Title:Joint Communications and Sensing Hybrid Beamforming Design via Deep Unfolding

Abstract:Joint communications and sensing (JCAS) is envisioned as a key feature in future wireless communications networks. In massive MIMO-JCAS systems, hybrid beamforming (HBF) is typically employed to achieve satisfactory beamforming gains with reasonable hardware cost and power consumption. Due to the coupling of the analog and digital precoders in HBF and the dual objective in JCAS, JCAS-HBF design problems are very challenging and usually require highly complex algorithms. In this paper, we propose a fast HBF design for JCAS based on deep unfolding to optimize a tradeoff between the communications rate and sensing accuracy. We first derive closed-form expressions for the gradients of the communications and sensing objectives with respect to the precoders and demonstrate that the magnitudes of the gradients pertaining to the analog precoder are typically smaller than those associated with the digital precoder. Based on this observation, we propose a modified projected gradient ascent (PGA) method with significantly improved convergence. We then develop a deep unfolded PGA scheme that efficiently optimizes the communications-sensing performance tradeoff with fast convergence thanks to the well-trained hyperparameters. In doing so, we preserve the interpretability and flexibility of the optimizer while leveraging data to improve performance. Finally, our simulations demonstrate the potential of the proposed deep unfolded method, which achieves up to 33.5% higher communications sum rate and 2.5 dB lower beampattern error compared with the conventional design based on successive convex approximation and Riemannian manifold optimization. Furthermore, it attains up to a 65% reduction in run time and computational complexity with respect to the PGA procedure without unfolding.