Federated Deep Reinforcement Learning-based task offloading strategy with game theory in vehicular edge computing

Abstract

The increasing complexity of Internet of Vehicles (IOV) applications poses significant challenges to vehicular onboard computing resources, leading to heightened latency and energy consumption. Task offloading techniques in vehicular edge computing (VEC) offer a promising solution by transferring computational tasks to distributed edge servers with enhanced processing power. However, in highly dynamic VEC scenarios, multiple vehicles tend to offload tasks concurrently, exacerbating system challenges. An inappropriate offloading strategy can result not only in increased system latency but also in severe privacy breaches. To address these issues, a federated deep reinforcement learning-based task offloading strategy with game theory (FDRLGT) is proposed to minimize total system delay and protect user privacy. Specifically, Deep Reinforcement Learning (DRL) is used to train a local offloading strategy model with local data, while Federated Learning (FL) aggregates local model parameters instead of raw data to ensure privacy. In multi-vehicle simultaneous task offloading contexts, we address the problem of policy homogeneity in FDRL, which can lead to locally suboptimal solutions. To overcome this, we design a game theory model integrated into the FDRL algorithm to enhance optimization. Simulation results demonstrate that the proposed FDRLGT algorithm enhances system efficiency, ensures privacy, and effectively reduces total system delay compared to other baseline algorithms.