Graph Neural Network-Based Reinforcement Learning for Cooperative Network Localization

We propose a Graph Transformer-based Reinforcement Learning Network Localization (GTRL) framework, which bridges graph representation learning with multi-agent decision making for cooperative Sensor Network Localization (SNL). Unlike conventional methods, GTRL casts localization as a cooperative policy learning problem, where a Graph Transformer (GT) encoder extracts spatial dependencies from graph-structured observations, and Multi-Agent Proximal Policy Optimization (MAPPO) learns per-node policies to refine positions. This paradigm shift enables agents to adaptively aggregate information and make distributed decisions, yielding robustness to noise, scalability to large networks, and efficiency under limited communication. Extensive experiments show that GTRL converges significantly faster than traditional methods, achieving higher accuracy with far fewer iterations. In addition, it generalizes well across diverse network layouts, anchor densities, communication ranges, and noise levels, outperforming state-of-the-art Graph Neural Network (GNN) localization methods.