Coordinated Offloading Cooperative Multi-Agent Reinforcement Learning for the Edge–Cloud Continuum

Abstract

The increasing demand for intelligent, low-latency services in edge–cloud continuum systems poses new challenges for dynamic and efficient task offloading. We propose a Multi Agent Reinforcement Learning (MARL) framework for dis tributed task offloading under partial observability, where each device offloads only a portion of its computation to a target edge server while executing the remaining portion locally (partial offloading). The problem is formulated as a Markov Decision Pro cess (MDP), where each device independently learns an optimal policy under partially observable system states. To address the non-convexity, high-dimensional state space, and continuous ac tion space, we develop a multi-agent Advantage-Weighted Actor Critic (AWAC)-enhanced Proximal Policy Optimization (PPO) based computation offloading algorithm. The proposed method combines the stable policy updates of PPO with an advantage weighting mechanism inspired by AWAC to emphasize higher quality actions during training. To encourage cooperation without inter-device communication, a shared global reward aligned with system cost is employed. Extensive simulations demonstrate rapid convergence and significant improvement in overall system cost compared with baselines, e.g., Deep Deterministic Policy Gradient (DDPG) and Deep Q-Network (DQN).