mmWave Beam Alignment using Hierarchical Codebooks and Successive Subtree Elimination

We propose a best arm identification multi-armed bandit algorithm in the fixed-confidence setting for mmWave beam alignment initial access called \ac{SSE}. The algorithm performance approaches that of state-of-the-art Bayesian algorithms at a fraction of the complexity and without requiring channel state information. The algorithm simultaneously exploits the benefits of hierarchical codebooks and the approximate unimodality of rewards to achieve fast beam steering, in a sense that we precisely define to provide fair comparison with existing algorithms. We derive a closed-form sample complexity, which enables tuning of design parameters. We also perform extensive simulations over slow fading channels to demonstrate the appealing performance versus complexity trade-off struck by the algorithm across a wide range of channel conditions