Abstract:

In the scenario of the space-air-ground integrated vehicle network (SAGVN),due to the limited battery capacity and energy,unmanned aerial vehicle (UAV) devices cannot provide long-term effective support for task offloading; due to resource costs,communication delays,and delay jitter,it is difficult for low-Earth orbit (LEO) satellites to provide stable high bandwidth communication services for the Internet of Vehicles(IoV) tasks.To address the resource optimization problem of UAVs and LEO satellites in the scenario of SAGVN,a solution for task offloading,power adjustment,and caching decision-making based on multi-task deep reinforcement and auxiliary learning (MTDRAL) is proposed.First,the task segmentation and transmission model,latency model,energy consumption model,server computation and caching model,and problem model are established.Then,by comprehensively considering the task processing delay,server energy consumption,and cache hit rate,an MTDRAL-based solution for task offloading and resource scheduling is proposed.Finally,comparative experiments are conducted between the proposed solution and four benchmark strategies:a random offloading strategy,a success-rate-based greedy decision strategy,a multi-network deep reinforcement learning offloading strategy based on the soft actor-critic algorithm,and a multi-network deep reinforcement learning offloading strategy based on the deep deterministic policy gradient algorithm.Experimental results show that when the server count is 14 and the number of vehicle terminals is 10,the proposed solution outperforms the four comparative strategies by 134.41%,31.32%,38.93%,and 29.49% in comprehensive scoring,respectively.The proposed solution has a good performance and can better meet the task offloading requirements in the scenario of SAGVN.

Key words: space-air-ground integrated vehicle network(SAGVN), resource optimization, task offloading, reinforcement learning