Dependent tasks joint scheduling and offloading for edge computing based on deep reinforcement learning

In the mobile edge computing (MEC) scenario, numerous complex applications consist of dependent tasks. Efficient offloading of these applications is essential for reducing latency and minimizing terminal energy consumption. However, existing studies typically employ a decoupled decision-making paradigm, where the task scheduling sequence is predetermined before independently determining the offloading location. This approach separates the scheduling and offloading processes, significantly limiting the exploration of the strategy space and hindering the identification of the global optimal solution. To address these limitations, we propose a joint scheduling and offloading algorithm based on Proximal Policy Optimization (JSO-PPO). We construct an integrated Markov Decision Process (MDP) model and introduce an action masking mechanism, unifying task scheduling and location offloading into a single end-to-end decision. Furthermore, to enhance the algorithm’s performance and stability, the JSO-PPO integrates Deep Dense Architectures in Reinforcement Learning (D2RL) for superior state representation and introduces an adaptive penalty term into its objective function for more stable convergence. Simulation results demonstrate that, compared to multiple existing algorithms, the proposed JSO-PPO achieves significant improvements in minimizing the weighted sum of application finish latency and user equipment energy consumption. These findings validate the efficiency and robustness of our joint optimization paradigm in dynamic and complex edge environments.