Meta-Learning For 6G Communication Networks With Reconfigurable Intelligent Surfaces

Channel acquisition is one of the main challenges in a reconfigurable intelligent surface (RIS) system due to the passive nature of an RIS. In order to accurately estimate RIS channels, a large number of pilot symbols are required, which could yield a severe performance degradation in terms of the spectral efficiency (SE). In this paper, a practical channel acquisition and passive beamforming technique is proposed using a limited number of pilot symbols in an RIS-assisted cellular network. In particular, the proposed technique relies on a meta-learning framework. To address practical RIS challenges, the problem of maximizing the instantaneous SE is formulated and a novel approach to solve this optimization problem is developed. The proposed algorithm enables an RIS to select the optimal phase shift matrix without the need for perfect channel state information. Also, the trained parameter resulting from the proposed meta-learning algorithm can be guaranteed to converge to an optimal solution. Simulation results show a comparable performance to an exhaustive search method with a few training symbols which validates the advantages of meta-learning for an RIS system.