Engines are bad(?) at endgame evaluation


  In endgame evaluation we should not always believe Chess engines blindly. they shows incorrect value in many endgame positions. such as the position below is valued by nearly +5 for white, by all top engines!

   but, this is a draw game by perfect play. interestingly, the perfect play is very easy for black - he has to move his king b8 and c7 repeatedly.  


Engines are bad at evaluating any position whose moves must have an impact greater than 10-15 moves later.  Within this window they are very accurate, but outside of it they are completely blind.

Endgames provide countless examples of computer misevaluations.  But also closed positions and openings.  Their moves are not bad, but you can't trust the number evaluation without doing some work with the engine.


Years ago there was a famous example of just how bad computers are at endgames. Black was up two rooks that were blocked behind a completely locked-up board. One of them was en prise- but of course taking it would open up the board and eventually black would use the remaining rook to win. It took decades of development for engines to realize that taking the rook was a bad thing.


 Another example:

 Houdini 3 shows +19 for white,

 but it's draw!!


Your second example is not very probable to say the least Smile

But there are even worse things, such as this one that the computer evaluated as winning for me (White) (this is from memory, but the main characteristics were the same). When you ask him the moves he just moves the kings forth and back without achieving anything - obviously there is no way to get through the position.


Irontiger, the position u showed is evaluated by 0.00, by houdini, Critter,    

and stockfish. only rybka 4 shows +1.91