Variable-Depth Learning Architecture to Adaptive Multi-User Channel Prediction

On the road to 6G, the growing number of antennas and mobility-related applications emphasize the urgency of addressing the channel aging issue. Traditional channel prediction methods are no longer valid due to simplified assumptions. Deep learning (DL)-based predictors, even with attractive performance improvements, generally lack the utilization of extra correlations and flexible inference with accuracy-efficiency trade-off. In this paper, by representing the topology structure of nearby users (UEs) as a graph, a multi-user channel prediction algorithm is proposed for learnable UE correlation extraction, called LUCE, to improve prediction performance. Instead of simple summation-based fusion, LUCE introduces a predefined geographic embedding and a learnable embedding to offset the position information of the cross-attention (CA) to learn and interact UE features on the graph automatically. Furthermore, a lightweight LUCE is proposed to leverage the homogeneity of CSIs by sharing sub-networks. In addition, a variable-depth learning architecture is proposed, called adaptive LUCE (AdaLUCE), and its number of blocks can be dynamically adjusted for each input by tuning a desired threshold. In particular, AdaLUCE utilizes a hierarchical triple residual architecture to promote its training efficacy and produce predictions at its internal blocks without increasing complexity, and adopts a fitting inspired criterion (FIC) to make real-time decisions for adaptive inference. In total, AdaLUCE is optimized by the weighted sum of multiple loss functions, and a multi-step training scheme is presented. We show that AdaLUCE preserves and releases computing resources for various speeds in sample-wise, improving both prediction accuracy and inference efficiency.