Neural Networks for Phase Shift Optimization of Reconfigurable Intelligent Surfaces Under Imperfect Channel State Information

Reconfigurable intelligent surface (RIS) is a key enabling technology for the sixth generation (6G) of mobile networks. It can focus the signal at an intended location (e.g., a user hotspot) through dynamically adjusting the phase shifts of its passive reflecting elements, thereby enhancing the signal quality and network coverage. However, the optimal configuration of the phase shift profile of RIS is challenging since it requires accurate channel state information (CSI), which is prohibitively expensive to acquire in practice because the number of reflecting elements in RIS is usually large. To address this limitation, in this paper, we train and test a fully-connected neural network (FCN) that estimates the optimal phase shift profile of RIS from noisy CSI measurements. We evaluate the performance of the proposed Machine Learning (ML) model in terms of different key performance indicators (KPIs), including the system bit error rate (BER) and throughput, phase shift estimation mean square error (MSE), and the training time of the neural network itself. Simulation results demonstrate that our proposed technique can significantly improve the performance in RIS-assisted wireless networks, reducing the gap to the optimal network throughput to below 1 %.