Multi-Agent Deep Reinforcement Learning in Flying Ad-Hoc Networks for Delay-Constrained Applications

In the near future, 6G networks will be able to provide its users with heterogeneous tailored end-to-end network services for edge AI-powered applications. However, in many 6G use case scenarios there may not be structured network infrastructures. One of the most appealing solutions to meet this problem is to bring computing, networking and storage facilities on site by means of UAVs. The target of this paper is the definition of a management framework for Flying Ad-Hoc Networks (FANET) aimed at bringing edge computing to remote areas, guaranteeing low latency to ground devices with heterogeneous requirements in terms of mean latency. The proposed framework allows zero-touch management of different network slices realized by partitioning the Computing Elements (CE) installed on each UAV of the FANET and wireless links connecting them to each other. An Inter-Slice Orchestrator is in charge of deciding this resource partition, while an Intra-Slice Orchestrator manages horizontal offload among UAVs for each slice via Reinforcement Learning, in order to reduce end-to-end latency and delay jitter.