Deep Reinforcement Learning-Based Computation Offloading and Optimal Resource Allocation in Industrial Internet of Things with NOMA

Multi-access mobile edge computing is considered as an important way to enable computationally intensive but latency-sensitive services in the future industrial IoT. In this paper, we consider a multi-access MEC network. Powered by WPT, each IID (IIoT devices) collects energy and then follows a binary offload policy, i.e., either executes the task locally or offloads to the MEC server for execution via a non-orthogonal multiple access (NOMA) approach. Where the channel gain from each IID (IoT device) to the edge computing server is time-varying. To address the goal of maximizing the total computation rate of all IIDs by jointly optimizing the selection of computation modes and the allocation of transmission time in dynamic channel scenarios, we propose a Deep Reinforcement Learning (DRL)-based computation offloading and resource allocation optimization method (DCOORA). The method is able to obtain near-optimal solutions under timevarying channels and, moreover, can solve the problem with very low computational complexity. Simulation results show that the hybrid framework based on online offloading of DRL obtains near-optimal computational performance for different network settings. In addition, it has very low computation time, which is particularly suitable for time-varying channel environments.