3D Adaptive Beamforming Approach with a Fine-Tuned Deep Neural Network

Adaptive beamforming is an essential smart antenna operation, with increasing importance for beyond 5G networks. However, the increased computational complexity of deterministic beamforming algorithms highlights their inability to respond to future demands. In this paper we approach the problem of mapping the angles of arrival of incoming signals on an 8x8 uniform planar array, to a desired complex feeding weight vector. We investigate a well-known deep learning model, the multi-layer perceptron, and we train it on a large dataset produced by the null-steering beamforming (NSB) algorithm. In order to propose the most efficient design, we fine tune the hyperparameters and each individual layer of the deep feedforward neural network architecture. The proposed model is able to reach excellent beam-steering accuracy, with signal to interference-plus-noise ratio levels similar to that of NSB, at a significantly lower response time. Finally, we provide a statistical analysis to compare the performance of each beamformer.