[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: computer-go: A problem with understanding lookahead
From: Dave Dyer <ddyer@xxxxxxxxxxxxxxxxx>
>
>I don't think it's "dead simple", but I don't see that it is a whole
>lot more complicated than chess either.
I don't agree at all. Evaluation in Go is qualitatively much harder,
at least given the existing theoretical framework in which we understand
the game.
In Chess, a piece is either on the board or off the board. A square
is either covered or uncovered. A pawn is passed or not. There are
many such easily computed metrics which are typically used in evaluation.
In go, there are no such metrics.
There are easily measured metrics in both games but it turns out they
are not very useful. (But they are not worthless either)
Yes, a piece is on board right now, but will it be in 2 moves? A pawn
is passed, but there is no easy way to tell if that is even an
advantage. If you are about to get checkmated, my passed pawn is like
a group of dead stones. My trapped piece is on the board right now,
but it might as well be dead. I don't see how Go is any different.
I agree with one thing however. When I say that it is "not a whole
lot more complicated" I mean this in a conceptual way. It does take
much more compute power to do a good full board evaluation (no
searching allowed) in Go than in chess. With no searching allowed,
you can't even do a good 1 ply search in either game I'm afraid.
I will argue forever that a good chess evaluation is no better than a
good Go evaluation, but no one is likely to believe this because chess
programs play so well. It's like no one bothers to factor in that the
search might have at least a tiny effect on the quality of the move
chosen.
Start with a master strength chess program, subtract about 200 elo
rating points for every ply of search it does until you get to a 1 ply
search, and see what you end up with. A program weaker that your Go
program by far.
Don