DEEP LEARNING–DRIVEN SIGNAL OPTIMIZATION FOR 6G WIRELESS SYSTEMS: MODELS, METHODS, PERFORMANCE EVALUATION

https://doi.org/10.5281/zenodo.17492225

Abstract

The sixth generation (6G) of wireless networks is envisioned to deliver unprecedented capabilities in terms of spectral efficiency, ultra-low latency, reliability, and energy efficiency. Meeting these requirements requires fundamentally new approaches to physical-layer signal optimization, as traditional convex and iterative optimization methods become computationally prohibitive in large-scale, dynamic environments. Deep learning (DL) has emerged as a promising tool to address these challenges, offering universal approximation, rapid inference, and the ability to integrate domain knowledge into optimization frameworks. Recent advances span supervised learning for beamforming, deep unfolding of iterative algorithms such as WMMSE, reinforcement learning for power control and interference management, and DL-based phase optimization for reconfigurable intelligent surfaces (RIS). DL also enables joint end-to-end optimization across beamforming, power allocation, and RIS configurations, while knowledge-driven paradigms embed communication-theoretic constraints to enhance interpretability and generalization. However, challenges remain in terms of robustness to imperfect channel state information, constraint enforcement, scalability, and deployment feasibility. This paper surveys and synthesizes DL-driven signal optimization models, methods, and performance evaluations, providing a comprehensive perspective on their role in shaping practical and efficient 6G wireless systems.