Iteration Process Analysis of Real-Valued Time-Delay Neural Network with Different Activation Functions for Power Amplifier Behavioral Modeling

The dynamic iteration process analysis of real-valued time-delay neural network (RVTDNN) for power amplifier (PA) behavioral modeling is provided in this paper. The performance with three activation functions of hyperbolic tangent sigmoid (tansig), logistic sigmoid (logsig) and rectified linear unit (ReLU) are investigated. It is discovered that RVTDNN with ReLU has the fastest convergence speed, and the ones with tansig and logsig have similar modeling accuracy which is higher than ReLU, but slower convergence speed, validated by experiments on a Doherty PA. The iteration process investigation of RVTDNN with different activation functions provides practical considerations for applying the model into different future 5G scenarios.