Deep Learning-Based Two-Timescale CSI Feedback for Beamforming Design in RIS-Assisted Communications

This letter proposes a deep learning-based two-timescale channel state information feedback framework called RIS-CsiNet for beamforming (BF) design and discrete phase shift design in reconfigurable intelligent surface (RIS)-assisted frequency-division duplexing systems, where the direct link between the base station (BS) and the user equipment (UE) is blocked. Given that the BS-RIS channel is quasi-static and the RIS-UE channel rapidly changes, the feedback is divided into two types: large-timescale and small-timescale feedback. At the beginning of the large-timescale feedback, the neural network (NN)-based encoder at the UE compresses and feedbacks the BS-RIS channel to the BS. On the basis of the received BS-RIS channel information, an NN-based hypernetwork adopted at the BS generates the weights of the NNs that directly produce the BF vector and RIS phase shifts according to the received feedback information of the RIS-UE channel. In the small-timescale feedback, only the RIS-UE channel is fed back by the encoder. The BF vector and phase shifts are produced by the NNs generated in the large timescale. Simulation results show that RIS-CsiNet considerably outperforms conventional algorithms in terms of achievable rate and complexity.