Multi-Agent Reinforcement Learning Based Cooperative Multitype Task Offloading Strategy for Internet of Vehicles in B5G/6G Network

With the development of intelligent transportation, various computation intensive and delay sensitive applications are emerging in the Internet of Vehicles (IoV). The B5G/6G(Beyond 5th generation mobile communication technology/ 6th generation mobile communication technology) network has the characteristics of ultra-low latency and ultramany connections. The deployment of the Network in Boxes (NIB) supporting B5G/6G network in the vehicle can realize the real-time communication with the edge server (ES) and offload the task to the ES. However, the current Multi-access Edge Computing (MEC) lacks research on cooperative processing among multiple edge servers (ESs), and the efficiency of data-intensive computation tasks is still insufficient. In this paper, we investigate the cooperative offloading of multi-type tasks among ESs in B5G/6G networks under dynamic environment. In order to minimize the delay of task execution, we regard cooperative offloading as a Markov Decision Process (MDP), and improve the convergence speed and stability of traditional SAC(Soft Actor Critic) algorithm by adaptive weight sampling mechanism. Finally, an offline centralized training distributed execution framework based on improved soft actor critical (ISAC)(OCTDE-ISAC)is proposed to optimize the cooperative offloading strategy. The experimental results show that the proposed algorithm is better than the existing algorithm in terms of latency.