Learning to Invert the Amplifier Nonlinearity

This paper presents several machine learning methods for digital predistortion (DPD) in the presence of a nonlinear channel, e.g., due to a high power amplifier (HPA). We propose separable parameter estimation for a transmission system in which the HPA is followed by a linear time-invariant (LTI) channel, e.g., arising due to the output multiplexing (OMUX) filter. Our proposal mitigates the problem of exponential growth in the parameter estimation and compensation in alternate methods, e.g., those based on Volterra models or alternate Memory Polynomial based models. We apply the Gauss-Newton (GN) algorithm as well as multi-layered Neural Network and polynomial curve-fit methods for deriving the inverse of Saleh’s model of memoryless nonlinearity. We conduct a detailed simulation study and conclude that the proposed iterative GN nonlinear least-squares approach outperforms alternative techniques in the literature.