DDPG-Driven Multi-Layer Task Offloading: Edge–Fog Collaboration and Learning Stability

The proliferation of Internet of Things (IoT) devices and the deployment of 5G/6G networks have made task offloading in multi-layer edge–fog–cloud systems increasingly challenging, especially under high device density. This paper proposes a Deep Deterministic Policy Gradient (DDPG)-driven multi-layer task offloading scheme that jointly optimizes state representation and edge–fog collaboration. Specifically, a 13-dimensional continuous state vector is constructed to capture task features, device states, and global resources, while a fog-assisted interaction mechanism is introduced to reduce local bias and improve coordination efficiency. A dynamic exploration strategy is combined with Prioritized Experience Replay (PER) to enhance sampling efficiency and stabilize training. Simulation results demonstrate that the proposed scheme achieves higher task success rates, lower energy consumption, and reduced waiting and wasted time compared with baseline algorithms, while ensuring stable learning under varying device densities.