Full adaptive neural predistortion for satellite non linear high power amplifiers

This paper proposes a novel on-line method to predistort the nonlinear high power amplifier (HPA) on-board the satellite by neural networks (NN). The proposed method consists of two NNs: a NN to model the HPA on-line and another NN to pre-distort it. The updating rules of the predistortion NN depend on the instantaneous parameters of the modeling NN. The ordinary gradient descent algorithm suffers from a long transient phase and a slow convergence speed since the two NNs evolves si-multaneously. Using the natural gradient descent exhibits good convergence speed and more accurate model and pre-distorter in a short time.