Enhancing Spectral Efficiency of Short-Packet Communications in STAR-RIS-Assisted SWIPT MIMO-NOMA Systems With Deep Learning

This paper proposes an adaptive user grouping (AUG) scheme for short-packet communication (SPC) in simultaneous transmitting and reflecting (STAR)-reconfigurable intelligent surface (RIS)-assisted multiple-input multiple-output (MIMO)-non-orthogonal multiple access (NOMA) systems with SWIPT. The information users with different channel conditions are optimally grouped while the energy user harvests the energy from the base station. Besides that, multiple STAR-RISs are deployed to assist the information users in improving the quality of received signals. We formulate the spectral efficiency (SE) maximization of the considered system to optimize the linear precoding matrix, phase shift of the reflection and transmission at STAR-RIS, energy beamforming matrix, and grouping variables. The formulated problem leads to a mixed binary integer programming which is challenging to solve optimally. To tackle this problem, we first relax the integer variable to be continuous and decouple the relaxed problem into two subproblems to alternately tackle the phase shift and beamforming parts. We then propose bisection search and low-complexity iterative algorithms to solve the phase shift and beamforming subproblems with guaranteed convergence at a relative optimum of each subproblem. Towards real-time optimization, we develop a convolutional neural network (CNN) to achieve the optimal solution of the relaxed problem via a quick-inference process. Numerical results demonstrate a SE improvement of 46% in the AUG scheme over the random user grouping one and 78% over the non-user grouping under various settings. Furthermore, the developed CNN model predicts optimal phase shift variables and beamforming matrices with high accuracy compared to conventional methods but in a shorter time.