Beam Training and Tracking with Limited Sampling Sets: Exploiting Environment Priors

Beam training and tracking (BTT) are key technologies for millimeter wave communications. However, since the effectiveness of BTT methods heavily depends on wireless environments, complexity and randomness of practical environments severely limit the application scope of many BTT algorithms and even invalidate them. To tackle this issue, from the perspective of stochastic process (SP), in this paper we propose to model beam directions as a SP and address the problem of BTT via process inference. The benefit of the SP design methodology is that environment priors and uncertainties can be naturally taken into account (e.g., to encode them into SP distribution) to improve prediction efficiencies (e.g., accuracy and robustness). We take the Gaussian process (GP) as an example to elaborate on the design methodology and propose novel learning methods to optimize the prediction models. In particular, beam training subset is optimized based on derived posterior distribution. The GP-based SP methodology enjoys two advantages. First, good performance can be achieved even for small data, which is very appealing in dynamic communication scenarios. Second, in contrast to most BTT algorithms that only predict a single beam, our algorithms output an optimizable beam subset, which enables a flexible tradeoff between training overhead and desired performance. Simulation results show the superiority of our approach.