Competitive Algorithms and Reinforcement Learning for NOMA in IoT Networks

This paper studies the problem of massive Internet of things (IoT) access in beyond fifth generation (B5G) networks using non-orthogonal multiple access (NOMA) technique. The problem involves massive IoT devices grouping and power allocation in order to respect the low latency as well as the limited operating energy of the IoT devices. The considered objective function, maximizing the number of successfully received IoT packets, is different from the classical sum-rate-related objective functions. The problem is first divided into multiple NOMA grouping subproblems. Then, using competitive analysis, an efficient online competitive algorithm (CA) is proposed to solve each subproblem. Next, to solve the power allocation problem, we propose a new reinforcement learning (RL) framework in which a RL agent learns to use the CA as a black box and combines the obtained solutions to each subproblem to determine the power allocation for each NOMA group. Our simulations results reveal that the proposed innovative RL framework outperforms deep-Q-learning methods and is close-to-optimal.