A self organising artificial neural network with problem dependent architecture

Self organizing maps (SOM) based on a regularly spaced array of nodes are well known. The optimum net site for a particular problem is not, in general, clear a priori. Even when a suitable size has been found, it is hard to get a uniform distribution of information across the net, despite the application of conscience mechanisms. As a result the final net contains passenger nodes which perform little or no useful purpose. This paper describes a three dimensional net developed to overcome the deficiencies of the tradition SOM. It evolves a physical node distribution in response to the data that is presented to it and eliminates passenger nodes. Although described in terms of mapping into three dimensions, the technique is in principle applicable to mapping into any number of dimensions.