State-Driven Adaptive Deep-Unfolded PGA Algorithm for Hybrid Beamforming in MIMO-JCAS Systems

In massive multiple-input multiple-output (MIMO) joint communication and sensing (JCAS) systems, hybrid beamforming (HBF) has attracted much attention because it can provide a favorable tradeoff between beamforming gain and hardware cost. However, HBF design in MIMO-JCAS systems is highly challenging. The main reasons are the strong coupling between the analog and digital precoders in joint communication-sensing optimization and the high-dimensional search space caused by large-scale antenna arrays. In this paper, a state-driven adaptive deep-unfolded hybrid beamforming algorithm is proposed for MIMO-JCAS systems. Specifically, the analog precoder update is redesigned in a manifold-based form to better match the geometry of the constant-modulus constraint, while the digital precoder update is enhanced by a learnable gradient-balancing mechanism to alleviate the dynamic imbalance between the communication-rate gradient and the sensing-error gradient. Furthermore, a lightweight state-driven control network is introduced to generate scaling factors for the hyperparameters according to the current iteration state, so that the unfolded model can adapt its update behavior during optimization. Different from conventional deep-unfolded methods with static hyperparameters during inference, the proposed method provides a more effective optimization strategy for the dynamic communication-sensing tradeoff in MIMO-JCAS hybrid beamforming. Simulation results demonstrate the effectiveness of the proposed state-driven adaptive deep-unfolded method. Compared with the conventional deep-unfolded projected gradient ascent (PGA) algorithm with 20 inner iterations, the proposed method improves the joint objective, while achieving faster convergence and stronger robustness.