quiescence search helps avoid the horizon effect in computer game playing. a well-known way of doing quiescence search in chess is to search a subtree considering only captures, king checks and responding to check. a transposition table is important to detect perpetual check and repetition, and also as an optimization to tree search. how much computation does quiescence search require? will the extra search terminate in a reasonable amount of time?
this quiescence search seems to elegantly pair well with a simple evaluation function which considers only material and checkmate.
searching less: omit pawn captures. or, omit pawn captures unless the captured pawn is close to promotion. close to promotion = passed pawn? also search moves which create a passed pawn (i think this has to be a capture, though it could be a capture of a pawn). the evaluation function should also take into account passed pawns.
searching more: also keep searching if a threat is made to capture the second most valuable piece (after the king) still left on the board. and responding to that threat. pawn promotion. advancement of a passed pawn. advancement of any pawn.
in general: search moves that significantly change static evaluation in one move, where "significant" could be absolute or relative.
of course, to beat such a system, steer the game toward positions which have a forced sequence which includes one quiet move that quiescence search does not see.
No comments :
Post a Comment