Computation Offloading Based on Deep Reinforcement Learning for UAV-MEC Network

The existing MEC technology is not suitable for the situations where the number of Mobile Users (MUs) increases explosively or network facilities are sparsely distributed, and general MEC solutions cannot fully address the MU emergency communication and task offloading requirements in post disaster emergency communication networks. Unmanned Aerial Vehicle (UAV) can play an important role in wireless systems because it can be flexibly deployed to quickly help disaster areas improve signal coverage and restore communication quality, enabling MUs in disaster areas to unload tasks normally. This article investigates the deployment of UAV-MEC systems in emergency situations of network communication interruption in the region after a disaster, and studies the optimization strategy for task offloading in edge environments. Based on the relevant advantages of UAV and MEC technology, we have constructed a UAV-MEC system model for emergency disaster relief scenarios. Under the constraints of UAV flight trajectory and MU calculation mode, the problem of minimizing the total system delay is formulated. Due to the non convexity of this problem, there are high-dimensional state spaces and continuous action spaces, making it difficult to find the optimal solution. This article proposes a Split DQN (SDQN) algorithm based on Reinforcement Learning (RL) and uses it to solve optimization problems. The results of simulation experiments show that our proposed SDQN algorithm can approach the global optimal solution of the optimization problem more closely than the benchmark algorithm, and thus reduce the total system delay. Thus, it is an effective optimization decision-making scheme.