Deep Reinforcement Learning-Based Routing Method for Low Earth Orbit Mega-Constellation Satellite Networks with...

Area of Science:

• Telecommunications Engineering
• Network Routing
• Artificial Intelligence in Networking

Background:

• Low-orbit satellite communication networks are crucial for 5G and 6G, offering wide coverage but facing challenges in stable path construction, load balancing, and congestion.
• Service Function Chain (SFC) constraints in 3GPP architectures further complicate routing in large-scale Low Earth Orbit Mega Satellite Networks (LEO-MSNs).

Purpose of the Study:

• To propose GDRL-SFCR, an end-to-end routing decision method that jointly optimizes transmission delay and network load balancing.
• To address the complexities of LEO-MSNs under Service Function Chain (SFC) constraints.

Main Methods:

• Developed an end-to-end routing method, GDRL-SFCR, integrating Graph Neural Networks (GNN) and Deep Reinforcement Learning (DRL).
• Utilized GNN for extracting node attributes and dynamic topology features within the NTN low-orbit satellite network architecture.
• Employed DRL with custom reward functions to train routing policies that satisfy SFC constraints, transmission delays, and network loads.

Main Results:

• GDRL-SFCR reduced end-to-end traffic transmission delay by over 11.3% compared to existing methods.
• Achieved a reduction in average network load by more than 14.1%.
• Increased traffic access success rate by over 19.1% and network capacity by twofold.

Conclusions:

• GDRL-SFCR effectively optimizes routing in LEO-MSNs, balancing delay and load under SFC constraints.
• The proposed GNN and DRL-based approach offers superior performance over graph theory and traditional RL methods for satellite network routing.