Optimizing for a different target than elo

[Uri Blass]

I would like to know if there are programmers who try to test for a different target than bigger elo.

Possible target that I would like to see is winning as fast as possible.


You can decide that every game is a set of 2 games from the same position and you decide that if the chess result in the 2 games is 1-1 by 2 wins then the winner is the side who needed less moves to mate and if both needed the same number of moves it is a draw.

I wonder if somebody tried to train NN for this target.
You can decide that all the possible positions after 3 plies are the starting positions that you test so you get also some obviously winning positions like 1.e4 Nf6 2.Qh5 when in case black win both games the winner is the side who mated in less moves.

I think that we need it to stop the stupidity of programs that are based on NN that do not care about the speed of winning and as a result do not play well with a big handicap.
I do not claim that losing as slow as possible is best when you play with a big handicap because it may be better to set traps for the opponents but it is clearly better than what NN's do today.