Sparsely Gated Mixture of Experts Neural Network For Linearization of RF Power Amplifiers

This article presents a piecewise neural network (NN) with dynamic sparsity for modeling and linearization of radio frequency (RF) power amplifiers (PAs). A mixture of experts NN (MENN) approach is employed to combine several smaller real-valued time-delay NNs (RVTDNNs) by means of a gating NN. Furthermore, we complement the MENN framework with top- $K$ sparse gating, such that only a subset of experts is activated during each sample inference, reducing the computational complexity at run time. An end-to-end training approach is presented, to optimize the gating alongside with specializing the expert NNs, enabling the experts to collaborate. We experimentally investigate the scaleability of the proposed model in terms of modeling accuracy and linearization performance, as well as run time and model complexity, using RF measurements with two different gallium-nitride Doherty PAs at 1.8 and 3.5 GHz, respectively. Our experiments confirm a significant reduction in run-time complexity due to the sparse gating, with only a small penalty on accuracy, linearization capability and scaleability. Furthermore, the proposed approach is shown to offer favorable complexity-performance trade-offs, outperforming the existing state-of-the-art.