Forward Neural Network Learning Theory

First, the paper introduces the sturcture of feedforward neural network and classical back-propagation (BP) learning algorithm, then systematic analysis and reserch are made to the various learning methods of RBFNN(Radical Basis Function Neural Network). The idea and algorithm frame of evolution strategy are presented and introduced to the learning algorithm of BP Neural network to improve the performance of it.There is no universal incremental learning algorithm for all feedfoward neural networks because of not only their varities on the stucture and learning algorithms, but also their dependence on former learning instances. A universal incremental-learning method is presented with regenerating instances to represent all the previous information stored in the net by using the information on the structure and weights. To solve the problem of conflicts between old memory and new instances, the regenerated instances are evaluated before increment learning.Making use of conformity between tne structure of RBFNN and the idea of pattern recognition, a new kind of RBF network is also presented with introduing the structure and the relevant learning algorithm of it systematically. The theoretical analysis and prove of relevant algorithms and conclusions is also provided.In the end, the simulation demonstrates the good result of the learning algorithm and the new network structure.