoted. But white has 8 pawns. – Lucenaposition Aug 03 '23 at 23:26

4 Answers4

10

The following two links seem to be exactly what you are looking for (many such positions in each one of them).

http://en.chessbase.com/post/can-we-still-deceive-computers-in-chess

http://www.chess.com/forum/view/general/positions-engines-get-wrong--please-contribute

A cute study, which engines are unable to solve last I checked (white to move and win):

6r1/p1p1p1pP/P1PpP1P1/8/2P1B3/2K3pp/3P2rq/R5bk w - - 1 0
GloriaVictis
  • 3,756
  • 2
  • 28
  • 47
  • I think the study might be cooked. The mainline goes 1.hxg8=N d5 2.Bf3 d4+ 3.Kb4 d3 4.Nh6 gxh6 5.g7 h5 6.g8=N h4 7.Nf6 exf6 8.e7 f5 9.e8=N f4 10.Nd6 cxd6 11.c7 d5 12.c8=N d4 13.Nb6 axb6 14.a7 b5 15.a8=B bxc4 16.Bae4 c3 17.Bxd3 c2 18.Bf1 d3 19.Kb3 c1=Q 20.Rxc1 Bd4 21.B1xg2# 1-0. But 12... dxc4 poses problems. The intended solution continues 13.Nb6 c3 14.Kc4 cxd2 15.Kxd3, but 14... axb6 seems to throw a spanner in the works. See for example this discussion in Dutch. – kahen Mar 06 '15 at 19:52
  • Google translate link (for the Dutch-impaired - such as me). – kahen Mar 06 '15 at 19:53
5

http://chesstempo.com/chess-tactics/90023 The final position of the solution is an ending with queen against two knights with pawns on both sides where the knights can set up a blockade. Looks like a draw but the engine evaluates that the queen side wins.

8/5ppk/7p/3pP3/5P2/5N1P/4NKP1/7q b - - 0 5
MichalRyszardWojcik
  • 3,300
  • 16
  • 39
4
1Q2r1k1/P6p/8/8/4q3/7P/6PK/8 w - - 2 6

White to move. 1.a8=Q Qxa8 2.Qg3+ may be a perpetual check that is hard to analyze. See the comments in the discussion of this tactics problem at http://chesstempo.com/chess-problems/103006

MichalRyszardWojcik
  • 3,300
  • 16
  • 39
  • Really, such a simple looking position? If computer finds this hard to analyze, then this would be one of the best answers for the question. – Cyriac Antony Nov 30 '18 at 10:19
3

Besides fortresses, another type of position the engines tend to get wrong is when the weaker side is about to promote a pawn but the promotion can be delayed indefinitely. 

[FEN "5Q2/8/8/8/2k5/7P/1p4PK/r7 w - - 0 1"]
[White "Anatoly Karpov"]
[Black "Anthony Miles"]
[Event " Biel"]
[Date "1990  (analysis, white to move)"]

For a human player it is easy to see that this is a draw, but the engine has to try all the different checks from all possible squares in all different orders. (I haven't tested this position with the newest engines, perhaps some of them can handle this?)

Dag Oskar Madsen
  • 9,031
  • 1
  • 36
  • 61