Meta-MADDPG: Achieving Transfer-Enhanced MEC Scheduling via Meta Reinforcement Learning

Abstract

With the assistance of mobile edge computing (MEC), mobile devices (MDs) can optionally offload local computationally heave tasks to edge servers that are generally deployed at the edge of networks. As thus, the latency of task and energy consumption of MDs can be both reduced, significantly improving mobile users’ quality of experience. Although considerable MEC scheduling algorithms have been designed by researchers, most of them are trained to solve specific tasks, leaving the performance in other MEC environments remaining dubious. To address the issue, this paper first formulates the optimization problem to minimize both task delay and energy consumption, and then transforms it into Markov decision problem that is further solved by using the state-of-the-art multi-agent deep reinforcement learning method, i.e., MADDPG. Furthermore, aiming at improving the overall performance in various MEC environments, we integrate MADDPG with meta-learning and propose Meta-MADDPG which is carefully designed with dedicated reward functions. The evaluation results are given to showcase the more satisfactory performances of Meta-MADDPG over the state-of-the-art algorithms when confronting new environments.