A Self Growing Binary Tree Neural Network for the Identification of Multiclass Systems

A self growing binary tree neural network is introduced for the on-line identification of multi-class systems. Instead of solving the traditional two class problem with a single neuron a time shrinking threshold logic unit is introduced such that the modified neuron has the capability of partitioning the raw data records into three different regions. Incorporating a Least Mean Squares (LMS) learning algorithm provides the capability of detecting and creating new classes and this allows clustering and partitioning of the raw data records into model classes and yields estimates of the parameters. The models which describe the behaviour of the system at different operating regions can be recovered by inspection of the connection weights of the individual neurons. Optimisation procedures for the on-line estimation are also proposed. Simulation studies are included to illustrate the concepts.