Dependent task offloading for air-ground integrated MEC networks: a multi-agent collaboration approach

The air-ground integrated mobile edge computing (MEC) network provides an efficient data processing platform for application services with timing-dependent subtasks in the Internet of Things (IoT). Nevertheless, in multi-user concurrency scenarios, due to the unbalanced node load and the complexity of service requests, it may lead to long user response latency. In this paper, we propose the air-ground collaborative task offloading (AGC_TO) scheme, aiming to minimize total task offloading delay and maximize node resource utilization. The approach considers the dynamic resource scenarios in an integrated subtask timing dependency constraints, unmanned aerial vehicles (UAV) flight service range change, offloading location optimization, CPU resource allocation and node load optimization, CPU resource allocation, and node load five key factors. Firstly, We employ a subtask preprocessing (STP) algorithm based on depth-first traversal to analyze the topology of timing-dependent tasks in terms of obtaining different sequences of string parallel divisions for single-user request tasks in a multi-user concurrency scenario. Subsequently, utilizing this sequence with dynamically changing global computational resource information, considering the mobility of the UAV as well as the dynamically available resources among nodes, the optimal offloading location of the subtask can be autonomously determined by the multi-agent collaborative offloading algorithm based on multiagent deep deterministic policy gradient (MADDPG). Numerical results demonstrate that the proposed scheme is significantly better than the representative benchmark in terms of task offloading latency and system resource utilization. Specifically, the proposed scheme outperforms the single agent offloading algorithms double deep-Q network based task offloading (DDQN_TO) and deep Q-network based task offloading (DQN_TO), reducing task offloading latency by 41.4% and 46.4%, and enhancing resource utilization by 8.4% and 8.6%, respectively.