Real-Time Downlink Resource Allocation in NOMA Systems: A Reinforcement Learning Approach

Non-orthogonal multiple access (NOMA) is an important multiple access technology for next generation wireless communications. This work focuses on realtime resource allocation in NOMA systems based on reinforcement learning (RL). Q-learning (QL) is an agile RL approach that can adjust its learning strategy to dynamic channel state, making it a perfect machine learning algorithm that can tell agents what to do to maximize its rewards. It does not require a given model of the environment and thus can work adaptively in different scenarios. However, the majority of existing works treated QL as a tool to solve an optimization problem. In this work, QL participates the entire resource allocation process in NOMA systems. As long as user equipment (UE) locations are given, it can optimize resource allocation to achieve a maximum sum rate. In particular, we demonstrate its effectiveness with simulation results in three NOMA schemes, including multi-user superposition transmission (MUST), pattern division multiple access (PDMA), and sparse code multiple access (SCMA) systems. The excellent tracking convergence property of the proposed schemes makes it an ideal choice to perform realtime resource allocation in wireless communications.