Fairness-Aware Multi-Agent Learning-based Task Offloading in Dynamic Vehicular Scenarios

Task offloading in mobile edge computing (MEC) is essential for reducing latency, balancing workload, and meeting task deadlines. In the scope of vehicular networks, the concept of a vehicular micro cloud (VMC) has been designed to handle such edge computing without the need for installed MEC infrastructure. However, the mobility vehicles and the need to fairly distribute workload make it vital to develop an adaptive and intelligent task offloading strategy. We propose a multi-agent twin delayed deep deterministic policy gradient (MATD3)-based task offloading strategy that enables vehicles to make decentralized, dynamic offloading decisions. Our solution significantly enhances overall fairness stability, and also improves delays to meet relevant deadlines. Our approach is evaluated in a single intersection scenario and a real-world traffic scenario from Nagoya. We compare our approach with a greedy and exhaustive baseline that sequentially offloads tasks to the current least-loaded vehicle. Compared to the baselines, our solution can deal with dynamic scenarios and provide long-term workload fairness in combination with reduced delays.