Adaptive Pareto-Optimal and Generalizable Multi-Objective Offloading and Resource Scheduling in Dynamic Mobile Edge Computing for Enhanced User Experience

Mobile Edge Computing (MEC) faces significant challenges in dynamic environments, balancing conflicting objectives like latency, energy, and Quality of Experience (QoE) amidst heterogeneous resources and multi-user competition. These issues are compounded by poor generalization and slow adaptation. This paper introduces APOG-MARL (Adaptive Pareto-Optimal and Generalizable Multi-Agent Reinforcement Learning), a novel framework built on an Adaptive-Contextual Multi-Objective Markov Decision Process (MOMDP). APOG-MARL integrates a hierarchical context-aware state representation for generalization, a multi-objective Pareto policy network for optimal trade-offs, a constraint-driven multi-agent collaboration mechanism for efficient resource management, and a meta-learning approach for rapid user preference adaptation. Extensive simulations demonstrate APOG-MARL's superior performance across varying network scales, dynamic user preferences, and high resource utilization scenarios. It achieves enhanced user QoE, significantly lower average task latency and total energy consumption, superior Pareto front quality, and robust resource utilization, consistently outperforming state-of-the-art baselines. APOG-MARL offers a powerful and practical solution for optimizing task offloading and resource scheduling in complex, dynamic MEC environments.