Towards Energy-Sustainable and Fair 6G: A Hybrid Learning Approach for IRS-Assisted MISO-NOMA Systems

The integration of intelligent reflecting surfaces (IRS) with non-orthogonal multiple access (NOMA) has emerged as a promising paradigm for 6G wireless systems, enabling spectral efficiency, user fairness, and energy sustainability. This paper presents a theoretical machine learning (ML) framework for resource allocation in IRS-assisted multiple-input single-output NOMA (MISO-NOMA) networks. The system model incorporates an IRS phase-shift matrix and user-specific quality-of-service (QoS) constraints, and the optimization objective is to maximize the weighted sum-rate subject to transmit power, discrete IRS phase resolution, and minimum signal-to-interference-plus-noise ratio (SINR) requirements. To address the non-convexity of the joint beamforming, power allocation, and phase-shift optimization, we propose a hybrid supervised–reinforcement learning approach, where offline supervised pre-training provides near-optimal initialization and online reinforcement learning ensures adaptive refinement under dynamic channel conditions. Theoretical analysis demonstrates significant gains in energy efficiency (up to 28%), Jain’s fairness index (> 0.92), and computational latency reduction (sub-millisecond inference) compared to conventional semidefinite relaxation (SDR) and successive convex approximation (SCA) methods. These results confirm that the proposed ML-based framework not only approaches optimal sum-rate performance but also scales efficiently with large IRS deployments, making it suitable for ultra-reliable low-latency communication (URLLC) and energy-conscious 6G applications.