부산대학교 도서관

Aiming to solve the problem of interruption of normal communication service caused by the damage of ground communication facilities after disaster, an Air-Ground Integrated Mobile Edge Network (AWMEN) offloading model was established under the constraints of communication security, energy consumption and coverage. The traditional method needs to be re-iterated every time the preset environmental state changes, which will waste a lot of communication resources and computing resources, greatly reduce the efficiency, and face the risk of data privacy disclosure. However, the deep reinforcement learning method under the federated learning framework will be more flexible and applicable to dynamic scenarios. A Markov decision process model is constructed based on the unmanned aerial vehicles (UAV) and the environment. The experience trajectory is designed by interacting with the external environment, and the optimal offloading strategy is obtained. The Twin Delayed Deep Deterministic Policy Gradient of behavior cloning (TD3-BC-R) is compared with baseline method (0-1 mode), Actor-Critic (AC-R), Deep Deterministic Policy Gradient (DDPG-R) and Twin Delayed Deep Deterministic Policy Gradient (TD3-R), the experiment shows that, The total time cost of TD3-BC-R is reduced by more than 1/3, and low latency transmission is also achieved.