A function approximation algorithm using sequential composition

A new method for approximating one dimensional functions is developed based on structural capabilities of multilayer feedforward neural networks. It possesses notable but unproven convergence properties which are examined in a series of examples. It is shown that it outperforms conventional networks for complicated one dimensional problems. An adaptive version of the algorithm whose approximation changes to best use the data available is also presented. Experiments indicate that this method is very stable in the presence of noise. For straightforward function approximation however, other conventional routines generally perform better. Nevertheless, noise stability, adaptability and other properties make the new method useful in context.