Introduction
As wireless networks evolve toward 5G and beyond, achieving high spectrum efficiency remains a primary challenge. Two technologies have emerged as frontrunners: Intelligent Reflecting Surfaces (IRS) and Non-Orthogonal Multiple Access (NOMA). IRS utilizes programmable passive elements to steer signals around obstacles, while NOMA allows multiple users to share the same spectrum by superimposing signals in the power domain. This paper proposes a machine learning framework to jointly optimize these technologies in a dynamic environment where user mobility complicates channel estimation and resource allocation.
System Model and Problem Statement
The research focuses on an IRS-aided Multiple-Input Single-Output (MISO) NOMA network. In this setup, a Base Station (BS) serves multiple mobile users grouped into clusters. The presence of physical obstacles often blocks the direct line-of-sight (LoS) path between the BS and users, making the IRS critical for establishing alternative reflective paths.
The core difficulty lies in the high coupling of variables: optimizing the IRS phase shift matrix, user clustering, and power allocation simultaneously is a non-convex, dynamic problem. Traditional mathematical optimization often fails to provide real-time solutions as user mobility constantly shifts the optimal configuration.
Proposed Machine Learning Framework
To address these complexities, the authors introduce a "divide and conquer" approach utilizing three distinct machine learning models:
- Mobility Prediction (LSTM): A Long Short-Term Memory (LSTM) network predicts future user positions based on historical coordinate data. This proactive approach ensures the system adapts to movement rather than merely reacting to it.
- Dynamic User Clustering (K-GMM): Using predicted positions and Channel State Information (CSI), a K-means-based Gaussian Mixture Model (K-GMM) partitions users into clusters to facilitate efficient NOMA operation.
- Resource Allocation (DQN): A Deep Q-Network (DQN) serves as the central agent. It observes the network state and determines the optimal IRS phase shifts and power allocation to maximize the total sum rate.
Key Results and Performance
The proposed framework demonstrates significant performance gains over traditional methods. A key finding is the superiority of the NOMA scheme over Orthogonal Multiple Access (OMA).
As shown in the evaluation, the IRS-NOMA system consistently outperforms IRS-OMA across various transmit power levels. Specifically, at 20 dBm, the NOMA approach yields a 44.99% higher sum rate, maintaining a nearly 40% advantage at higher power levels. Furthermore, the study highlights that increasing the number of IRS reflecting elements enhances the sum rate, although the marginal utility decreases as the array size grows.
Conclusion and Future Outlook
By integrating mobility prediction with reinforcement learning, this framework effectively manages the complexities of IRS-NOMA networks. While the current study relies on simulated environments and assumes perfect CSI, it provides a robust foundation for future real-world implementations. The results prove that intelligent, machine-learning-driven tuning of IRS elements is essential for unlocking the full potential of next-generation wireless spectrum efficiency.
Summary updated