Coordinated Offloading: Meta-Guided Multi-Agent Reinforcement Learning for Edge-Cloud Continuum

The increasing demand for intelligent and low latency services in edge-cloud continuum systems poses new
challenges for dynamic and efficient task offloading. In this paper, we propose a meta-guided multi-agent reinforcement learning (MARL) framework for distributed task offloading under partial observability. The problem is formulated as a Markov Decision Process (MDP), a solution where each device independently learns an optimal policy under partially observable system states. To address the lack of global coordination in decentralized settings, a lightweight meta-controller is introduced to observe system-level metrics and periodically broadcast a utility signal. This signal is incorporated via reward shaping to guide agent learning without introducing additional communication overhead. The proposed method offers a promising direction for achieving balanced decision-making and scalable coordination in complex edge-cloud environments.