Fast-Convergence Reinforcement Learning for Routing in LEO Satellite Networks

Area of Science:

• Computer Science
• Network Engineering
• Artificial Intelligence

Background:

• Low Earth Orbit (LEO) satellite networks face routing challenges due to dynamic topology and varying transmission needs.
• Traditional routing algorithms like Open Shortest Path First (OSPF) struggle with the rapid link state changes inherent in LEO networks.

Purpose of the Study:

• To develop a novel routing algorithm for LEO satellite networks that ensures fast convergence.
• To improve the adaptability and efficiency of routing in dynamic satellite communication environments.

Main Methods:

• Proposed a Fast-Convergence Reinforcement Learning Satellite Routing Algorithm (FRL-SR) where each satellite acts as an agent.
• Agents use reinforcement learning to select forwarding ports based on network link status.
• Implemented a mechanism using 'hello' packets for rapid policy updates upon network state changes.

Main Results:

• FRL-SR demonstrates faster information perception and convergence compared to traditional reinforcement learning algorithms.
• The algorithm effectively masks LEO network topology dynamics and adaptively adjusts forwarding strategies.
• Experimental results show FRL-SR surpasses the Dijkstra algorithm in average delay, packet arrival ratio, and network load balance.

Conclusions:

• FRL-SR offers a significant improvement for routing in LEO satellite networks.
• The proposed algorithm provides a robust solution for handling dynamic network conditions.
• FRL-SR enhances overall network performance metrics crucial for satellite communication.