Reinforcement learning-based robust routing strategies against cascading failures in LEO satellite networks

Area of Science:

• Space Systems Engineering
• Network Communications
• Artificial Intelligence

Background:

• Low Earth orbit (LEO) satellite networks are vital for global communication but face challenges from dynamic topology and traffic loads.
• Existing static routing methods struggle with the unpredictable nature of LEO constellations, leading to vulnerabilities like cascading failures.

Purpose of the Study:

• To develop a robust and adaptive routing strategy for LEO satellite networks using reinforcement learning.
• To improve network resilience against traffic imbalances and node failures.

Main Methods:

• A realistic time-varying spatiotemporal load model incorporating tidal traffic and spatial phase differences was developed.
• A virtual node topology mapping mechanism was established.
• The routing problem was framed as a Markov decision process, and a distributed Q-learning algorithm was designed for adaptive link weight reconfiguration based on residual capacity and load-reciprocal priority.

Main Results:

• The proposed reinforcement learning strategy significantly outperformed static routing baselines in maintaining connectivity during node failures.
• The method demonstrated substantial robustness gains at the critical breakdown threshold.
• The algorithm achieved rapid convergence, superior cost-effectiveness, and high reliability with lower normalized construction costs.

Conclusions:

• Reinforcement learning offers a promising solution for robust routing in dynamic LEO satellite networks.
• The developed strategy enhances network resilience, connectivity, and cost-effectiveness.
• This research provides insights for designing effective routing strategies in future LEO constellations.