Open Access

ARTICLE

A Multi-Agent Deep Reinforcement Learning-Based Task Offloading Method for 6G-Enabled Internet of Vehicles with Cloud-Edge-Device Collaboration

Fangxiang Hu¹, Qi Fu^1,2,*, Shiwen Zhang¹, Jing Huang¹

1 School of Computer Science and Engineering, Hunan University of Science and Technology, Xiangtan, China
2 School of Electronic Information Engineering, Changsha Institute of Technology, Changsha, China

* Corresponding Author: Qi Fu. Email:

Computers, Materials & Continua 2026, 87(3), 75 https://doi.org/10.32604/cmc.2026.074154

Received 03 October 2025; Accepted 06 February 2026; Issue published 09 April 2026

Abstract

In the Internet of Vehicles (IoV) environment, the growing demand for computational resources from diverse vehicular applications often exceeds the capabilities of intelligent connected vehicles. Traditional approaches, which rely on one or more computational resources within the cloud-edge-device computing model, struggle to ensure overall service quality when handling high-density traffic flows and large-scale tasks. To address this issue, we propose a computational offloading scheme based on a cloud-edge-device collaborative 6G IoV edge computing model, namely, Multi-Agent Deep Reinforcement Learning-based and Server-weighted scoring Selection (MADRLSS), which aims to optimize dynamic offloading decisions and resource allocation. The scheme first designs an improved multi-agent proximal policy optimization (MAPPO) algorithm, decoupling centralized training from distributed execution for multiple terminal vehicle agents. Specifically, the centralized training of terminal vehicles is migrated to the high-performance edge layer, while lightweight decision-making networks are retained at the terminal vehicles to enable efficient and dynamic task offloading decisions. Additionally, a server-weighted scoring selection (SS) algorithm is proposed, which integrates two key metrics—short-term server load and geographical proximity—to select the optimal server and allocate communication resources. The proposed scheme improves the quality of experience (QoE) while balancing energy consumption. Simulation results demonstrate that the MADRLSS scheme significantly outperforms existing benchmark methods in terms of task offloading efficiency and stability, maintaining QoE consistently above 82% and effectively enhancing service quality in complex vehicular scenarios.

---

Keywords

---