It would be hopeless to use a simple parametric (e.g. linear or
quadratic) family of regression functions 
in order to treat the
wide variety of voltage security problems, which are highly
non-linear. Thus we turn to non-parametric regression techniques able
to cope with a broad range of non-linearities. According to our
experience, regression trees and multilayer perceptrons
are two efficient and complementary such approaches [9].
Below we describe the former and merely summarize the advantages and
drawbacks of the latter better known approach.