Reinforcement Learning-Based Latency Minimization for Vehicular Edge Computing Networks

Pushing computing functions to network edges can save backhaul network bandwidth and meet diverse latency requirements, thus supporting computation-intensive and delaysensitive tasks. However, challenges such as unstable network conditions, vehicle mobility, and strict latency requirements make it crucial to decide on local task provision. This paper focuses on the cloud-edge latency optimization problem (LOP) based on task offloading and resource allocation in the vehicle edge network, taking into account vehicle mobility, time-varying channels, and signal blocking. First, the LOP is formulated as a mixed-integer nonlinear programming (MINLP) model. We reformulate this challenging optimization problem as a Markov game and incorporate a unique delay task dropout penalty mechanism to efficiently tackle this challenging optimization problem. This mechanism effectively reduces the negative impacts of latency on task execution and system performance. Then, we design an intelligent decision-making algorithm based on a multiagent deep deterministic policy gradient. The proposed algorithm doesn't require prior knowledge of uncertain parameters and works without relying on dynamic models. Extensive simulation tests show that our proposed algorithm converges quickly and enhances system performance more effectively than the three alternative solutions.