Explainable Multiagent Deep Reinforcement Learning for Joint Task Offloading and Resource Allocation in Distance and Channel-Aware NOMA Vehicular Edge Networks
Jianqiang Hu, Lin Chen, Shigen Shen, Tian Wang
Abstract
With the rapid development of intelligent transportation and vehicular edge computing (VEC), efficient, fair, and interpretable task offloading has become a key challenge in dynamic and resource-constrained environments. Non-orthogonal multiple access (NOMA) can enhance connectivity and spectrum efficiency. However, conventional resource allocation strategies typically rely solely on channel gain ordering while overlooking spatial factors and fairness. In addition, the lack of transparency in multi-agent deep reinforcement learning (MADRL) decision-making raises concerns regarding transparency and trustworthiness. To address these challenges, we propose a NOMA-based task offloading framework that integrates distance and channel-aware resource allocation, and we design a distributed multi-agent decision-making algorithm based on potential games (DACA-MAD4PG), further incorporating Shapley Additive Explanations (SHAP) to improve interpretability. The proposed framework is significantly different from existing NOMA-based task offloading approaches in the following three aspects. First, it introduces a distance and channel-aware joint resource allocation mechanism to enhance both efficiency and fairness in vehicular edge computing. Second, an exact potential game is incorporated to guarantee system stability and the existence of Nash equilibria. Last, SHAP is integrated to provide post hoc interpretability, thereby improving transparency in multi-agent decision-making. Experiments based on real-world DiDi trajectory data demonstrate that the proposed approach significantly reduces task latency, improves service success rate, cumulative reward, and fairness, and outperforms several baselines, thereby providing a stable and interpretable solution for VEC task offloading.