Graph-Neural Multi-Agent Coordination for Distributed Access-Point Selection in Cell-Free Massive MIMO

TL;DR

APS-GNN is introduced, a scalable distributed multi-agent learning framework that decomposes APS into agents operating at the granularity of individual AP-UE connections and achieves one to two orders of magnitude lower inference latency than centralized MARL approaches due to its fully parallel and distributed execution.

Abstract

Cell-free massive MIMO (CFmMIMO) systems require scalable and reliable distributed coordination mechanisms to operate under stringent communication and latency constraints. A central challenge is the Access Point Selection (APS) problem, which seeks to determine the subset of serving Access Points (APs) for each User Equipment (UE) that can satisfy UEs' Spectral Efficiency (SE) requirements while minimizing network power consumption. We introduce APS-GNN, a scalable distributed multi-agent learning framework that decomposes APS into agents operating at the granularity of individual AP-UE connections. Agents coordinate via local observation exchange over a novel Graph Neural Network (GNN) architecture and share parameters to reuse their knowledge and experience. APS-GNN adopts a constrained reinforcement learning approach to provide agents with explicit observability of APS' conflicting objectives, treating SE satisfaction as a cost and power reduction as a reward. Both signals are defined locally, facilitating effective credit assignment and scalable coordination in large networks. To further improve training stability and exploration efficiency, the policy is initialized via supervised imitation learning from a heuristic APS baseline. We develop a realistic CFmMIMO simulator and demonstrate that APS-GNN delivers the target SE while activating 50-70% fewer APs than heuristic and centralized Multi-agent Reinforcement Learning (MARL) baselines in different evaluation scenarios. Moreover, APS-GNN achieves one to two orders of magnitude lower inference latency than centralized MARL approaches due to its fully parallel and distributed execution. These results establish APS-GNN as a practical and scalable solution for APS in large-scale CFmMIMO networks.

Figures (10)

• Figure 1: Message-passing structure among agents: yellow dashed arrows for agents that share the same UE states and blue dashed arrows for agents that share the same AP.
• Figure 2: Local reward structure for each agent. Each agent associated with an AP-UE pair receives a local reward composed of two signals: the AP activation signal, and the UE SE satisfaction signal.
• Figure 3: Architecture of APS-GNN. Each agent encodes its local temporal channel history using a GRU, which is concatenated with the most recent channel magnitude and projected into a $64$-dimensional embedding. The resulting node features participate in two rounds of message passing using Attention-based Graph Convolution layers over same-UE and same-AP edges, enabling spatial coordination among neighboring agents. After concatenation, ReLU, and LayerNorm operations, the final $384$-dimensional embedding is linearly projected to a $2$-dimensional output representing the binary AP-UE connection decision.
• Figure 4: Spatial distribution of APs and UEs within the service area. The plot illustrates UE and AP locations, along with the strongest and second-strongest AP-UE associations.
• Figure 5: Performance comparison of the algorithms in minimizing the number of activated APs while meeting the target SE across different medium-scale scenarios with $20$ APs and $6$ UEs. Error bars indicate the standard deviation.