New Machine Learning Approach for Low Overhead Multi-Beam Prediction

This paper investigates the problem of initial beam alignment that is crucial to beam-based communications used in 5G and envisioned for 6G. Our proposed method includes a beam sweeping using a learned codebook, and a beam prediction by a categorization neural network (NN) based on the measurements from the sweeping. Different from the straightforward supervised learning for the optimal beam index, we first design a compact representation of the optimal beam to reduce the complexity of the NN. Then, we show that learning the single optimal beam index can be problematic for realistic scenarios, and propose to use a new probabilistic spatial power distribution as the output. We show that our method is better suited for future wireless generations and can provide various information related to beam management by simulations.