Transfer Learning for Distributed Intelligence in Aerial Edge Networks

Unmanned aerial vehicle (UAV)-enabled edge computing is a promising paradigm to provide task offloading service for computation-intensive applications. Catering for highly dynamic edge networks, learning agents are placed on the UAVs to perceive environmental changes and adjust edge service strategies. These multiple agents build distributed intelligence in edge service scheduling. However, the agents, which independently learn service strategies from scratch, always cost high computing resources, cause extra delays, and bring a critical challenge to UAVs' service efficiency. To address this challenge, we propose a transfer learning empowered aerial edge network (AEN), which uses multi-agent machine learning to draw optimal service strategies, while leveraging transfer learning to share and reuse knowledge between the UAVs for saving resource costs and reducing training latency. Furthermore, we develop efficient transfer learning schemes where the edge resources are optimally scheduled for knowledge sharing and task offloading. Numerical results indicate that the proposed schemes maximize edge service utility while significantly improving machine learning efficiency.