On Mar 13, 2007, at 23:27 , Peter Drake wrote:
Hmmm -- p. 735 of Russell & Norvig's AI text contains a strong argument
that "nearest-neighbor methods cannot be trusted for high- dimensional
data".
AFAIK that's because when you add more and more dimensions and try to
calculate the distance
between two data points you get less and less different values the higher
the number of
dimensions. You have so many dimensions that have the same distances
(between different sets
of data points I mean) that the few that are different get lost. Or
that's at least that's what
I've been told.
Urban
- --
http://bettong.net
Explanation makes sense. My comment is maybe out of context, sorry about
that but just ask this anyway. In other context? EA search, I had "idea" of
using dynamic distance functions help local optimum problem. Kind of
accepting fact that don't know how distance should be calculated. For fixing
that distance function changes in search time dimension, and weights
different dimensions differently, so algorithm has even possibility to
advance from local optimum in longer search run and so on don't converge
local optimum. I don't know/remember any studies related dynamic distance(in
my case dynamic fitness) function but kind of intrested if there is as it is
kind of simple idea. It is kind of biologically realistic to think that
enviroment changes and values different attributes differently depending
situation.
This allows such behaviour that small change in one dimension can cause big
difference between data-points without knowing anything about data-point
semantics.
t. Harri
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
