Online Computation Offloading in NOMA-Based Multi-Access Edge Computing: A Deep Reinforcement Learning Approach

作     者：Nduwayezu, Maurice Quoc-Viet Pham Hwang, Won-Joo 

作者机构：Inje Univ Dept Informat & Commun Syst Gimhae 50834 South Korea Pusan Natl Univ Res Inst Comp Informat & Commun Busan 46241 South Korea Pusan Natl Univ Sch Biomed Convergence Engn Yangsan 50612 South Korea 

出 版 物：《IEEE ACCESS》 (IEEE Access)

年 卷 期：2020年第8卷

页      面：99098-99109页

核心收录：

基　　金：Institute of Information and Communications Technology Planning and Evaluation (IITP) - Korea Government (MSIT) [2020-0-01450] Pusan National University National Research Foundation of Korea [2019R1C1C1006143] Funding Source: Korea Institute of Science & Technology Information (KISTI), National Science & Technology Information Service (NTIS) 

主　　题：NOMA Resource management Task analysis 5G mobile communication Wireless communication Machine learning Edge computing 5G networks deep reinforcement learning (DRL) multi access edge computing (MEC) non-orthogonal multiple access (NOMA) online computation offloading 

摘      要：One of the missions of fifth generation (5G) wireless networks is to provide massive connectivity of the fast growing number of Internet of Things (IoT) devices. To satisfy this mission, non-orthogonal multiple access (NOMA) has been recognized as a promising solution for 5G networks to significantly improve the network capacity. Considered as a booster of IoT devices, and in parallel with the development of NOMA techniques, multi-access edge computing (MEC) is also becoming one of the key emerging technologies for 5G networks. In this paper, with an objective of maximizing the computation rate of an MEC system, we investigate the computation offloading and subcarrier allocation problem in Multi-carrier (MC) NOMA based MEC systems and address it using Deep Reinforcement Learning for Online Computation Offloading (DRLOCO-MNM) algorithm. In particular, the DRLOCO-MNM helps each of the user equipments (UEs) decides between local and remote computation modes, and also assigns the appropriate subcarrier to the UEs in the case of remote computation mode. The DRLOCO-MNM algorithm is especially advantageous over the other machine learning techniques applied on NOMA because it does not require labeled data for training or a complete definition of the channel environment. The DRLOCO-MNM also does avoid the complexity found in many optimization algorithms used to solve channel allocation in existing NOMA related studies. Numerical simulations and comparison with other algorithms show that our proposed module and its algorithm considerably improve the computation rates of MEC systems.