Delay-Aware Large-Small Model Collaboration over LEO Satellite Networks
Multi-agent reinforcement learning balances computation and communication across satellite networks.
A new paper from researchers at (affiliations unspecified) introduces a delay-aware collaboration scheme that pairs small AI models on resource-constrained remote sensing satellites with large models on more capable computing satellites, all connected via low Earth orbit (LEO) satellite networks. The key innovation is a joint optimization of two interdependent decisions: which tasks to offload from sensing satellites to computing satellites, and how to route data across inter-satellite links to minimize total service delay.
The team formulated this as a decentralized partially observable Markov decision process (Dec-POMDP) and solved it using a multi-agent reinforcement learning (MARL) algorithm. Their approach uses offline policy training to pre-learn effective routing strategies, then an online bisection search to iteratively adjust offloading decisions in real time. Simulation results show the scheme reduces service delay by up to 31.85% compared to baseline methods, making it a promising solution for real-time Earth observation and other latency-sensitive satellite applications.
- Proposes a large-small model collaboration framework for LEO satellite networks, balancing computation and communication loads.
- Uses multi-agent reinforcement learning with offline policy training for routing and online bisection search for offloading decisions.
- Achieves up to 31.85% reduction in service delay compared to benchmarks in simulation.
Why It Matters
Enables faster AI inference on satellite networks, critical for real-time Earth observation and disaster response.