Re: computer-go: A problem with understanding lookahead

["David Mechner" <mechner@xxxxxxxxxxx>]

Forget the combinatorics.

The critical difference between go and chess, as far as evaluation goes, is
that in go dead pieces can stay on the board, even until the end of the
game.

Go is a game of territory. To evaluate, you have to be able to count
territory. Dead stones are enemy territory. To be able to count territory,
you have to know which stones are dead. Therefore to able to evaluate a
position, you have to know which stones are dead.

But knowing which stones are dead is HARD. Not only can't our programs
figure it out quickly enough for practical use in a static evaluation
function, they can't figure it out AT ALL.

There are simple cases where a group is obviously alive or obviously dead,
where fast heuristics might give reasonably accurate answers. But in the
majority of (important) cases, there is no working model of how to go about
solving the problem.

Until serious advances are made in the state of the art of life and death
analysis, talk of a reasonable evaluation function is fantasy, and talk
about using big search in go is a fantasy.

-David Mechner
http://cns.nyu.edu/~mechner/