An Explainable and Objective-Aware DRL Framework for MEC-Enabled Vehicular Networks

Abstract

Multi-access Edge Computing (MEC) resource allocation in vehicular (V2X) networks is a multi-objective control problem that jointly optimizes latency, load balancing, and mobility constraints under highly dynamic conditions. Although deep reinforcement learning (DRL) can adapt to such environments, its decision-making process often lacks transparency, making it difficult to trust in practice. In this work, we propose an eXplainable Artificial Intelligence (XAI) assisted DRL framework for MEC orchestration composed of two main components: the eXtended Maskable Proximal Policy Optimization (XMPPO) architecture and the Margin-Aware Policy Explanation (MAPE) framework. To support multi-objective learning without collapsing objectives into a single reward signal, we design an objective-aware XMPPO variant with decomposed value estimation. Moreover, we propose the MAPE framework that employs a perturbationbased XAI methodology to identify minimal sufficient subsets of state features while preserving the logit margin of a frozen DRL policy. Experiments conducted in the Simu5G simulator show that the proposed XMPPO minimizes the total response time (TRT) compared to the standard PPO while jointly optimizing all delay-related factors. In addition, the proposed MAPE framework provides a better performance-complexity-interpretability tradeoff compared to the classical SHapley Additive exPlanations (SHAP) XAI scheme. Experimental results show that feature-level explainability helps audit the state representation and improve transparency in learning-based MEC orchestration. By selecting the most decision-relevant features, our approach reduces total response time TRT by 7%, percentile 95 latency p95 by 9%, and deadline miss rate DMR by 5%, while improving fairness by 12% compared with the full-state baseline.