A Deep Reinforcement Learning Approach for Dependent Task Offloading in Multi-Access Edge Computing

The resource-constrained edge devices are struggling with the rising computational demands of modern user applications with many dependent tasks. In this paper, we investigate the dependent task offloading strategy by constructing graph theory for the dependent tasks so that the explicit priority relations based on deep reinforcement learning (DRL) are drawn. In particular, we propose a novel task-priority-transformer dependent task offloading (TPTDTO) method with the refined priority features of dependent tasks and the application context information. We first construct a priority representation for the dependent tasks which enables the efficient task offloading and energy consumption reduction. To improve the intelligent offloading decision efficiency, a proximal policy optimization (PPO) method is utilized to interact with the multi-access edge computing (MEC) system. Specifically, we leverage the enhanced representation capability of the transformer model to encode the dynamic MEC application topology and utilize a deep neural network to capture the task priority features. Moreover, the transformer model is adopted to address the scalability issue. Simulation results demonstrate that the proposed algorithm can achieve competitive performance compared with that of the existing baselines for dependent task offloading.