Multi-task scheduling in vehicular edge computing: a multi-agent reinforcement learning approach

Vehicular Edge Computing (VEC) represents a computational paradigm designed to optimize vehicular performance and enhance vehicle services. It entails strategically deploying computational and storage resources close to vehicles, facilitating rapid task execution with minimal latency. However, traditional task offloading methods usually consider offloading the tasks to the Road Side Units (RSUs) entirely and neglect the dependencies between the tasks. In this paper, we focus on addressing the dependent task scheduling problem to efficiently offload multiple computation-intensive vehicle applications to RSUs. Each application generated by the vehicle consists of a series of tasks with dependencies, which can be represented by a Directed Acyclic Graph (DAG). Furthermore, we define the multitask scheduling problem as an NP-hard optimization problem. To efficiently solve this problem, we model it as an Exact Potential Game (EPG) problem among the vehicles and prove the existence of Nash equilibrium. Then, we propose a Task Priority Sorting Algorithm (TPSA) to convert the task DAG graph into a prioritized task sequence. Utilizing the mapping of task sequence to the state space of reinforcement learning algorithms, we design a Dependency Task Offloading Algorithm based on Multi-Intelligent Deep Reinforcement Learning (DEMA) to address the offloading decisions for each task. Finally, numerical simulations are conducted to evaluate the performance of the proposed DEMA algorithm.