An incremental network construction algorithm for approximating discontinuous functions

Traditional neural network training techniques do not work well on problems with many discontinuities, such as those that arise in multicomputer communication cost modeling. We develop a new algorithm to solve this problem. This algorithm incrementally adds modules to the network, successively expanding the 'window' in the data space where the current module works well. The need for a new module is automatically recognized by the system. This algorithm performs very well on problems with many discontinuities, and requires fewer computations than traditional backpropagation.<>