Transform data according to PCA

[sigvaldm]

One could have a function or class before using e.g. multivariate local regression to change the axes of the data using PCA, and then scale them to have unit standard deviation along each axis. There should be some rather transparent way to rotate the data back. This could also replace the method that is today built into RBFnet (just scaling each axis without any transformation first).

[sigvaldm]

Another alternative is to try using a bandwidth matrix like in multivariate kernel density estimation.