Adaptive Filters Versus Machine Learning Based Beam Tracking Techniques for Millimeter-Wave Wireless Communications Systems

Abstract

This work compares the data-based using (bidirectional long short-term memory (BiLSTM)) and model-based using (extended Kalman filter (EKF) and least mean squares (LMS)) strategies to track the communication beams for millimeter wave (mm-wave) vehicular communications. This work utilizes the Deep MIMO dataset for adaptive filtering and machine learning (ML) as training data for the proposed system. BiLSTM networks are commonly used in beam tracking due to their effectiveness in processing sequential data and capturing temporal dependencies depending on their features like sequential data handling and their ability to process data in both forward and backward directions; they can handle the vanishing gradient problem that often occurs in traditional RNNs, the ability to enhance feature learning, and it can be more robust to noise. This work evaluates the simulation results using the average error and outage probability concerning signal-to-noise ratio (SNR) and the number of broadcast antennas. The results show that the ML performance is higher than that of the EKF and LMS, with the LMS having the lowest performance. Furthermore, the work illustrates the mean squares error (MSE) of the angle of arrival (AoA) using the time index, which yields better performance results over time for ML than the MSE values of the EKF and LMS, which rise due to more mistakes.

Keywords

• Vehicular mm-wave communication
• beam tracking
• adaptive filters
• machine learning
• BiLSTM