Deep Learning-Based Energy Efficiency Maximization in Massive MIMO-NOMA Networks With Multiple RISs.

In this paper, we study a deep learning (DL)-based energy efficiency maximization (EEM) problem in massive multiple-input multiple-output (MIMO)-non-orthogonal multiple access (NOMA) networks with multiple reconfigurable intelligent surfaces (RISs). These RISs are deployed randomly at the edge of the near radius to assist the base station (BS) in communicating with near and far users. We formulate the problem of jointly optimizing the precoding matrix and phase shift of the RISs to maximize the overall energy efficiency subject to the quality-of-service of each user, phase shift of RISs, and power budget of the BS constraints. To address this challenging non-convex problem with mixed-integer constraints, the original problem is decoupled into phase shift and beamforming sub-problems, then addressed them separately. We introduce the bisection search algorithm to address the challenge of the phase shifts optimization problem. For the beamforming optimization, it is transformed into an equivalent non-convex problem but with a more tractable form. Then, we propose an iterative algorithm based on the inner approximation method for its solution. To support real-time optimization, we design a deep learning framework to predict optimal solutions of phase shifts at RISs and precoding matrix under different parameter settings. Simulation results show that the proposed DL-based approach can predict the optimal solution with high accuracy in a short time compared to the conventional approach. Additionally, the effect of the maximum power budget at the base station, the number of RISs, and BS's antennas are evaluated thoroughly.