We report on the use of multilayer networks as 'estimators' of functio
ns, or their application to problems of nonlinear regression. A set of
target functions with adjustable complexity is used. The results of e
mpirical investigation into the extent to which function estimation is
possible, and the relationship between the network parameters, are pr
esented. This could provide the designer of a neural-based system with
some intuition about the design choices to be made.