Model-based Online Millimeter-wave Channel Sensing with Learned Empirical Priors

We consider the problem of adaptive sensing for multi-path channel estimation in millimeter-wave (mmWave) communications system with single RF chain. Current adaptive sensing approaches either focus on estimating the single dominant path or assume apriori knowledge of the number of multi-path components. A key challenge in Bayesian adaptive sensing is the choice of prior which captures the geometric structure of mmWave channels and leads to tractable adaptive sensing methods. In this work, we propose to learn a flexible empirical prior through supervised training that allows for test-time adaptations to the specific channel environment. We realize our framework using a novel recurrent neural network architecture which learns the prior implicitly as part of the recurrent update rule. At test time, our approach alternates between adaptively designing the beamformer to acquire measurement, estimating the channel, and refining the hyperparameters of the learned empirical prior based on history. We demonstrate our approach produces interpretable beamformers and generalizes to channel configurations with fewer multi-path components than it was trained upon.