Behavioral Modeling of Power Amplifiers with Modern Machine Learning Techniques

Abstract

In this study, modern machine learning (ML) methods are proposed to predict the dynamic non-linear behavior of wideband RF power amplifiers (PAs). Neural networks, k-nearest neighbor, and several tree-based ML algorithms are first adapted to handle complex-valued signals and then applied to the PA modeling problem. Their modeling performance is evaluated with measured data from two base station PAs. Gradient boosting is seen to outperform the other ML approaches and to give comparable performance to the generalized memory polynomial (GMP) reference model in terms of both the normalized mean squared error (NMSE) and adjacent channel error power ratio (ACEPR). This is the first study in the open literature to consider modern ML approaches, besides neural networks, for PA behavioral modeling.