Task Scheduling in Vehicular Networks: A Multi-Agent Reinforcement Learning Based Reverse Auction Mechanism

To maximize the utilization of system resources and improve the overall execution effectiveness of task offloading in vehicle networking is the aim of this paper, which enables vehicles to access various intelligent services through the Internet. Due to the limited computing power and energy of vehicles, task offloading can improve the performance and efficiency of complex and time-consuming tasks by leveraging the edge servers. However, task offloading also involves the trade-off and coordination among multiple vehicles and servers, which poses a great challenge for task and resource allocation. To tackle this problem, we propose a novel method that combines multi-agent reinforcement learning and reverse auction mechanism, which allows vehicles and servers to learn and optimize their own bidding strategies for task offloading, and achieve a fair and efficient outcome. Proximal policy optimization (PPO) algorithm with long short-term memory (LSTM) networks is employed to train our multi-agent reinforcement learning model. We conduct extensive experiments in our open-source Python-based simulation environment, VehicleJobScheduling, which can realistically model the vehicle networking scenario and provide various evaluation metrics. The study reveals that the PPO+LTSM strategy utilizes LSTM to capture temporal information, enabling more rational decision-making and achieving the most effective load balance.