MEC-Hopper: DRL-Based Adaptive Multi-Hop Service Provisioning in UAV-Assisted MEC

The advent of 5G has brought network slicing and Multi-access Edge Computing (MEC) to the fore, which has enabled provisioning of various services at the network edge. With Unmanned Aerial Vehicles (UAVs), these edge services can be offered in areas with limited accessibility to the traditional network, without compromising service requirements. Due to cost, power and resource constraints, it is desirable to let only a few UAVs provide a service. UAVs flying around to cover different zones may result in longer service delays to the ground User Equipments (UEs). On the other hand, hovering the UAVs over fixed zones and letting a UAV offload jobs to any other UAV would require line-of-sight communication among all UAVs. In our work, we address the above challenges and develop an efficient solution. In the proposed framework, MEC-Hopper, the UAVs hover over zones iteratively defined by clusters of ground UEs, with only a subset of the UAVs instantiating the services while others act as relays for offloading jobs, facilitating multi-hop communications. We develop a centralized Deep Reinforcement Learning (DRL) algorithm to select a subset of UAVs for hosting the services, and a multi-hop latency-aware job distribution algorithm to fairly distribute the jobs, providing acceptable low latency. Experimental results demonstrate MEC-Hopper’s ability to outperform a baseline static edge service provisioning framework in terms of the latencies incurred. Moreover, we exemplify the resilience of the proposed framework against a Denial-of-Service (DoS) attack, without necessitating re-training or implementing additional mitigation strategies. Additionally, MEC-Hopper’s longer hovering time demonstrates the advantages of choosing a subset of UAVs for service provisioning.