Data Detection and Resource Allocation for Distributed Two-Way Cell-Free Network: A Graph Attention Network-Based Approach

This paper introduces a unified message-passing framework for two-way cell-free relaying networks, achieving paradigm-level integration of uplink data detection and downlink resource allocation through a coherent distributed architecture. Departing from conventional decoupled designs that treat the uplink and downlink as independent problems, our framework considers both communication directions under a unified architectural paradigm, while avoiding the need to explicitly solve a highly complex joint uplink–downlink optimization problem. Specifically, the uplink and downlink are implemented through specialized yet philosophically consistent message-passing approaches, enabling fully parallel execution of data detection and resource allocation via distributed message passing among access points (APs). This design eliminates centralized coordination while maintaining architectural coherence across the network. For the uplink, we propose a distributed Expectation Propagation (EP) detector, enhanced with a gated recurrent unit (GRU), termed DEP-GRU, to perform iterative Bayesian inference and local refinement at each AP. This effectively improves detection accuracy while maintaining low inter-AP communication overhead. For the downlink, we introduce a graph attention network (GAT)-based resource allocation strategy to jointly optimize AP selection, beamforming, and power control. By modeling each beam as a graph node, the GAT captures local interactions and efficiently allocates resources in a fully distributed manner. To avoid the need for labeled data, we employ an unsupervised, physics-aware training approach, which directly enforces system power constraints and fairness requirements during training. Simulation results demonstrate that the proposed joint design significantly improves the symbol error rate and achievable throughput, validating its effectiveness in a distributed two-way cell-free relaying network.