A Multi-agent Reinforcement Learning based Offloading and Resource Allocation for Vehicle Edge Computing

This paper investigates the the challenges of resource allocation optimization for computation offloading in vehicular edge networks. It is assumed that multiple moving vehicles select Mobile Edge Computing (MEC) servers on Road Side Units (RSUs) to offload their computation tasks, for which, MEC servers allocate appropriate resources. To achieve the best balance between minimizing latency and resource consumption to ensure system performance, this study utilizes a Multi-Agent Reinforcement Learning (MARL) framework, where each vehicle and MEC server acts as an agent and learns optimal strategies through interactions with the environment. To improve performance, a new Global Heterogeneous MARL (GHMARL) framework is purposely designed based on existing MARL approaches. The proposed framework trains two Critic networks, namely, a global Critic network to ensure service quality and a local Critic network to ensure resource utilization efficiency. Extensive simulation results demonstrate the performance superiority of this approach over benchmark algorithms.