Reinforcement Learning Based Matching for Computation Offloading in D2D Communications

As a promising technology for efficient use of computing and storage resources, mobile edge computing (MEC) has attracted much attention for its utilization of data and computation resources close to resource-poor mobile devices. However, the wireless channel states between mobile devices vary rapidly and are difficult to predict, which makes it difficult to determine the computation offloading strategy. In this paper, we propose a parallel reinforcement learning based computation offloading (RLCO) scheme, which enables offloading nodes to learn the computation offloading strategy distributedly based on the historical channel states. In our proposed RLCO, each offloading node is considered as an agent to learn how to satisfy the latency of computation and transmission and the SNR of both offloading node and cooperative node while minimizing the energy consumption based on Q-learning. After training, the Q-values of each agent are used to form their preference lists used in the matching stage. Simulation results demonstrate that the proposed RLCO has a good performance in reducing the system energy consumption.