Dynamic User Clustering for MIMO-NOMA Network: Deep Reinforcement Learning Approach

To enable the base station to serve more users, we implement multiple-input multiple-output non-orthogonal multiple access, which is capable of serving multiple users with a single beam. Therefore, it is crucial to group users according to their channel correlation before performing beamforming. Heuristic methods are commonly used, as the user clustering problem is NP-hard. However, these methods do not sufficiently utilize the information obtained through exploration. We use discrete soft actor-critic deep reinforcement learning for user clustering to achieve better performance. The performance of the heuristic method is affected by the predefined number of clusters as well, which makes it impractical. To address this issue, we modify the output of the clustering algorithm with the probability of a user belonging to a cluster and introduce several virtual clusters to generate empty clusters. This enhancement allows certain clusters to remain unselected by any user, leading to the formation of empty clusters, and enabling dynamic adjustment of the number of clusters based on real-time network conditions. Simulation results have demonstrated that our proposed dynamic clustering algorithm outperforms traditional clustering methods and leads to the optimal number of clusters.