The Improvement Of BP Network Generalization Ability

The generalization ability of BP network is the ability to adapt to the new sampleafter the BP neural network’s learning.As the inherent training algorithm of BP neuralnetwork led to the unsatisfactory of network generalization capability.How toimprove the generalization of the BP network is a problem worthy of study. Previousstudies have shown several major factors that impact the BP network generalizationability.They are the quantity and quality of the training samples, the networkparameters and the network’s internal structure. From the above three points, thispaper using different improvements of them, and improving the generalization abilityof the network. Specificly, for a given multi-dimensional samples,there are often havesome associations between the sample variables.So it can combine the geneticalgorithm,use the global search ability and progressive optimization features of it,and constantly look for the combination of optimized variables. Thus it improves thegeneralization ability of the network. For the initial weights and threshold ofnetwork,it also using genetic algorithms to operate,furtherly improving thegeneralization ability of the network. Finally,it is the internal restructuring of the BPnetwork to think. In general, the output of the network layer and input layer is fixed,and the number of hidden layer nodes has been a hot issue. The article first parameterfitting on a set of data, attaining the fitting formula of a BP neural network hiddenlayer nodes.Then using this formula to calculate the hidden layer nodes and trainingthe BP neural network. While the calculated number of hidden nodes is alwayshigh,here it combining with SVD to decomposite the hidden layer output matrix andmake dimensionality reduction. Within the permissible error range, it removing thesmaller eigenvalue of hidden nodes to, thus completing the reduction of hidden layernodes to improve the network generalization ability.