Consider a chess game between a Test program and a Reference program in which Test loses. Exhaustively explore all possible alternate moves whenever Test was to move (and then play out the rest of the game) (explore only one ply away from the original (lost) game) to find the last possible move which would have changed the outcome of the game (to draw or win for Test). It is possible no such move exists (if Reference is way too strong, but we are mostly interested of when Test and Reference are simply two different instances of the same program).
Next, use machine learning to more quickly find the last possible "saving" move, using features such as the evaluation score and principal variation.
No comments :
Post a Comment