UAV-assisted dependency-aware computation offloading in device-edge-cloud collaborative computing based on improved actor-critic DRL

Unmanned aerial vehicle (UAV)-assisted mobile edge computing (MEC) has become a popular research topic, addressing challenges posed by the pressure of cloud computing and the limited service scope of MEC. However, the limited computing resources of UAVs and the data dependency of specific tasks hinder the practical implementation of efficient computational offloading (CO). Accordingly, a device-edge-cloud collaborative computing model is proposed in this study to provide complementary offloading services. This model considers stochastic movement and channel obstacles, representing the dependency relationships as a directed acyclic graph. An optimization problem is formulated to simultaneously optimize system costs (i.e., delay and energy consumption) and UAV endurance, taking into account resource and task-dependent constraints. Additionally, a saturated training SAC-based UAV-assisted dependency-aware computation offloading algorithm (STS-UDCO) is developed. STS-UDCO learns the entropy and value of the CO policy to efficiently approximate the optimal solution. The adaptive saturation training rule proposed in STS-UDCO dynamically controls the update frequency of the critic based on the current fitted state to enhance training stability. Finally, extensive experiments demonstrate that STS-UDCO achieves superior convergence and stability, while also reducing the system total cost and convergence speed by at least 11.83% and 39.10%, respectively, compared with other advanced algorithms.