Re: computer-go: A problem with understanding lookahead

["David Mechner" <mechner@xxxxxxxxxxx>]

It's a very good point. There's no a priori reason why limited depth
lookahead should work, or why in general looking n+1 ply should be any
better than looking n ply. Tom Anantharaman (one of the developers of deep
thought) told me that in chess, people used to be skeptical that there would
be a continued benefit from increasing search depth, but it has turned out
as a matter of empirical fact that playing strength has improved roughly
linearly with search depth throughout the history of chess.

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


> I have a serious problem with lookahead. I don't understand why it works.
> Why should the backed up values of a, say, 7 ply full width alpha-beta
> minimax search be any better than the values my evaluation function gives
> me by simply evaluating the positions at ply 1. I can understand that if
> the end of the game is in sight then lookahead is useful. I can
> understand  that if I use a quiesence measure or some such and only
> evaluate positions when I know that my evaluation function will work best
> then lookahead is useful. But if neither of these is the case then does
> lookahead help. Is there some way in which the backing up of values and
> minimaxing somehow reduces the error in the evaluations.
>
> I really am seriously stuck with this one.