Modifying Upstart for Use in Multiclass Numerical Domains

One of the research topics pursued by scientists specializing in artificial neural networks deals with the question of how to determine a neural network’s architecture. In the work reported here, we resurrect Frean’s Upstart (1990) that grows the network one neuron at a time. This algorithm is known to have some useful properties; however, it was originally developed only for applications with two classes and with training examples described by boolean attributes. To extend the usefulness of Upstart, we suggest modifications that facilitate the use of this paradigm in multiclass domains with numeric examples.