Neural Network Enhanced Analog Beam Selection Scheme for Terahertz Systems

With the rapidly increasing demand for communication, terahertz wave communication has gradually stands out with its higher rate, lower power consumption, and secure communication. Furthermore, to reduce the use of radio frequency chains, hybrid beamforming for MIMO system is proposed. In conventional method, in order to optimal the uplink sum rate, exhaustive search algorithms are commonly used to select the best codeword for analog beamforming. However, exhaustive search algorithms also cause too much complexity to be implemented in engineering. In this paper, an iterative sub-optimal algorithm is firstly proposed to avoid the computation of matrix inversion. Moreover, we propose a data-driven method based on RBF-NN of analog beam codebook selection to further reduce the complexity. Specifically, with training data coming from samples of the terahertz channel, the analog beam codebook selection problem is considered as a multiclass-classification problem. Using the dataset, we built a statistical classification model via RBF-NN method which can select suitable analog beams for each user, with low complexity and near optimal sum rate. Analysis and simulation results reveal that, compared with the conventional method, as long as the training data are sufficient, the proposed method reduce complexity by several orders, with near-optimal performance.