Generative AI-Based Dependency-Aware Task Offloading and Resource Allocation for UAV-Assisted IoV

Abstract

In recent years, the Internet of Vehicles (IoV) has emerged as a pivotal driving force within intelligent transportation systems, offering users immersive interactive experiences. Meanwhile, unmanned aerial vehicles (UAVs) have demonstrated substantial potential for widespread application within the IoV domain, attributed to their high flexibility, low cost, and ease of deployment. However, as the complexity of IoV tasks increases, complex dependencies among tasks give rise to notable delay issues, which are further exacerbated by the uneven distribution of computational resources. In response to the previously mentioned challenges, we suggest a strategy for resource distribution and task offloading aided by UAVs for IoV. Firstly, by constructing a complex task dependency model, tasks are topologically sorted to clarify the dependencies among tasks, thereby optimizing task execution order. Secondly, focusing on the core issues of task offloading and resource allocation, we present the multi-agent deep deterministic policy gradient (MADDPG) algorithm to devise a dependency-aware scheduling strategy. This strategy integrates task dependencies and UAV mobility characteristics, enabling intelligent decision-making for UAV trajectory planning and task scheduling by analyzing actor and critic network action rewards at each timeslot. To further tackle non-convex optimization problems, we design a federated learning (FL)-based intelligent data caching and computation offloading (Fed-IDCCO) algorithm, leveraging deep reinforcement learning (DRL) techniques. This approach handles large-scale and continuous state and action spaces to obtain optimal task offloading strategies within IoV environments. This methodology not only effectively reduces task processing delays and energy consumption but also significantly enhances the overall system performance. Extensive experimental results demonstrate that, compared to several existing benchmark algorithms, the suggested method offers unique benefits in diminishing delays in task processing, lowering energy usage, controlling costs, and improving cache hit rates.