Self-organization of architecture by simulated hierarchical adaptive random partitioning

A simulation environment based on the concept of hierarchical random partitioning for simultaneously self-organizing the architecture and connection weights of neural networks to approximate multivariate mappings is presented. The constructed approximation can be modeled as a modular, feedforward neural network with two hidden layers. The proposed environment shows good generalization even for small data sets and computes a confidence index for its predicted output. The simulation environment has a fast, automatic learning process and is based on a sound mathematical foundation