SVD-Based Neural Network Structure Optimization And Its Application

Feed-forward networks including BP and RBF networks have the most widespread applications and their generalization ability is the most important network performance. It is found that the generalization ability depends largely on a reasonable network complexity. Therefore, based on the theory of singular value decomposition, the paper proposes new structure optimization strategies for BP and RBF networks. The basic ideas are analyzing the output matrixes of trained networks with SVD theory, removing the hidden units having smaller contribution according to the contribution level principle, and obtaining appropriate scales finally. But because of the differences between BP and RBF networks, the two strategies are slightly different in optimization methods. Further, the optimized networks are used in thermal system modeling and achieve good results. The simulation results show the new strategies are effective in improving neural network generalization abilities.