Approximation property of multi-layer neural network (MLNN) and its application in nonlinear simulation

Presents the approximation properties of multilayer neural networks (MLNN) and their application in nonlinear simulation and analysis. The authors give a direct proof of the approximation ability for a single input-output MLNN with hidden units with sigmoid activation functions, and also give the relationship between the best polynomial approximation and the number of MLNN hidden units. Based on the analysis, the authors propose an MLNN model with hybrid sigmoid-Gaussian activation functions. To verify the idea, they present experiments and results of nonlinear simulation and analysis by MLNN for a solid-state power amplifier. These results prove that the proposed method has general application in nonlinear engineering simulations