A Redirected Learning Architecture for Non-linear Digital Pre-distortion

Abstract

This paper introduces the redirected learning architecture (RLA) for estimating non-linear digital pre-distortion models for non-linear devices such as power amplifiers. Existing architectures can be classified as direct learning architectures (DLA) or indirect learning architectures (ILA). DLAs first learn a model of the device and then determine a pre-distorter by either directly inverting the device model or estimating a pre-inverse of the model. The RLA is similar to a DLA, but rather than learning a model of the device, the RLA uses fixed-point iteration to determine a set of input/output pairs which characterize the device. These pairs are then used to estimate the pre-distorter by redirecting learning from the device to the pre-distorter. The fixed-point iteration is shown to converge under a mild condition. Simulation of a class AB power amplifier reveals improved suppression of harmonic distortion compared to a pth order inverse approach.