DNNini-UPSCA: Optimal Beamforming and Power Control for D2D Networks via DNN and UPSCA

Existing optimization and deep learning approaches for resource management optimization cannot achieve satisfactory performance and computation time trade-offs for interference networks with arbitrary device locations. This paper aims to address this issue. Specifically, we investigate an orthogonal frequency-division multiplexing (OFDM)-based wide-band device-to-device (D2D) network with multiple multi-antenna transceiver pairs accessing various subcarriers. We optimize resource management (RM), i.e., beamforming and power control, to maximize the worst-case achievable rate among all transceiver pairs under the beamforming and power constraints. The RM problem is a challenging large-scale non-convex problem. We propose joint and separate designs offering different worst-case achievable rate and computation time trade-offs by solving the RM problem and a resultant power control (PC) problem for separately optimized beamforming using optimization and deep learning techniques. First, we propose two parallel successive convex approximation (PSCA)-based parallel iterative algorithms to obtain stationary points of the RM and PC problems. Both can achieve completely parallel and closed-form per-iteration updates, significantly reducing the computation times. Next, we propose two data and model-driven deep learning methods that optimally select good channel state information (CSI)-adaptive initial points for RM-PSCA and PC-PSCA and effectively unroll the two PSCA-based algorithms into neural networks with the algorithm parameters as tunable parameters. Both further reduce the underlying algorithms’ computation times while achieving appealing worst-case achievable rates based on training over vast samples. Numerical results demonstrate the proposed approaches’ superior advantages over the state of the art.