GTPO: A Stabilizing Task Offloading Technique Based on GTrXL and PPO in IoT Edge Computing Environments

Resource-intensive applications in the Internet of Things are becoming increasingly popular due to user demand. To enhance system processing capabilities, these applications can perform offloading in the edge computing environment. However, this environment faces challenges due to numerous dynamic that can impact the task processing based on these applications. To address the task offloading problem, we first modeled these dependent tasks as dependent Directed Acyclic Graph (DAG) tasks. Next, we built a simulation of the edge computing environment to reflect the dynamic real-world environment. Finally, we proposed an algorithm called GTPO based on deep reinforcement learning, combining the enhanced Gated Transformer-XL (GTrXL) and Proximal Policy Optimization (PPO). Our experimental results showed that GTPO showed excellent stability, adaptability and achieved the lowest average execution latency time compared to other baselines when making task offloading decisions.