Near-Field Beam Training for XL-MIMO Systems: A Multi-Armed Bandit-Based Approach

In this paper, we propose a multi-armed bandit (MAB) based two-stage beam training scheme for extremely large-scale multiple-input-multiple-output. The proposed scheme leverages MAB to extract statistical channel state information (CSI) from historical beam training data, thereby reducing the beam training overhead. Specifically, in the first stage, we model the problem of angular sweeping as a MAB problem, and adopt a UCB strategy to choose the discrete fourier transform (DFT) vectors with large UCB values, where the UCB value includes the average beam gain and uncertainty terms, which can better find the DFT vectors with large beam gain. In the second stage, using the angular information in the first stage to get the candidate polar domain codewords, the polar domain sweeping is also modeled as a MAB problem. Then a UCB strategy is also utilized to find the codewords with large UCB values. These two UCB-based algorithms are proven to have logarithmic regret, which means that these algorithms converge to codewords with the largest average beam gain. Simulations validate the good performance of the proposed beam training scheme.