Zero-Shot Learning for Beam Management in LEO Satellite Communications

Beam management is one of the most challenging issues in low-earth orbit (LEO) satellites, where the antenna direction is dynamic, and the storage, computing, and communication resources are limited. In this work, we develop a zero-shot learning approach to maximize the average data rate of a user by optimizing beam tracking policy in both spatial and temporal domains. In the spatial domain, we develop a graph recurrent neural network (GRNN) with only 60 training parameters to predict the next beam direction. Compared with an existing recurrent neural network with more than 30, 000 parameters, the GRNN can reduce the storage requirement remarkably. In the temporal domain, we design a Twin deep Q-network (DQN) to determine the time to trigger beam tracking. To improve the generalization ability of GRNN and Twin DQN, we apply meta-learning to train them with different antenna directions and test them in unseen scenarios. Simulation results show that our zero-shot learning approach does not need new data samples in unseen scenarios. Thus, it does not introduce extra computing and communication overheads. Additionally, the achievable average data rate is around 10% to 20% higher than different benchmarks.