Multi-agent reinforcement learning with graph representation for green edge–cloud computation offloading

Edge–Cloud Computing (ECC) stands as a widely adopted distributed computing architecture that facilitates the offloading of computation-intensive tasks from Internet of Things (IoT) devices to edge servers. The growing emphasis on energy conservation and environmental protection raises the concerns of green edge–cloud computation offloading technology. However, conventional computation offloading methods have difficulties in making real-time offloading decisions and adapting to dynamic environmental changes, such as the communication channels. In response to these challenges, we propose a multi-agent reinforcement learning method with graph representation to address the edge–cloud computing offloading schedule problem. Our approach constructs a multi-agent computation offloading reinforcement learning scenario and utilizes graph neural networks to represent the connectivity features between devices and edge servers. Experimental results demonstrate that our proposed method outperforms other algorithms in reducing system energy consumption and response delay. Furthermore, the time-consuming of our approach is significantly shorter compared to heuristic genetic algorithms, with a reduction of 10–20 times.