Wednesday, July 16, 2008

Unsolved problems in computer chess

Computers play certain positions poorly, for example endgames with more pieces than the available tablebase.

In a related vein, computers sometimes lose on time, or by an unsound sacrifice, in drawn pawn-fortress positions.

There are no good techniques for a computer to automatically discover good chess strategy on its own from first principles. We seek a theory or algorithm for estimating game position values under a time constraint. Automatically learn the best evaluation function that runs in less than a given number of machine cycles. Fine tune the search strategy as well.

Finally, there is the psychological Endgame Swindle task mentioned earlier. What is the best move to defeat a human?

All these tasks computers do worse than humans.

No comments :