an idea for testing programs in endgame

[Uri Blass]

1)take a big database of chess games that ended in a draw and give the program to statically evaluate the final position.

find all the cases when the program evaluated the position statically as more than 1 pawn advantage for one side and order these position based on score.

It is possible that in small part of the pgn there is a win in the position that the players agreed to a draw so analyze the relevant positions in order to take these positions out of the set.

Now you have a set of positions that chess programs evaluate wrong.
You can think about rules to improve your evaluation to make the evaluation of these positions closer to draw.

Uri

[sje]

A similar idea would be to take a collection of games ending in a resignation where the resigning side is evaluated to be even or ahead. These games are nearly guaranteed to have an error either by the player or by the analyzing program.