Study On RBF Neural Network Improvement And Its Applications

RBF neural network has some advantages, such as simple structure, high learning speed, better fitting precision, better generalization ability and converging to global optimization etc. So it is often used in function approach, classification, time sequence prediction and so on. Although RBF neural network is easy to construct, its structure is certain or its algorithm is very complicated. It is a waste of network resources and lowers learning speed, so it is possible to reduce learning time through some improving work. In order to improve the properties of traditional RBF neural network, this paper proposes two modified RBF networks--DLCRBF and HUCRAN.In DLCRBF network, direct linear connections between input layer and output layer are added, and data is clustered with nearest neighbor clustering algorithm which makes the network learn on line. Then, the steepest descendent algorithm is used to modify the net parameters to get to optimization. All these work makes DLCRBF network better than classified RBF networks both in learning speed and generalization ability.Also this paper combines RAN, correlation pruning strategy and PCA data pretreatment method, and puts forward an improved network—HUCRAN. It has the advantages of RAN which are constructing simple structure network, learning on line and inputting data once, and it also has the advantages of correlation pruning strategy which are simplifying network structure and improving generalization ability. HUCRAN can offset respective faults of RAN and correlation pruning algorithm.DLCRBF and HUCRAN are respectively used in ethylene and propylene quality predictions and finance early warning. The simulation results show that the two improved networks have better performances than original networks, and they can be used in modeling engineering objects in which the relationship between secondary variables and main variables is very complicated and highly nonlinear. The simulations also verify that the improvement work in this paper is feasible and practical.