Scalable and Energy-Efficient Deep Reinforcement Learning for Mobile Edge Computing

Abstract

Scalable decentralized control for Mobile Edge Computing (MEC) remains an open challenge. Existing learning-based approaches either centralize decision-making or rely on state spaces that grow with the number of devices, limiting their applicability in dense sixth-generation (6G) deployments. We propose a decentralized Multi-Agent Deep Reinforcement Learning (DRL) framework for task offloading that addresses these limitations simultaneously. The problem is formulated as a Partially Observable Stochastic Game (POSG) with a bounded, aggregated observation space that remains fixed as the number of User Devices (UDs) increases, enabling true scalability without retraining. A cooperative, performance-oriented reward eliminates hand-tuned energy–latency trade-off coefficients, jointly promoting energy efficiency and fairness across heterogeneous UDs. Simulation results confirm that the proposed framework outperforms multiple benchmark schemes without centralized coordination or prior knowledge of system dynamics. Specifically, it achieves 11.8% higher task success rate, uses up to 61.6% less energy, and uniquely maintains near-perfect Jain’s fairness index as the network scales from 2 to 100 UDs.