Machine Learning based Analog Beam Selection for Concurrent Transmissions in mmWave Heterogeneous Networks

Abstract

In millimeter-wave (mmWave) heterogeneous networks (HetNets), a variety of mmWave base stations (mBSs) are usually deployed with massive MIMO to form directional analog beams. Each mobile user equipment (MUE) can be served by multiple mBSs simultaneously with concurrent transmissions. However, as the number of mBSs and MUEs increase, it becomes a big challenge for the mBS to quickly and precisely select the analog beams. Thus, this paper propose an machine learning (ML) method to improve the analog beam selection. First, we use stochastic geometry to model the distribution of HetNets, where the probabilities that multiple mBSs serve every MUE are further derived and get the average throughput (AT) for mmWave HetNets. Based on ML, we adopt the support vector machine (SVM) to iteratively select the analog beam, where a promotional sequential minimal optimization (Pro-SMO) algorithm is proposed to train data sets of all the links, where the computational complexity and algorithm convergence are also discussed. Simulation results at last proofed that the proposed ML algorithm not only gets a higher AT than the traditional channel estimation (CE) algorithm, but also achieves a very substantial reduction of calculation complexity.