Adaptive Neural Network Optimization for Fully Connected STAR-RIS-Aided NOMA Networks

In this paper, we examine the potential of fully connected simultaneously transmitting and reflecting reconfigurable intelligent surfaces (STAR-RIS) to enhance future sixth-generation (6G) networks, focusing on their integration with non-orthogonal multiple access (NOMA). By supporting multiple users exploiting the same frequency band, NOMA, when combined with STAR-RIS, can significantly improve both spectral efficiency and reliability. We propose a practical machine learning framework, namely adaptive neural network optimization (ANNO) to effectively configure STAR-RIS in complex, multi-solution landscapes. The ANNO framework leverages the adaptive features of neural networks, enabling dynamic responses to varying channel conditions and user requirements. Unlike conventional regularization approaches, we design the ANNO to minimize overfitting while maintaining accuracy. The numerical results demonstrate that the ANNO framework achieves substantial improvements in convergence speed and spectral efficiency compared to traditional methods, particularly in dense urban environments.