Study On The Application Of RBF Neural Network Learning Algorithm For Pattern Classification

Due to high complexity and non-linear among samples in pattern recognition system, the patterns are not easily classified correctly. Neural network becomes a common method used in solving pattern classification systems via network learning, in which the internal rule could be hidden within the architecture and parameters of neural network. RBF(radial basis function) has the advantages of fast learning, not easily getting in the local minimum, so as to more attentions and more applications instead of BP learning algorithm. General RBF neural network is easily to construct, but the excessively long time of learning process and the wasting of network resource is lead because of the fixed structure and high complexity. An improved RBF neural network learning algorithm (MPIRAN, maxerror-pruning-mproved-RAN) is presented first. This algorithm is described as follows: choose the pattern which generates the maximal error during learning process to alter the novelty, and adjust the centers of hidden units using a parameter called similarity when the novelty isn't satisfied, and in order to obtain simpler structure, adopt FPE pruning strategy. Then based on MPIRAN learning algorithm, to avoid the inefficiency of FPE, the MRIRAN (maxerror -RBFLN -improved-RAN) is presented. This algorithm utilizes the RBFLN (radial basis functional link network) structure, increases the influence on output layer from input layer to improve the learning accuracy, and adds the adjustment of both centers and widths of hidden units using similarity. In order to testify the validity, the algorithm is applied on the problem of component classification of building material first, then for testify its generalization, it is applied on the problem of two-dimension XOR classification. The result of simulation shows that the algorithm could realize correct classification of patterns, improve better learning and testing performance while only need smaller network architecture, and provide a potent approach for solving the problem of pattern classification.