An Adaptive and Lightweight Hybrid Beamforming Design for Massive MIMO

Traditional deep unfolding-based hybrid beamforming designs for Joint Communication and Sensing (JCAS) often suffer from prohibitive computational overheads derived from extensive step-size multiplications and are limited by fixed task weights, which prevent the system from adapting to dynamic channel and sensing environments. To address these challenges, this letter proposes a lightweight intelligent beamforming algorithm comprising two synergistic modules. First, at the algorithmic architecture level, we propose a Multiplication-Free Projected Gradient Ascent (MF-PGA) mechanism that transforms step-size scaling multiplications into hardware-friendly shift-and-add operations based on an Additive Powers-of-Two (APoT) discretization scheme, while preserving the full-precision gradient computation essential for beamforming accuracy. Second, at the optimization strategy level, we propose a meta-learning-based Balancer Neural Network (BalancerNN) that dynamically generates optimal task weights according to real-time channel conditions. Step-size parameter storage requirements are reduced by 87.5%, and all step-size-related multiplications are eliminated during inference. Furthermore, by mitigating the dominance of gradients under high signal-to-noise ratios through dynamic weighting, the method improves sensing accuracy by approximately 30%, achieving a balance between communication and sensing.