Deep reinforcement learning for partial offloading with reliability guarantees
Ji Li, Zewei Chen, Xin Liu · 2021
Edge computing is a new paradigm that supporting resource-constrained edge devices to better complete tasks. Making offloading decisions for edge devices plays a crucial role in edge computing. In this paper, we consider the online offloading decision for randomly arriving delay-sensitive tasks in the case of a time-varying channel and queue system. Under the constraints of delay and transmission reliability, we formulate a task offloading problem to maximize the number of successfully processed tasks. In order to solve the problem, we first model the task management of user as a queue model, which makes our system closer to the reality. Then we formulate the problem as a Markov decision process (MDP). Finally, we incorporate twin delayed deep deterministic policy gradient (TD3) algorithm and long short-term memory (LSTM) and propose a deep reinforcement learning (DRL) based distributed algorithm to solve the MDP. The simulation results show that compared with other advanced algorithms, our proposed method has the smallest dropped rate of the offloading system under the condition of poor channel conditions and the task has deadline constraints.