A Fractal Radial Basis Function Neural Net for Modelling

A new recursively self-constructing fractal learning algorithm for an artificial neural network is presented. The network architecture is based on a tree-like structure of Radial Basis Function (RBF) networks. This paper examines some of the problems with RBF nets, and suggests a possible solution. Neurons in RBF nets have local receptive fields. The new learning algorithm ‘grows’ new sub-networks within the fields of existing units producing modelling errors greater than the tolerated limit. The network complexity is thus matched to the complexity of the system being modelled. The local nature of the neurons allows the modelling tolerance to be stored locally, enabling the network to give a varying confidence estimate for outputs from various areas of the input space. The learning algorithm continually undergoes a train–test–train procedure. This prevents overtraining and can be used to create representative training data automatically.