The Training Algorithm And Performance Study Of Radial Basis Function Neural Network

Among various neural network models, Radical Basis Function Neural Network (RBFNN) boasts for its application of the principle method, simple structure and a solid mathematical base, so it is widely applied in pattern recognition, function approximation, nonlinear systems model, and timing analysis. This paper specifies the structure and principle of RBFNN and makes an in-depth research on different methods of its algorithm as follows:(1) After categorizing and analyzing the existing algorithms of RBFNN, designing some simulation experiments to several algorithms, the paper works out the basic principles of various algorithms and their advantages respectively, thus laying a foundation for the follow-up.(2) From the perspective of non-linear regression in-depth analysis of generalization theory of RBFNN, the paper makes a theoretical deduction on approximation error and generalization error of radical basic neural network, and then draws some useful conclusions which, under certain circumstances, may be used as a guideline in confirming radical basic neural network.(3) As is known to us, the training process of RBFNN, in fact, conforms to the process of confirming the hidden layer structure and connection weight. With the in-depth research and analysis of the existing algorithms and theories, this paper proposes three improved training algorithms of RBFNN, namely,①The data center selecting algorithm based on Kohonen network and OLS algorithm. The algorithm uses Kohonen competitive network to filter to training sample set, to solve the traditional OLS algorithm to put the overall training sample set as a candidate subset as a result of the process of calculating the amount of orthogonal amazing, algorithm efficiency is very low disadvantage. ②The density method based data center selecting algorithm. The algorithm uses the density method of statistics to classify the training sample set to avoid the flaw that a man-made the specified categories number is too blind to category result.③The data centers selection algorithm based on the improved APC-III algorithm. The algorithm uses dynamic data centers based on the sample distribution, while the original APC-III algorithm uses a unified data center, which was not suitable in unevenly distributed data.Simulation experiments are designed respectively for these three improved algorithms to prove their validity.