A Modified BP Algorithm With Additional Momentum Adaptive Leariing Rate For Power-amplifier Design

This paper mainly introduces the methods of using BP neural network to design thenonlinear Doherty power amplifier model. It is from two aspects to complete the design ofDoherty power amplifier such as power efficiency and relationship between input power andoutput power. BP neural network has the advantages of simple structure, strong operability,and can approximate any nonlinear continuous function, has an ability to solve the problem ofapproximation of nonlinear objective function effectively and get a wide used in the modelingof the amplifier. The traditional BP neural network model turn a set of samples of theinput/output problem into a nonlinear optimization problem, but as its limits, power efficiencyand power output of the network output cannot meet the requirements of actual poweramplifier in the power amplifier design application. Aiming at the shortage of traditional BPneural network, a BP improved algorithm based on adding momentum term adaptive learningrate adopted to model the nonlinear behavior of Doherty Power Amplifier.In order to build a more accurate model of neural network, it is necessary to select theappropriate parameters for improved BP algorithm, that is to choose the appropriatemomentum factor and learning rate. The big value of learning rate will cause the networkoscillation instability, small value of learning rate can avoid the instability, but increases thetraining time. By repeating training for the network, we found that when the momentumfactor value is0.9can reduce the volatility trend effectively, the method of adaptive learningrate can adjusted learning rate in real time, when the error increases then reduce the learningrate and when the error reduces then increase the learning rate to speed up the networkconvergence time.Training data and testing data are extracted from Single Tone of the DPA. Simulation resultsare presented by the error convergence curves and the results show that the modified BPalgorithm model produce a faster convergence speed. The error of two methods of curvefitting results are compared and the results show that using the modified BP algorithm canget more accurate curve fitting than the traditional BP neural network method. Through aseries of experimental results contrast, the modified BP neural network model can get moreaccurate result of the amplifier design.