Odd Simple Position

[Martin Brown]

I found a curiosity in a trivial won endgame where Shredder10 played a rather strange move (with or without bitbases enabled). The position is:

[D]8/5pkp/8/2K5/3Nq3/8/8/8 b - - 0 47

And black surprisingly plays QxN rather than h5 or f5.
Following QxN whites response is even wierder. The king runs away from the queen rather than capturing it and the evaluations are wacky. The queen moves to be adjacent to the king until they hit the edge of the board and white has no option but to capture. The evaluation is like misere chess!

-279.93 Kc6 Qc5+ Kb7 Qb6+ Kc8 Qc7+ KxQ
-279.95 Kb5 Qc5+ Ka4 Qb4+ KxQ
-279.99 KxQ

And then after KxQ it immediately sees the right answer!

Other modern engines see the correct lines without much difficulty. Please can someone explain why this position foxes an otherwise strong engine?

Crafty 19.19 gets the right order but fails to get the right mate depth right of the king moving. Crafty 20.14 solves it almost instantly.

Thanks for any enlightenment.

[bob]

I found a curiosity in a trivial won endgame where Shredder10 played a rather strange move (with or without bitbases enabled). The position is:

[D]8/5pkp/8/2K5/3Nq3/8/8/8 b - - 0 47

And black surprisingly plays QxN rather than h5 or f5.
Following QxN whites response is even wierder. The king runs away from the queen rather than capturing it and the evaluations are wacky. The queen moves to be adjacent to the king until they hit the edge of the board and white has no option but to capture. The evaluation is like misere chess!

-279.93 Kc6 Qc5+ Kb7 Qb6+ Kc8 Qc7+ KxQ
-279.95 Kb5 Qc5+ Ka4 Qb4+ KxQ
-279.99 KxQ

And then after KxQ it immediately sees the right answer!

Other modern engines see the correct lines without much difficulty. Please can someone explain why this position foxes an otherwise strong engine?

Crafty 19.19 gets the right order but fails to get the right mate depth right of the king moving. Crafty 20.14 solves it almost instantly.

Thanks for any enlightenment.

There are lots of issues, particularly if you use EGTBs or anything that emulates their behaviour. A program might well not play KxQ if you have just a few EGTB files, but you do have KPPK because it would realize that playing KxQ drops into an EGTB positiont hat is mate in N, while with KQPP vs K (if you don't have that EGTB) it would be well behind, but it might not see the mate.

I tried this on my local machine, and by using just 3-4 piece endgame tables, Crafty will play QxN instantly as it leads to a mate. With 5 piece tables, it plays h5 which leads to a shorter mate (h5=M15, Qd4=M18).

[hgm]

Shredder must have recognizers that recognize the KPPK as won, and award an enormous score to it, much more than merely being a Queen ahead. And it aparently does not search deep enough to see the checkmate with the Queen, so leaving the Queen on the board cannot give it a better score.

[muxecoid]

I can imagine a human playing like that in blitz with Fischer time control. Instead of risking blundering a fork you simplify.

[michiguel]

I can imagine a human playing like that in blitz with Fischer time control. Instead of risking blundering a fork you simplify.

Moreover, you will eliminate the possibility to lose on time. But the best thing is to put either the queen or the king at one square distance diagonally. It will take 3 moves for the knight to check. Then you put your king again at the same distance. That is the best way not to blunder with seconds on the clock.

Miguel