Environment-Aware Deep Reinforcement Learning Framework for Adaptive Task Offloading in Fog Networks
R. Rajarajan, Mostafha Alwbaidy, Rajasekar Deepa, D. Kesavan, Srinivas Samala · 2025
In recent years, fog and edge computing have become crucial for bringing computation, storage, and networking closer to end devices to reduce latency and improve real-time processing. However, traditional methods such as Hybrid Offloading fail in overloaded or structurally inflexible settings. To overcome this, an Environment-Aware Deep Reinforcement Learning-based Offloading (DRL-off) framework was proposed. Initially, a DRL-based agent embedded in each fog node independently learns the optimal offloading policies using state observations, including resource usage, latency, and neighbor availability. Furthermore, the agent dynamically decides to execute tasks locally, offload to nearby nodes, or escalate them to the cloud. A reward function then balances latency, energy, and successful execution, whereas a two-tier safety mechanism preserves real-time execution for time-sensitive applications. Subsequently, a gossip-based communication mechanism is employed to enable lightweight coordination among fog nodes, in which nodes periodically exchange minimal state information with their neighbor nodes. Finally, the experimental results showed that DRL-Off consistently outperformed baseline models, such as cloud offloading, by achieving a task success rate (92.3 %) and a reduction in average latency (43.2 %) across diverse environments.