arXiv:2507.11064v2 Announce Type: replace-cross
Abstract: Reducing feedback overhead in beyond 5G networks is a critical challenge, as the growing number of antennas in modern massive MIMO systems substantially increases the channel state information (CSI) feedback demand in frequency division duplex (FDD) systems. To address this, extensive research has focused on CSI compression and prediction, with neural network-based approaches gaining momentum and being considered for integration into the 3GPP 5G-Advanced standards. While deep learning has been effectively applied to CSI-limited beamforming and handover optimization, reference signal allocation under such constraints remains surprisingly underexplored. To fill this gap, we introduce the concept of channel prediction-based reference signal allocation (CPRS), which jointly optimizes channel prediction and DM-RS allocation to improve data throughput without requiring CSI feedback. We further propose a standards-compliant ViViT/CNN-based architecture that implements CPRS by treating evolving CSI matrices as sequential image-like data, enabling efficient and adaptive transmission in dynamic environments. Simulation results using ray-tracing channel data generated in NVIDIA Sionna validate the proposed method, showing up to 36.60% throughput improvement over benchmark strategies. Read More
