NOMA Assisted Multi-Task Multi-Access Mobile Edge Computing via Deep Reinforcement Learning for Industrial Internet of Things

Multiaccess mobile edge computing (MA-MEC) has been envisioned as one of the key approaches for enabling computation-intensive yet delay-sensitive services in future industrial Internet of Things (IoT). In this article, we exploit nonorthogonal multiple access (NOMA) for computation offloading in MA-MEC and propose a joint optimization of the multiaccess multitask computation offloading, NOMA transmission, and computation-resource allocation, with the objective of minimizing the total energy consumption of IoT device to complete its tasks subject to the required latency limit. We first focus on a static channel scenario and propose a distributed algorithm to solve the joint optimization problem by identifying the layered structure of the formulated nonconvex problem. Furthermore, we consider a dynamic channel scenario in which the channel power gains from the IoT device to the edge-computing servers are time varying. To tackle with the difficulty due to the huge number of different channel realizations in the dynamic scenario, we propose an online algorithm, which is based on deep reinforcement learning (DRL), to efficiently learn the near-optimal offloading solutions for the time-varying channel realizations. Numerical results are provided to validate our distributed algorithm for the static channel scenario and the DRL-based online algorithm for the dynamic channel scenario. We also demonstrate the advantage of the NOMA assisted multitask MA-MEC against conventional orthogonal multiple access scheme under both static and dynamic channels.

I. System Model and Problem Formulation

The growing exploitation of Internet of Things (IoT) in industrial section has yielded a variety of computation-intensive yet delay-sensitive applications, e.g., unmanned warehouse and unmanned factories. However, due to the cost and hardware issues, conventional IoT devices (IoTDs) are usually equipped with limited computation resources, leading to a poor performance when running the computation-intensive tasks. Thanks to the recent advances in multiaccess radio networks, the paradigm of multiaccess mobile edge computing (MA-MEC) has provided a promising approach to address this issue [1]. With MA-MEC, an IoTD can offload part of its computation tasks to several nearby edge-computing servers (ECSs) equipped with a sufficient amount of computation resources, which effectively reduces the latency in completing the tasks. The advantage of MA-MEC and its applications have attracted a lot of research interests [2]–[7]. In particular, the joint management of communication and computation resources plays a crucial role to the performance of computation offloading, e.g., energy efficiency [8]–[12]. Thanks to the recent advances in deep learning (DL), many research efforts have been devoted to exploiting DL for computation offloading [16]–[24].