Meta-Learning-Enhanced GNN Approach for Signal Compression in Uplink Cell-Free MIMO Networks

In this work, we investigate the uplink signal compression problem for fronthaul-constrained cell-free (CF) multiple-input multiple-output (MIMO) networks. Considering the dynamic nature of wireless communications, e.g., the varying number of accessing users and the changing rate requirements, we propose a scalable and adaptive approach by integrating graph neural network (GNN) with meta-learning. In particular, the problem of maximizing the achievable uplink sum rate is formulated by optimizing the beamforming matrix of each access point (AP), while accounting for the requirement of user rate. To tackle this non-convex optimization problem, we develop a joint-Reptile-enhanced GNN approach. Therein, the uplink CF-MIMO network is modeled as a heterogeneous graph with cooperative node features, and the joint-Reptile approach is employed to obtain adaptive and generalizable meta-parameters, enabling rapid adaptation to dynamic CF-MIMO networks. Simulation results demonstrate that the proposed approach exhibits better adaptability compared to conventional initialization benchmarks. Moreover, it can achieve higher sum rate performance across various signal dimensions, while maintaining scalability.