Stockfish 1.6 - does not recongnize a lost pawn endgame

[pgeorges]

In the following position, Stockfish sets the score to draw (0.0). Any player see in half second it is lost for white (and other engines also). I think Stockfish does not get that there is a far pawn on the side, which nearly always lead to a win by decoying the opponent king.

[d]8/6p1/2p4p/pp2k3/P7/2PK2P1/7P/8 w KQkq - 0 1

Pascal Georges

[George Tsavdaris]

In the following position, Stockfish sets the score to draw (0.0). Any player see in half second it is lost for white (and other engines also). I think Stockfish does not get that there is a far pawn on the side, which nearly always lead to a win by decoying the opponent king.

[d]8/6p1/2p4p/pp2k3/P7/2PK2P1/7P/8 w KQkq - 0 1

My Stockfish 1.6 does get the loss for white instantly, so i guess you have a somehow flawed version of Stockfish.


Analysis by Stockfish 1.6 JA:

1.a4xb5 c6xb5 2.Kd3-c2 a5-a4 3.Kc2-b2 Ke5-d5
-+ (-2.74) Depth: 6 00:00:00 1kN
1.Kd3-c2 b5xa4 2.Kc2-b2 Ke5-e4 3.Kb2-a3
-+ (-2.54 !) Depth: 6 00:00:00 1kN
1.Kd3-c2 Ke5-e4 2.Kc2-b2 h6-h5 3.Kb2-b3 g7-g5 4.c3-c4
-+ (-2.10) Depth: 7 00:00:00 4kN
1.Kd3-c2 Ke5-e4 2.Kc2-b2 h6-h5 3.Kb2-b3 Ke4-d3 4.h2-h4 g7-g6 5.a4xb5 c6xb5
-+ (-2.38) Depth: 8 00:00:00 7kN
1.Kd3-c2 Ke5-e4 2.Kc2-b2 h6-h5 3.Kb2-b3 Ke4-d3 4.h2-h4 g7-g6 5.a4xb5 c6xb5
-+ (-2.38) Depth: 9 00:00:00 13kN
1.Kd3-c2 Ke5-e4 2.Kc2-b2 h6-h5 3.Kb2-b3 Ke4-d3 4.h2-h3 g7-g5 5.h3-h4 g5-g4 6.a4xb5 c6xb5
-+ (-2.38) Depth: 10 00:00:00 26kN
1.Kd3-c2 Ke5-e4 2.Kc2-b3 Ke4-d5 3.g3-g4 g7-g5 4.h2-h3 Kd5-c5
-+ (-2.58 --) Depth: 11 00:00:00 72kN
1.Kd3-c2 Ke5-e4 2.Kc2-b3 Ke4-d5 3.g3-g4 g7-g5 4.Kb3-a3 Kd5-c4 5.Ka3-b2 b5-b4 6.c3xb4 Kc4xb4 7.h2-h3 c6-c5 8.Kb2-c2 Kb4xa4
-+ (-2.70) Depth: 12 00:00:00 142kN
1.Kd3-c2 g7-g5 2.Kc2-b3 g5-g4 3.Kb3-a3 Ke5-d5 4.Ka3-b3 Kd5-c5 5.a4xb5 c6xb5 6.Kb3-a3 Kc5-c4 7.Ka3-b2 a5-a4
-+ (-2.78) Depth: 13 00:00:00 334kN
1.Kd3-c2 g7-g5 2.Kc2-b3 g5-g4 3.Kb3-a3 Ke5-d5 4.Ka3-b3 Kd5-c5 5.a4xb5 c6xb5 6.Kb3-b2 Kc5-c4 7.Kb2-c2 h6-h5 8.Kc2-b2 Kc4-d5 9.Kb2-b3
-+ (-2.74) Depth: 14 00:00:00 461kN
1.Kd3-c2 g7-g5 2.a4xb5 c6xb5 3.Kc2-b3 Ke5-e4 4.c3-c4 b5-b4 5.h2-h3 Ke4-d4 6.h3-h4 g5xh4 7.g3xh4 h6-h5
-+ (-2.94 --) Depth: 15 00:00:01 559kN
1.Kd3-c2 g7-g5 2.a4xb5 c6xb5 3.Kc2-b3 Ke5-e4 4.c3-c4 b5-b4 5.h2-h3 Ke4-d4 6.h3-h4 g5xh4 7.g3xh4 h6-h5
-+ (-3.79 --) Depth: 16 00:00:01 1013kN
1.Kd3-c2 g7-g5 2.a4xb5 c6xb5 3.Kc2-b3 Ke5-e4 4.Kb3-c2 a5-a4 5.Kc2-d2 g5-g4 6.Kd2-c2 Ke4-f3 7.Kc2-b2 Kf3-g2 8.c3-c4 b5xc4 9.Kb2-c3 Kg2xh2 10.Kc3xc4 Kh2xg3
-+ (-5.25) Depth: 17 00:00:02 2241kN
1.a4xb5 c6xb5 2.c3-c4 b5-b4 3.c4-c5 a5-a4 4.c5-c6 Ke5-d6 5.c6-c7 Kd6xc7 6.Kd3-c4 a4-a3 7.Kc4-b3 h6-h5 8.Kb3-c2 Kc7-d6 9.Kc2-b3 Kd6-d5 10.h2-h3 Kd5-d4 11.Kb3-a2 Kd4-d5
-+ (-4.56) Depth: 17 00:00:03 2681kN
1.a4xb5 c6xb5 2.c3-c4 b5-b4 3.c4-c5 a5-a4 4.c5-c6 Ke5-d6 5.c6-c7 Kd6xc7 6.Kd3-c4 a4-a3 7.Kc4-b3 h6-h5 8.Kb3-c2 Kc7-d6 9.Kc2-b3 Kd6-d5 10.h2-h3 g7-g5 11.Kb3-a2 Kd5-c4 12.g3-g4 h5-h4 13.Ka2-b1 Kc4-d5
-+ (-6.06) Depth: 18 00:00:04 4314kN
1.Kd3-c2 g7-g5 2.Kc2-b3 g5-g4 3.Kb3-a3 Ke5-e4 4.Ka3-b3 Ke4-f3 5.a4xb5 c6xb5 6.c3-c4 b5xc4+ 7.Kb3xc4 Kf3-g2 8.Kc4-d3 Kg2xh2 9.Kd3-d2 a5-a4 10.Kd2-d1 Kh2xg3
-+ (-5.73) Depth: 18 00:00:05 4871kN
1.Kd3-c2 g7-g5 2.Kc2-b3 Ke5-e4 3.a4xb5 c6xb5 4.c3-c4 b5-b4 5.c4-c5 Ke4-d5 6.h2-h4 g5xh4 7.g3xh4 Kd5xc5 8.Kb3-b2 Kc5-c4 9.Kb2-c1 a5-a4 10.Kc1-b2 h6-h5 11.Kb2-b1 Kc4-c3 12.Kb1-a2 Kc3-c4 13.Ka2-b2 Kc4-c5 14.Kb2-a2 a4-a3 15.Ka2-b3 Kc5-d5 16.Kb3-c2 Kd5-c4
-+ (-6.06) Depth: 19 00:00:07 7015kN

[mcostalba]

In the following position, Stockfish sets the score to draw (0.0). Any player see in half second it is lost for white (and other engines also). I think Stockfish does not get that there is a far pawn on the side, which nearly always lead to a win by decoying the opponent king.

[d]8/6p1/2p4p/pp2k3/P7/2PK2P1/7P/8 w KQkq - 0 1

Pascal Georges

Are you using a gcc compiled version of Stockfish 1.6 ?

There have been many reports of 'strange' behaviours when compiling with gcc under Linux.

[royb]

In the following position, Stockfish sets the score to draw (0.0). Any player see in half second it is lost for white (and other engines also). I think Stockfish does not get that there is a far pawn on the side, which nearly always lead to a win by decoying the opponent king.

[d]8/6p1/2p4p/pp2k3/P7/2PK2P1/7P/8 w KQkq - 0 1

Pascal Georges

Are you using a gcc compiled version of Stockfish 1.6 ?

There have been many reports of 'strange' behaviours when compiling with gcc under Linux.

With my gcc compile of SF-1.6 under SCID 4.1, SF-1.6 crashes in about 2 seconds on this position.

I should add that SF-1.5.1 also crashes quickly with this position using SCID 4.1.

Roy

[mcostalba]

In the following position, Stockfish sets the score to draw (0.0). Any player see in half second it is lost for white (and other engines also). I think Stockfish does not get that there is a far pawn on the side, which nearly always lead to a win by decoying the opponent king.

[d]8/6p1/2p4p/pp2k3/P7/2PK2P1/7P/8 w KQkq - 0 1

Pascal Georges

Are you using a gcc compiled version of Stockfish 1.6 ?

There have been many reports of 'strange' behaviours when compiling with gcc under Linux.

With my gcc compile of SF-1.6 under SCID 4.1, SF-1.6 crashes in about 2 seconds on this position.

Roy

Under Windows both with Intel (JA compiles) and with MSVC it works without problems and clearly sees a win for black in less then a second.


So I would think that gcc miscompiles Stockfish 1.6

Only tomorrow I will be able to verify this and so find the fix....also this time we will need a 1.6.1, we are never lucky with the .0 release

[pgeorges]

Are you using a gcc compiled version of Stockfish 1.6 ?

There have been many reports of 'strange' behaviours when compiling with gcc under Linux.

Yes, I am using GCC 4.3.2 under Linux. And by removing the option -O3 in the Makefile, Stockfish sees the position lost for white immediatly. So I guess it is a bug in GCC compiler.

Pascal

[mcostalba]

Are you using a gcc compiled version of Stockfish 1.6 ?

There have been many reports of 'strange' behaviours when compiling with gcc under Linux.

Yes, I am using GCC 4.3.2 under Linux. And by removing the option -O3 in the Makefile, Stockfish sees the position lost for white immediatly. So I guess it is a bug in GCC compiler.

Pascal

I have a report from a private email where it seems that gcc 4.2.3 works flawless.

Please try -O2 instead of -O3 and check if it works for you.

[pgeorges]

With my gcc compile of SF-1.6 under SCID 4.1, SF-1.6 crashes in about 2 seconds on this position.

I should add that SF-1.5.1 also crashes quickly with this position using SCID 4.1.

Roy

Scid 4.1 ? Not released yet !

Under Scid, both Stockfish 1.5.1 and Stockfish 1.6 works for me (even if for the latest evaluation is not always perfect with GCC).

Pascal Georges

[pgeorges]

Please try -O2 instead of -O3 and check if it works for you.

Indeed, with -O2 Stockfish gets the position immediately. Note that I may be wrong when stating it is a compiler issue. I remember that when switching -On to various values it would reveal some bugs in my own code. Maybe it is also the case here. For example it may be a different way of packing structs (32 bits alignment), hence a different code (this could explain crashes, or SF picking the wrong data hence the wrong score when you are lucky enough to avoid a segfault).

Pascal Georges

[Hart]

I've tried this position with a dozen engines but it fails every time with Rybka 3, RobboLito 0.085g3, Stockfish 1.5.1 (intel x64), and Stockfish 1.6 (x64 ja).
Also, I don't know if this means anything but this FEN shows up as a Chess 960 position when I input it into Arena. Maybe a problem reading the FEN? I noticed the original FEN has the KQkq tags. Does that mean castling should be allowed when it cannot?