Predictive Beamforming With Active Inference Under Continuous Time-Varying Channels

In dynamic communication environments, where channel conditions change continuously over time, beamforming employing massive multiple-input multiple-output (MIMO) in the millimeter wave (mmWave) band requires frequent beam re-alignment. To reduce the beam search overhead, hierarchical codebooks have been introduced to narrow down candidate beams efficiently. However, in such environments, a repeated and significant drop in signal-to-noise ratio (SNR) often occurs due to re-searching for the optimal beam, which remains a challenge. To address this issue, we propose a method that combines hierarchical codebooks with active inference to predict the optimal beam rather than search for it. Our approach infers the underlying channel state and selects the beam accordingly, aiming to minimize temporary SNR degradation caused by repeated beam training. Simulations in a scenario involving a mobile terminal moving around a single base station demonstrate that our method achieves a higher average SNR than classical beam training using the divide-and-conquer algorithm. Furthermore, the frequency of sharp SNR drops is reduced. These results show that our proposed predictive beamforming method offers high adaptability and stable communication quality under continuously time-varying channel conditions.