Abstract
Multi-access edge computing (MEC) aims to extend cloud service to the networkedge to reduce network traffic and service latency. A fundamental problem inMEC is how to efficiently offload heterogeneous tasks of mobile applicationsfrom user equipment (UE) to MEC hosts. Recently, many deep reinforcementlearning (DRL) based methods have been proposed to learn offloading policiesthrough interacting with the MEC environment that consists of UE, wirelesschannels, and MEC hosts. However, these methods have weak adaptability to newenvironments because they have low sample efficiency and need full retrainingto learn updated policies for new environments. To overcome this weakness, wepropose a task offloading method based on meta reinforcement learning, whichcan adapt fast to new environments with a small number of gradient updates andsamples. We model mobile applications as Directed Acyclic Graphs (DAGs) and theoffloading policy by a custom sequence-to-sequence (seq2seq) neural network. Toefficiently train the seq2seq network, we propose a method that synergizes thefirst order approximation and clipped surrogate objective. The experimentalresults demonstrate that this new offloading method can reduce the latency byup to 25% compared to three baselines while being able to adapt fast to newenvironments.