Supervised Learning of Neural Networks for Active Queue Management in the Internet

Area of Science:

• Computer Science
• Network Engineering
• Artificial Intelligence

Background:

• Active Queue Management (AQM) is crucial for mitigating network congestion.
• Traditional AQM mechanisms like PIα face challenges with complex network traffic patterns.
• Self-similar network traffic, often modeled by fractional Gaussian noise, requires advanced management strategies.

Purpose of the Study:

• To develop a machine learning model that replicates the behavior of the AQM PIα mechanism.
• To investigate the effectiveness of a neural network-based AQM for managing network traffic.
• To enhance network performance by proactively dropping packets before buffer overflow.

Main Methods:

• Utilizing neural networks to model AQM behavior.
• Generating training data incorporating the self-similarity of network traffic using fractional Gaussian noise.
• Conducting quantitative analysis through simulations.

Main Results:

• The proposed neural network-based AQM mechanism was evaluated.
• Simulations analyzed key performance metrics: queue length, packet rejection rates, and waiting times.
• The AI-driven AQM demonstrated its utility in managing network resources.

Conclusions:

• Neural network-based AQM presents a viable and effective approach to network traffic management.
• The proposed model successfully mimics AQM PIα behavior while handling self-similar traffic.
• This AI-driven solution offers improved network performance through intelligent queue management.