Mixture of Experts Neural Network for Behavioral Modeling and Linearization of RF Power Amplifiers

This article presents a novel neural network (NN) structure for modeling and linearization of RF power amplifiers, with particular aim on improved flexibility and ability to scale. NN-based models have superior nonlinear modeling capabilities as compared with polynomial models making them an attractive alternative for digital predistortion. However, when it comes to scaling the models to larger sizes and greater accuracies, a single large NN may become bulky to execute due to the many dependent operations involved. The proposed Mixture of Experts model combines several smaller time-delay NNs by means of a combining gate NN. An end-to-end training approach optimizes the gating alongside with specializing the expert NNs, enabling the experts to collaborate. The proposed solution is compared with various piece-wise polynomial models regarding their ability to scale in terms of modeling accuracy, linearization performance using measurements with a gallium nitride (GaN) Doherty PA and a GaN LMBA PA. The MENN is shown to achieve similar modeling accuracy as a single large NN, while converging faster and utilizing fewer hidden layers. In addition, we demonstrate the flexibility of the MENN to adopt different configurations without compromising performance. This makes MENN well suited to be mapped flexibly to corresponding NN processing hardware, as described in detail in this article.