New Ontogenic Neural Network Classificator Based on Probabilistic Reasoning
Adrian Horzyk · 2003
Many training and configuration methods of today suffer from imprecisely defined training parameters and difficult configuration manners. The determination of a neural network topology and all training parameters is sometimes very difficult and only experts can take advantageous of them. This paper introduces a new ontogenic method for fully automatic configuration of neural network topology and all parameters of thus network. The method enables us to create a universal neural network classificator fully automatically. The presented method uses some probabilistic estimations of training data to create minimal 3-layer partially connected neural network topology that fully satisfies training data requirements. Optionally, this method can automatically carry out some reductions of minor features of training data without loosing correct classification and good generalization properties of the neural network. Finally, the described configuration method is very fast, user-independent, efficient and economical in comparison to other methods of today.