Research on Dependent Task Offloading Based on Deep Reinforcement Learning

The dependent task offloading in mobile edge computing (MEC) is an NP-hard problem. The existing solutions suffer from optimizing only a single objective, ignoring inter-task dependencies, poor dynamic adaptation and slow convergence. In this paper, we propose a time consumption table based dependent task offloading algorithm (TBDTOA), which aims to jointly optimize the product of task latency and energy consumption by predicting the impact of offloading decisions to achieve a more efficient task offloading strategy. First, the dependent task offloading problem is transformed into an optimization problem under the Markov decision process (MDP). Second, the task offloading strategy is efficiently solved using a deep reinforcement learning technique based on time consumption tables. Finally, experimental results show that the algorithm outperforms the other three baseline schemes in terms of latency, energy consumption and cost.