Based On Lssvr Improved Rbf Neural Network Algorithm And Its Application

The Artificial Neural Networks (ANN) is an information processing system implemented by computer software. It is essencially an abstracted mathematical model for simulation of human brain functions such as memory, association and reflection. Relying on the prefereble adaptive abillity, self-organized capabilities and fault-tolerence abillity, the ANN had a series of outstanding applicable achivements in various sectors and areas such as function approximation or regression analysis, pattern recognition and data classification, clustering, prediction, diagnosis, and process control, with a whole new research approch. However, the anticipation on the higher level of ANN was further inspired by its own success and will be acomplished by improving and enrich its mathmatical model, as the very core and foundational work to be done.To expatiate the work, the rudiments of the two core section of the ANN—Support Vector Machine (SVM) and Radial Basis Function (RBF) were firstly introduced briefly and systematiclly, as well as their current development status. The thoery, structure, arithmetic and application of SVM and RBF were analyzed, with the mathematical deducing process of the main arithmetic provided.One of the focal points of the research aims to elevate the mechanical learning effeciency and adaptive capacity. On the basis of summarizing the current research achivements, this paper presented an improved RBF neural network based on Least Squares Support Vector Regression (LSSVR), which has been also proved to possess better stability and notable function of noise reduction. In order to obtain optimized initiating structure of the network, the LSSVR was chosen to determine the center of network and the number of the hidden nodes. The problems of RBFNN, such as the difficulty in center selection and initiating network structure creation, were sucessfully resolved. The supporting vectors deduced through LSSVR were taken as the center of the network, and their number was regarded as the number of the hidden nodes of the RBF network, of which optimal initiating structure and network parameter was then established. The parameter was finely adjusted by gradinet descent algorithm to endow stronger noise reduction ability to the network, as well as the optimal performance.At the end of this paper the LSSVR-RBF neural network was applied in stock price prediction. Compared to the result of RBF neural network, it has notably lesser forecasting error, demonstrating the significant promotion of the mechanical learning efficiency by the work of this paper.