Communication-Efficient Cooperative Localization: A Graph Neural Network Approach

Yinan Zou, Christopher G. Brinton, Vishrant TripathiWiopt 2025 · DOI

TL;DR

This paper introduces VQ-MPNN, a novel graph neural network approach for cooperative localization in wireless networks. It significantly reduces communication overhead by using vector quantization to compress node state information, achieving localization accuracy comparable to existing methods while using an order of magnitude less data.

What Problem Does This Paper Address?

Accurate localization of devices is crucial for many wireless applications, especially in environments where GPS is unavailable (e.g., underwater, dense forests). Cooperative localization leverages inter-node distance measurements and message exchanges to improve accuracy. However, in communication-constrained environments, transmitting the large messages required by traditional methods becomes a bottleneck. Existing approaches struggle to balance low localization error with low communication overhead, particularly in networks with complex, "loopy" topologies. This paper aims to bridge this gap by developing a communication-efficient method that performs well on loopy graphs and drastically reduces the amount of data transmitted.

Methods

The core of the proposed solution is the Vector Quantized Message Passing Neural Network (VQ-MPNN). Here's a breakdown of its key components:

  • Graph Representation: The wireless network is modeled as a directed graph. Nodes represent devices (anchors with known positions and agents with unknown positions), and edges represent distance measurements between them.
  • Feature Encoding: Initial node positions and distance measurements are used as node and edge features, respectively. These are then encoded into node and edge states using Multi-Layer Perceptrons (MLPs).
  • Vector Quantization (VQ) for Compression: This is the key innovation. Instead of transmitting the full, high-dimensional node states, VQ-MPNN uses a shared, trainable codebook. Each node's state is mapped to the closest codeword in this codebook, and only the index of that codeword is transmitted to neighbors. This drastically reduces the message size.
  • Message Passing: Nodes exchange these compressed codeword indices. Upon receiving indices, nodes use a decoder to reconstruct approximate node states. These reconstructed states, along with their own states and edge states, are used to generate messages.
  • Node State Update: Each node integrates incoming messages from its neighbors to update its own node state iteratively over several communication rounds.
  • Position Estimation: After the communication rounds, the final node state is fed into a dedicated estimation network to predict the node's 2D position.
  • End-to-End Training: The entire VQ-MPNN, including the VQ codebook, encoder, decoder, and message passing networks, is trained end-to-end to co-optimize localization accuracy and communication efficiency.

Datasets

The paper does not explicitly mention pre-existing public datasets. Instead, it describes a simulation setup for training and evaluation:

  • Training Data: 3000 network realizations (simulated scenarios) were generated.
  • Validation Data: 300 network realizations were used for validation.
  • Test Data: 300 network realizations were used for final testing.

The simulations involve a 2D static scenario of a 50m x 50m area with 9 anchor nodes and 20 agent nodes. Two types of measurement noise were considered: additive white Gaussian noise (AWGN) and range-dependent Gaussian noise.

Key Insights and Ideas

  • VQ as a Communication Bottleneck Solver: The paper convincingly demonstrates that vector quantization is a highly effective method for compressing messages in cooperative localization, achieving an order-of-magnitude reduction in communication cost without a significant hit to accuracy.
  • End-to-End Learning for Co-Design: By training the entire system end-to-end, VQ-MPNN can jointly optimize localization performance and message compression, a significant advantage over methods that treat these as separate problems.
  • GNNs for Loopy Graphs: The graph neural network architecture inherently handles loopy graph topologies, a common challenge for traditional belief propagation methods, leading to more robust and accurate localization.
  • Generalizability: The trained VQ-MPNN models show good generalizability, performing well even when the number of agents changes from the training configuration. This is crucial for real-world dynamic networks.
  • Trade-off Optimization: VQ-MPNN offers a superior trade-off between localization error (RMSE) and communication cost compared to existing methods, making it highly practical for resource-constrained wireless systems.

Key Results

  • Communication Efficiency: VQ-MPNN achieves a communication cost that is an order of magnitude lower than other advanced methods like MPNN and significantly lower than traditional BP-based approaches (PLBP, SP-ADMM, NBP). For example, in Figure 5(a), VQ-MPNN with K=2^9 or K=2^11 requires significantly fewer communication bits for comparable RMSE.
  • Localization Accuracy: The proposed VQ-MPNN achieves localization errors (RMSE) comparable to or even better than state-of-the-art methods like MPNN and significantly outperforms traditional methods like PLBP and SP-ADMM, especially in loopy graph scenarios.
  • Robustness: The approach demonstrates robustness across different noise models (AWGN and range-dependent noise).
  • Scalability: The VQ-MPNN architecture is scalable and can function even when nodes are added or removed from the network.
  • Training Efficiency: The paper notes that training VQ-MPNN does not require excessive computational resources or time, converging in approximately three hours on a server equipped with an NVIDIA Tesla V100S.

Limitations and Future Work

  • Static Networks and Independent Noise: The current work focuses on static networks and assumes independent noise for distance measurements.
  • Future Work: The authors suggest extending the approach to mobile networks and considering correlated noise models. They believe data-driven methods like VQ-MPNN can significantly outperform traditional approaches in these more complex scenarios, which are difficult to model analytically.

Summary updated

BibTeX
@inproceedings{zou2025communication,
  author = {Yinan Zou and Christopher G. Brinton and Vishrant Tripathi},
  title = {{Communication-Efficient Cooperative Localization: A Graph Neural Network Approach}},
  booktitle = {Wiopt},
  year = {2025},
  doi = {10.23919/wiopt66569.2025.11123335},
}
Figure