[Paper Review] Knowledge-Driven Resource Allocation for D2D Networks: A WMMSE Unrolled Graph Neural Network Approach

Yang Hao, Nan Cheng+6 more|arXiv (Cornell University)|Jul 12, 2023

Software-Defined Networks and 5GComputer Science3 citations

TL;DR

This paper proposes a knowledge-driven graph neural network (GNN) that unrolls the weighted minimum mean square error (WMMSE) algorithm to enable low-latency, scalable resource allocation in device-to-device (D2D) networks. By embedding WMMSE’s iterative structure into GNN layers, the model achieves superior generalization and robustness across dynamic topologies and mobility without retraining, outperforming standard GNNs in scalability and sample efficiency.

ABSTRACT

This paper proposes an novel knowledge-driven approach for resource allocation in device-to-device (D2D) networks using a graph neural network (GNN) architecture. To meet the millisecond-level timeliness and scalability required for the dynamic network environment, our proposed approach incorporates the deep unrolling of the weighted minimum mean square error (WMMSE) algorithm, referred to as domain knowledge, into GNN, thereby reducing computational delay and sample complexity while adapting to various data distributions. Specifically, the aggregation and update functions in the GNN architecture are designed by utilizing the summation and power calculation components of the WMMSE algorithm, which leads to improved model generalization and interpretabiliy. Theoretical analysis of the proposed approach reveals its capability to simplify intricate end-to-end mappings and diminish the model exploration space, resulting in increased network expressiveness and enhanced optimization performance. Simulation results demonstrate the robustness, scalability, and strong performance of the proposed knowledge-driven resource allocation approach across diverse communication topologies without retraining. Our findings contribute to the development of efficient and scalable wireless resource management solutions for distributed and dynamic networks with strict latency requirements.

Motivation & Objective

• Address the challenge of low-latency, scalable resource allocation in dynamic D2D networks with high user mobility and varying topologies.
• Overcome the limitations of conventional deep learning models in handling variable-sized networks due to fixed input/output dimensions.
• Improve generalization and reduce sample complexity by embedding domain-specific WMMSE algorithmic knowledge into a GNN architecture.
• Enable end-to-end learning that generalizes across diverse communication topologies and mobility patterns without retraining.
• Theoretical and empirical validation of the model’s expressiveness, optimization performance, and robustness under distribution shift.

Proposed method

• Design a graph neural network (GNN) where aggregation and update functions are derived from the summation and power calculation components of the WMMSE algorithm.
• Unroll the WMMSE algorithm into a deep GNN architecture, using its iterative structure as inductive bias to guide message passing and node updates.
• Integrate domain knowledge from WMMSE into the GNN’s learnable parameters to enhance interpretability and reduce model exploration space.
• Train the model end-to-end using WMMSE outputs as supervision, enabling fast inference with millisecond-level latency.
• Use a dynamic environment simulation with moving devices and varying topologies to evaluate generalization under distribution shift.
• Compare sample efficiency and performance against standard end-to-end GNNs using varying training dataset sizes.

Experimental results

Research questions

• RQ1 Can unrolling the WMMSE algorithm into a GNN architecture improve scalability and generalization in D2D networks with variable user counts and topologies?
• RQ2 How does incorporating WMMSE algorithmic structure into a GNN affect sample complexity and convergence speed compared to standard GNNs?
• RQ3 To what extent does the knowledge-driven GNN maintain performance under dynamic conditions such as user mobility and topology changes?
• RQ4 Does the proposed method generalize better than end-to-end GNNs when test data distribution shifts from training data, especially in sparse or dense topologies?
• RQ5 What is the theoretical justification for the improved expressiveness and optimization performance of the unrolled GNN architecture?

Key findings

• The proposed knowledge-driven GNN maintains over 95% performance under low-speed mobility, while the end-to-end GNN degrades significantly under high-speed movement.
• The model generalizes well across diverse topologies, including fully connected and sparse graphs, without retraining, as shown in performance close to the optimal WMMSE solution.
• In sparse topologies, the model’s performance remains robust and closely tracks WMMSE, whereas standard GNNs degrade with increasing interference density.
• The network achieves strong performance even with small training datasets, indicating low sample complexity due to high algorithm alignment with WMMSE.
• Theoretical analysis confirms that the model simplifies end-to-end mappings and reduces the model exploration space, enhancing optimization and expressiveness.
• Performance in the weighted sum rate scenario generalizes better than in the rate-only scenario due to richer data distribution learning during training.