Reinforcement Learning for User Clustering in NOMA-enabled Uplink IoT

The model-driven algorithms have been investigated in wireless communications for decades. Presently, the model-free methods based on machine learning techniques are rapidly being developed in the field of non-orthogonal multiple access (NOMA) to dynamically optimize multiples parameters (e.g., number of resource blocks and QoS). With the aid of SARSA Q-learning and Deep reinforcement Learning (DRL), in this paper we proposed a user clustering based resource allocation with uplink NOMA techniques in multi-cell systems. It performs user grouping based on network traffic to efficiently utilise the available resources, we apply SARSA Q-learning to light and DRL to heavy network traffic. To characterize the performance of the proposed optimization algorithms, achieved the capacity for all the users is used to define the reward function. The proposed SARSA Q-learning and DRL algorithms are capable of assisting base-stations to efficiently assign available resources to IoT users considering different traffic conditions. As a result, simulation outcomes show that both the algorithms, SARSA Q-learning and DRL performed better than orthogonal multiple access (OMA) in all the experiments and converged with maximum sum-rate.