Fast Meta Learning for Adaptive Beamforming

Abstract

This paper studies the deep learning based adaptive downlink beamforming solution for the signal-to-interference-plus-noise ratio balancing problem. Adaptive beamforming is an important approach to enhance the performance in dynamic wireless environments in which testing channels have different distributions from training channels. We propose an adaptive method to achieve fast adaptation of beamforming based on the principle of meta learning. Specifically, our method first learns an embedding model by training a deep neural network as a transferable feature extractor. In the adaptation stage, it fits a support vector regression model using the extracted features and testing data of the new environment. Simulation results demonstrate that compared to the state of the art meta learning method, our proposed algorithm reduces the complexities in both training and adaptation processes by more than an order of magnitude, while achieving better adaptation performance.