Intelligent Recovery for Large-Scale LEO Networks
Deep Reinforcement Learning for User Reconnection in Software-Defined Large-Scale LEO Satellite Networks
A study of how Deep Reinforcement Learning and Software-Defined Networking can restore user connectivity after satellite node failures—while balancing computational speed, resource efficiency, and solution quality.
Dynamic satellite networks require rapid, autonomous decision-making.
simulation experiments
algorithms compared
primary performance metrics
DQN decision stages
What happens when
a satellite suddenly fails?
LEO satellite networks move rapidly and constantly change topology. Damage caused by space debris, component failure or malicious attacks can disconnect large numbers of users almost instantly.
Traditional exact algorithms can find optimal reconnection schemes, but computational cost grows dramatically as networks scale. Greedy heuristic methods are fast, but may become trapped in locally optimal solutions.
From node failure
to network recovery.
The architecture combines an LEO satellite layer, an SDN-based control plane and terrestrial users. The controller maintains a global network view and coordinates user reconnection after node failure.
A four-stage
decision process.
The reconnection task is modeled as a Markov Decision Process. A DQN agent learns which satellite should receive each disconnected user while minimizing total system cost.
Detect
Identify the damaged satellite and determine which users have lost service.
Observe
Build a global SDN view of satellite load, connectivity and resource availability.
Decide
DQN selects legal reconnection actions according to the long-term reward and system cost.
Reconnect
Restore service and output satellite load, user associations and total system cost.
Performance under
increasing network scale.
Backtracking, DRL and heuristic algorithms were tested across multi-user scenarios.
Lower runtime
DRL versus Backtracking
at larger user scale.
Above optimal cost
DRL approaches the theoretical Backtracking optimum.
Lower system cost
DRL versus the Heuristic approach at larger scale.
Runtime Comparison
As user scale increases, the computational cost of exact Backtracking rises rapidly. DRL maintains substantially lower runtime while remaining relatively close to heuristic inference speed.
System Cost / Energy Consumption
Backtracking represents the theoretical optimum. DRL remains close to that optimum while producing lower cost than the heuristic approach.
DRL becomes more valuable as network scale increases.
Percentage differences reported in the paper, comparing smaller and larger user scenarios.
Space infrastructure is becoming part of the communication network.
Satellite networks extend communications beyond terrestrial infrastructure.
Moving nodes continuously reshape available communication paths.
Space and terrestrial systems operate as one integrated network.
Not simply the fastest.
Not simply the lowest-cost.
The balance between both.
Exact search can achieve the theoretical optimum but becomes computationally expensive as the network scales. Heuristic algorithms respond rapidly but sacrifice allocation quality. The proposed DRL approach operates between these extremes—approaching optimal system cost while maintaining practical inference speed.
Deep Reinforcement Learning for User Reconnection in Software-Defined Large-Scale LEO Satellite Networks
Read the complete system model, optimization formulation, DQN architecture and experimental results.
PAPER