Stable and Transferable Wireless Resource Allocation Policies Via Manifold Neural Networks

Zhiyang Wang, Luana Ruiz, Mark Eisen, Alejandro Ribeiro

TL;DR

This paper proposes using Manifold Neural Networks (MNNs) to solve resource allocation problems in massive wireless networks. By modeling large-scale networks as continuous manifolds rather than discrete graphs, the authors prove that these neural policies are stable against system noise and can be seamlessly transferred from small training setups to much larger operational networks.

What Problem Does This Paper Address?

In modern wireless systems, allocating resources (like power) to mitigate interference is a high-dimensional, non-convex nightmare. While Graph Neural Networks (GNNs) have shown promise, existing theoretical results often fail to guarantee stability or performance when the network size becomes extremely large.

Wireless environments are also inherently "noisy"—channel measurements fluctuate, and network topologies change. A practical resource allocation policy must be:

  1. Stable: Small changes in channel conditions shouldn't lead to wild swings in the allocation policy.
  2. Transferable: A model trained on a small network (e.g., 50 users) should still work effectively on a massive network (e.g., 1,000 users) without retraining.

Methods

The authors bridge the gap between discrete graphs and continuous spaces by modeling the limit of a very large wireless network as a manifold $\mathcal{M}$.

Key components include:

  • Laplace-Beltrami Operator ($L$): This is the manifold equivalent of the graph Laplacian. it characterizes how information (interference) propagates across the continuous surface of the network.
  • Manifold Neural Networks (MNNs): These are defined as a cascade of manifold convolutional filters and non-linearities. Crucially, these filters depend on the eigenvalues of the Laplace-Beltrami operator.
  • Frequency Difference Threshold (FDT) Filters: To ensure stability, the authors use FDT filters. These filters group "close" eigenvalues together and treat them similarly, preventing small perturbations in the manifold's structure from causing large errors in the network's output.

The researchers mathematically prove (Theorem 1) that MNNs using these filters are stable under absolute perturbations to the manifold operator, providing a theoretical foundation for why these models generalize so well.

Datasets

The study uses simulated wireless network environments rather than static datasets:

  • Setup: Transmitters and receivers are dropped randomly into a 2D space.
  • Channel Model: Includes large-scale pathloss and random fast fading (Rayleigh distribution).
  • Training: A GNN policy is trained using unsupervised learning on a relatively small network of $m=50$ transmitter-receiver pairs.
  • Evaluation: The trained policy is then tested on newly generated networks ranging from 50 to 1,000 pairs to test scalability and transferability.

Key Insights and Ideas

  • The Manifold Limit: Treating a network as a manifold allows researchers to use the tools of continuous geometry to solve problems that are computationally intractable in discrete graph settings as $m \to \infty$.
  • Stability-Discriminability Trade-off: There is a fundamental tension in filter design. Small Lipschitz constants ($B$) and large eigenvalue groupings ($\alpha$) make a model very stable but less "discriminative" (less able to distinguish between different signals). The authors note that non-linearities in the MNN help recover some of this lost discriminative power.
  • Spectral Separation: The use of FDT filters (illustrated in Figure 1) is the "secret sauce." By giving similar responses to eigenvalues that are close together, the network becomes robust to the "jitter" caused by environmental noise.
An α-FDT filter that separates the spectrum of L by grouping eigenvalues that are less than α apart.

Key Results

The numerical experiments validate the theoretical proofs:

  • Superior Transferability: As shown in Figure 2, the GNN/MNN policy trained on only 50 nodes consistently outperforms standard baselines like WMMSE, equal power allocation, and random allocation, even as the network scales up to 1,000 nodes.
  • Robustness to Noise: The policies remained stable when log-normal noise was added to the channel states.
  • Architecture Impact: Figure 3 shows that while stability decreases slightly as the network gets deeper (more layers) or wider (more filters), the error remains small, validating the stability bounds established in the paper.
Sum-of-rate achieved by GNN trained on small network and execute on larger networks compared with other baseline methods.
Difference between the sum-of-rate ratios on the original wireless network setting and the perturbed one.

Limitations and Future Work

  • Manifold Assumptions: The proofs rely on the manifold being smooth and compact. Real-world wireless deployments might have "sharp" edges or non-smooth densities that don't perfectly fit this model.
  • Filter Complexity: The stability results are specific to FDT filters. Exploring whether other types of manifold filters (like Chebychev or Cayley filters) offer similar stability-transferability benefits is a potential next step.
  • Real-world Data: While the simulations are grounded in standard wireless models, testing on real-world signal traces would further strengthen the case for MNNs in production hardware.

Summary updated

Abstract

We consider the problem of resource allocation in large scale wireless networks. When contextualizing wireless network structures as graphs, we can model the limits of very large wireless systems as manifolds. To solve the problem in the machine learning framework, we propose the use of Manifold Neural Networks (MNNs) as a policy parametrization. In this work, we prove the stability of MNN resource allocation policies under the absolute perturbations to the Laplace-Beltrami operator of the manifold, representing system noise and dynamics present in wireless systems. These results establish the use of MNNs in achieving stable and transferable allocation policies for large scale wireless networks. We verify our results in numerical simulations that show superior performance relative to baseline methods.
BibTeX
@inproceedings{wang2022stable,
  author = {Zhiyang Wang and Luana Ruiz and Mark Eisen and Alejandro Ribeiro},
  title = {{Stable and Transferable Wireless Resource Allocation Policies Via Manifold Neural Networks}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2022},
  doi = {10.1109/icassp43922.2022.9746313},
}
Figure