ATARI: A Graph Convolutional Neural Network Approach for Performance Prediction in Next-Generation WLANs

Area of Science:

• Computer Science
• Electrical Engineering
• Wireless Networking

Background:

• High-density Wi-Fi deployments face severe performance degradation due to interference, exacerbated by wider channels in new standards like 802.11n/ac.
• Channel Bonding (CB) increases network capacity but complicates channel assignment in dynamic environments.
• Traditional analytical or system models struggle with accuracy and computational cost for predicting Wi-Fi performance in complex scenarios.

Purpose of the Study:

• To develop a novel, data-driven approach for accurately and efficiently predicting Wi-Fi performance in dense, dynamic deployments.
• To overcome the limitations of existing models in handling the combinatorial complexity of channel assignment with Channel Bonding.
• To leverage graph neural networks (GNNs) for improved Wi-Fi performance prediction.

Main Methods:

• Utilized a Graph Neural Network (GNN) model trained on deployment topology and wireless interaction data.
• Exploited the graph structure of the network to capture intricate wireless interactions.
• Evaluated the GNN model's performance against naive and other Machine Learning (ML) approaches.

Main Results:

• The GNN model achieved high accuracy in predicting Wi-Fi performance.
• Preserving the graph structure during learning resulted in a 64% performance increase over a naive approach.
• The proposed method showed a 55% improvement compared to other ML approaches when using all training features.

Conclusions:

• A data-driven GNN approach offers a significant improvement for predicting Wi-Fi performance in dense deployments.
• Graph structure preservation is crucial for effective GNN-based wireless performance prediction.
• This method provides an accurate and computationally efficient solution for dynamic channel assignment challenges in modern Wi-Fi networks.