Comment Re:Interesting computer Chess? (Score 1) 230
A human player of Kasparov's stature may well be able to "keep up" without being able to match the computer's ability to evaluate 2.4 million positions a second. Two things explain the human's ability to equal or outperform the computer.
First, only a tiny fraction of those positions need to be evaluated. All but a small fraction of them result from the human's making one of a large number of foolish next moves -- for example, a pointless king move when the king is under no immediate threat and in a safe position. And at each move up the line all but a small fraction of the moves are again not worth considering. So the computer is wasting time evaluating millions of positions that don't require evaluation.
Second, the quality of the evaluations is at least as important as the quantity of evaluations. Kasparov and his peers (and near peers) can apply prodigious amounts of theoretical knowledge and experience to evaluate the subtleties of a position. Winning evaluation algorithms are far more complex than adding up the standard point values (queen = 10 or 9 1/2, rook = 5, bishop = 3, etc.) for the pieces of each side at a certain position eighteen moves down the line. Pieces must be evaluated according to complex circumstances. Doubled rooks are worth more, and how much more depends on the precise situation. A queen doubled behind a rook is more valuable than a rook behind a queen (the rook can be exchanged, the queen generally can't). When both sides have just one bishop, the bishop's value to each side may depend on whether the bishops are of the same or opposite colors when one side or the other has a material advantage. In late middle game and end game situations, a pair of bishops is worth more than a bishop and a knight or two knights. Knights at the edge of the board or -- worse -- in or beside a corner -- are worth less. Blocked bishops and rooks are worth less. Isolated pawns and doubled pawns are worth less. A piece vulnerable to exchange is worth less to the player with a material disadvantage. Numerous other considerations affect the value of the pieces. The adjustments in value are subtle and extremely hard to make mechanically; they require the judgement that the grandmaster can bring to bear. A computer must put specific values on each piece in each situation, probably using increments of 0.1 point (where pawn = 1.0 at start). This is no easy task.
Beyond revaluation of the pieces according to circumstances, values must be places on positional factors. Does the position yield the initiative to the opponent, who can (say) check your king or attack a piece? How many pieces are aimed at the enemy king? How well defended are the kings? Has the ability to castle been lost? How many king moves, and how many enemy pawn moves, will it take to reach a promotion square that the king must defend? Many other far more subtle positional factors must be evaluated. The possible situations are so numerous that I doubt the ability of any computer algorithm to account for all of the point adjustments that must be made, and to assign realistic values to each variation or complication. But humans know how to do this without resorting to a mechanical algorithm of the sort a computer must use.
First, only a tiny fraction of those positions need to be evaluated. All but a small fraction of them result from the human's making one of a large number of foolish next moves -- for example, a pointless king move when the king is under no immediate threat and in a safe position. And at each move up the line all but a small fraction of the moves are again not worth considering. So the computer is wasting time evaluating millions of positions that don't require evaluation.
Second, the quality of the evaluations is at least as important as the quantity of evaluations. Kasparov and his peers (and near peers) can apply prodigious amounts of theoretical knowledge and experience to evaluate the subtleties of a position. Winning evaluation algorithms are far more complex than adding up the standard point values (queen = 10 or 9 1/2, rook = 5, bishop = 3, etc.) for the pieces of each side at a certain position eighteen moves down the line. Pieces must be evaluated according to complex circumstances. Doubled rooks are worth more, and how much more depends on the precise situation. A queen doubled behind a rook is more valuable than a rook behind a queen (the rook can be exchanged, the queen generally can't). When both sides have just one bishop, the bishop's value to each side may depend on whether the bishops are of the same or opposite colors when one side or the other has a material advantage. In late middle game and end game situations, a pair of bishops is worth more than a bishop and a knight or two knights. Knights at the edge of the board or -- worse -- in or beside a corner -- are worth less. Blocked bishops and rooks are worth less. Isolated pawns and doubled pawns are worth less. A piece vulnerable to exchange is worth less to the player with a material disadvantage. Numerous other considerations affect the value of the pieces. The adjustments in value are subtle and extremely hard to make mechanically; they require the judgement that the grandmaster can bring to bear. A computer must put specific values on each piece in each situation, probably using increments of 0.1 point (where pawn = 1.0 at start). This is no easy task.
Beyond revaluation of the pieces according to circumstances, values must be places on positional factors. Does the position yield the initiative to the opponent, who can (say) check your king or attack a piece? How many pieces are aimed at the enemy king? How well defended are the kings? Has the ability to castle been lost? How many king moves, and how many enemy pawn moves, will it take to reach a promotion square that the king must defend? Many other far more subtle positional factors must be evaluated. The possible situations are so numerous that I doubt the ability of any computer algorithm to account for all of the point adjustments that must be made, and to assign realistic values to each variation or complication. But humans know how to do this without resorting to a mechanical algorithm of the sort a computer must use.
