Task Migration Strategy in Vehicular Networks Based on Reinforcement Learning

With the research and application of edge computing in vehicular networks, computing tasks can be offloaded from vehicles to roadside edge servers to reduce system service latency. However, as vehicles move, the computing tasks need to be migrated from one edge server to another. Predicting the vehicles movement trajectory and formulating a reasonable task migration plan for this is a key challenge that needs to be addressed. Traditional computing offloading methods cannot be directly applied in vehicular networks. Therefore, this paper constructs a vehicular task offloading system based on a multi-layered computing network, introduces a Markov mobility model to describe the vehicle movement trajectory, and solves the optimal migration path problem. Since this problem is NP-hard, a solution method based on a constrained Markov model is proposed, along with an Actor-Network Primal-Dual Deep Deterministic Policy Gradient (ANPD-DDPG) algorithm based on reinforcement learning to achieve the optimal solution. Finally, in simulation experiments, the proposed method is compared with existing research, showing about a 33% reduction in system delay and migration cost. The characteristics of the ANPD-DDPG algorithm in terms of convergence speed and system delay are also analyze.