Thanks for your interesting contributions. I have a few points in response:
On multiple processors: point taken. This is on our list for future
versions. So far we have been focusing on building various kinds of
functionality but postponing the speed issue. It sounds like you are
making some progress by sampling your data -- which is good. I might
add that you can actually strengthen the models some by using
stratified random sampling to make emphasis more uniform across your
predictor space, as opposed to wasting a bunch of time on the
oversampled parts of the space. You have probably discovered this.
On the autocorrelation issue: Because NPMR is using a local model,
you can already model the autocorrelation with NPMR by just including
coordinates as predictors -- you don't need to calculated a distance
matrix with a local model. You can visualize this easily by fitting a
response variable to x and y as predictors, then looking at a contour
map of the surface. That surface is built on the autocorrelation --
if it wasn't there, the surface would just be flat or irregularly
bumpy. The difficulty as I see it is that many of the interesting
drivers are themselves autocorrelated, which means you risk throwing
the baby out with the bath water.
I believe that this built-in modeling of autocorrelation is a feature
of NPMR that hasn't been explored in the literature. It would be
interesting to compare different modeling strategies for dealing with
autocorrelation wtih NPMR. Please post a link to the pdf here if you
or someone else does this!
-Bruce