Optimal Beamforming and Power Control for D2D Networks with Arbitrary Device Locations via PSCANet and S-PSCANet

Existing optimization and deep neural network (DNN) approaches for beamforming and power control cannot achieve satisfactory performance and computation time trade-offs for interference networks with arbitrary device locations, due to optimization algorithms' expensive computation costs and DNNs' universal approximation and lack of generality. To address this issue, we propose new beamforming and power control approaches by elegantly leveraging optimization and deep learning techniques. We investigate the maximization of the worst-case achievable rate of multiple transceiver pairs under power constraints in orthogonal frequency-division multiplexing (OFDM)-based wideband device-to-device (D2D) networks with multiple subcarriers and arbitrary device locations. First, we obtain each transceiver's transmit and receive beamforming using the standard singular value decomposition (SVD) method. Then, we propose an iterative algorithm named PSCA to obtain joint power control for all transceivers using the parallel successive convex approximation (PSCA) method. PSCA allows parallel and closed-form per-iteration updates and has a short per-iteration computation time. Next, we propose two PSCA-driven deep unrolling neural networks, namely PSCANet and S-PSCANet to reduce the overall computation time. PSCANet and S-PSCANet marry PSCA's parallel computation mechanism with the parallelizable neural network architecture and effectively optimize PSCA's algorithm parameters based on vast samples of random channel fading coefficients and device locations. Moreover, S-PSCANet successfully resolves the training problem due to vanishing or exploding gradients when unrolling many PSCA iterations. Numerical results demonstrate the superior advantages of PSCANet and S-PSCANet over the existing approaches.