Application Research Of Neural Network For Data Classification

It is a trend at present to use neural network in pattern recognition. This paper is aimed at the classification research of vising artificial neural network(ANN) in data pattern and testing the validity of adapting neural networks to model behavior of the compound components of building materials pattern recognition. According to the character of neural network, an improved neural network model is constructed, and its performance is analyzed and proved.Because of the data classification's complexity, the methods such as statistical pattern recognition, structural pattern recognition and neural networks have some disadvantage in the classification. In this paper, an improved neural network is used in the classification. First, based on RBF (Radial Basis Function) neural network, an improved four-layer feed-forward neural network is presented. The input and output is changed, so the parameter is decreased. During the learning procedure, the training ways are used differently according to different characteristics of the patterns. Then taking account the specific feature of classification problem, a new training algorithm is proposed. In this algorithm, a suitable error function named regional mapping error function, which depends only on the misclassified training patterns, is defined, and the advantage of clustering algorithm and RAN (Resource Allocating Network) is also combined in this algorithm, so the network parameters is adaptively adjusted during training process. The validity of this network is illustrated using an example taken from the pattern of a synthetic two-class problem and the component analysis of CaO-Al₂O₃-SiO₂ system. Simulation shows that this approach can be successfully used in the pattern recognition and can get good generalization ability.