Phantom Draw - test position

[Martin Brown]

[D] r2r2k1/pp3pbp/2p1b1p1/8/P2qPP2/1P1pnQ2/3N1RBP/2R1N2K b - - 0 25

In this quirky position black makes a mistake and plays 25. ... Qb2
White's only plausible response is 26. Rb1 and then strong engines miss blacks crucial winning continuation of 26. ... Qd4 blinded by white 27. Rc1 which offers a black (winning) a potential draw by repetition.

So the main engine line continues with the bizarre swapoff line:
26. ... Nxg2 27. Rxb2 Nxe1 28. Qd1 Bxb2 29. Qxe1
which is way off the mark

The only two engines I have tried that can find 26. ... Qd4 at this point are Crafty and Shredder. All the others stubbornly think 27. Rc1 draws and will sacrifice the queen to avoid it leaving white in much better shape.

I think the correct main line for black should be to simplify...
26. ... Qd4 27. Qg3 Bg4 28. Nef3 Bxf3 29. Rxf3 etc.

I spotted this because the game score blundercheck annotation made no sense. Shredder and Crafty return polluted search values but do play the right moves when the time comes but everything else is completely blind to the correct continuation even after 26. Rb1 is played out on the board!

Not only that but most of them cannot see into the position even after the transition tables have been preloaded with the right continuation lines.

Most odd. I presume this position with tactical blindness is exceptional in some way otherwise it suggests a way to play anti-computer chess.

I hope I got the quote a diagram thing right - been a while since I posted here (doesn't look rigth in preview mode). Thanks for any enlightenment.

[MikeGL]

That's strange, in my case, most engines consider 26. ... Qd4.
Both FireBird and Critter1.2 considers 26. ... Qd4 instead of 26. ... Nxg2.
I believe other strong engines will easily consider Qd4 too.
It's common in todays engines to vary their main line when they detect 3-fold-repetition, they usually make another move which is just as strong as the repeated move, just to avoid a draw.

btw, what interface and OS are you using?

[Martin Brown]

Chessbase v10 and major engines which fell for Nxg2 included

Junior7, Fritz8, Fritz10, Rybka2.32 and Stockfish 1.51

OK some of them are perhaps a bit long in the tooth but I was a bit surprised by Fritz 10 failing to annotate the game correctly.

I will give the engines you have suggested a whirl. Thanks!

[UncombedCoconut]

OK some of them are perhaps a bit long in the tooth but I was a bit surprised by Fritz 10 failing to annotate the game correctly.

They certainly are. Newer Stockfishes, Critter, Komodo, Toga, Spark, even Arasan solve this instantly. If you say the diagram looks funny, maybe the position you showed your engines isn't the same as the one you posted?

[Michel]

Newer Stockfishes, Critter, Komodo, Toga, Spark, even Arasan solve this instantly

Did you actually check this? I just tried Toga 1.4.1SE and it doesn't see Qd4 after
Qd2, Rb1.

The bug is due to the fact that engines often do not distinguish between repetitions
in search and in the game history.

This causes the engine to misannotate games.

[UncombedCoconut]

Yes, but it turns out I had the older Toga II 1.3.1 for some reason. My Arasan was also old (13.1); the rest were at their latest versions.

[Martin Brown]

The bug is due to the fact that engines often do not distinguish between repetitions in search and in the game history.

This causes the engine to misannotate games.

Not only misannotate games but also to be vulnerable to best move avoidance in game play if a partial repetition sequence can be forced. The Fritz10 engine played Nxg2 losing the queen to avoid moving it back to d4 and creating a phantom draw via Rc1.

It looks to me like the logic should be that the side with material advantage should never make the final repetition, but must consider a first repetition in full to avoid this sort of gotcha. I was surprised that most of the engines I tried also quickly lost the plot even when the cache was preloaded with correct game continuation after Qd4.

[Michel]

Not only misannotate games but also to be vulnerable to best move avoidance in game play if a partial repetition sequence can be forced. The Fritz10 engine played Nxg2 losing the queen to avoid moving it back to d4 and creating a phantom draw via Rc1.

But in normal circumstances an engine playing black would not have played Qb2 in the first place. Did Fritz10 really make this move by itself?

If the engine played white it would not care that black makes a bad move after Rb1.

It looks to me like the logic should be that the side with material advantage should never make the final repetition, but must consider a first repetition in full to avoid this sort of gotcha. I was surprised that most of the engines I tried also quickly lost the plot even when the cache was preloaded with correct game continuation after Qd4.

In search it is allowed to consider a position a draw after the first repetition provided the earlier identical position is also part of the search. But this is not allowed if the earlier identical position is part of the game history.

[UncombedCoconut]

Yes, but it turns out I had the older Toga II 1.3.1 for some reason. My Arasan was also old (13.1); the rest were at their latest versions.

I just remembered that SCID sends a FEN rather than move history. That's the problem.

[UncombedCoconut]

Yes, but it turns out I had the older Toga II 1.3.1 for some reason. My Arasan was also old (13.1); the rest were at their latest versions.

I just remembered that SCID sends a FEN rather than move history. That's the problem.