Derin öğrenme, yeniden yapilandirilabilir akilli yüzeylerle buluşuyor: 6G ve ötesine doğru yeni sistem tasarimlari ve çözümleri

Sixth-generation (6G) wireless networks are envisioned to meet unprecedented demands in spectral efficiency, ultra-low latency, high reliability, and pervasive connectivity. Beyond conventional transceiver design and spectrum expansion, intelligent control of the radio environment has emerged as a crucial enabler, with Reconfigurable Intelligent Surfaces (RIS) offering low-cost and energy-efficient means to manipulate propagation. At the same time, data-driven methods such as Machine Learning (ML) and Deep Learning (DL) provide powerful tools for modeling, optimization, and adaptive decision-making under uncertainty. The convergence of RIS and DL therefore represents a transformative paradigm for next-generation wireless communication, sensing, and localization. This thesis makes several key contributions. First, a comprehensive survey of RIS-assisted communication systems and their integration with DL techniques is presented, providing an overview of the state-of-the-art and identifying open research challenges across resource allocation, beamforming, localization, and sensing. Building upon this foundation, a supervised learning-assisted partitioning framework is introduced for uplink Non-Orthogonal Multiple Access (NOMA), where RIS elements are intelligently distributed to users. The scheme achieves comparable ergodic sum rate and fairness to advanced iterative algorithms while significantly reducing computational overhead, with theoretical outage probability analysis and BER results validating performance under diverse quality-of-service and SIC conditions. To address dynamic scenarios, a Deep Reinforcement Learning (DRL) approach is then incorporated, enabling online partitioning of RIS elements without retraining. The proposed DRL framework attains near-optimal achievable rates and sustained fairness while reducing computational time by up to 90\%, highlighting its scalability and adaptability in time-varying environments. The third contribution introduces a novel RIS partitioning scheme that jointly addresses RIS identification (RIS-ID) and beamforming, allocating elements between detection and communication tasks. Simulation and theoretical results confirm that the scheme enhances SNR while preserving reliable identification accuracy, thereby establishing a unified trade-off between sensing and communication. Finally, a DL-based localization framework is proposed, leveraging Sionna-based ray-traced channels and optimized RIS phase configurations. The two-stage design achieves a mean localization error of 3.06 m, significantly improving upon the non-optimized RIS case (4.05 m) and the RSS-based baseline (5.80 m). Moreover, the optimized scheme enhances fine-grained accuracy, raising the probability of sub-meter localization from 1.21\% to 17.72\% of the available receiver positions. These results confirm the effectiveness of RIS-assisted learning-based localization in multipath-rich non-line-of-sight environments. Collectively, these contributions establish a unified RIS–DL co-design framework that advances communication, control, identification, and localization in realistic wireless environments. The presented studies provide both theoretical insights and empirical evidence that RIS, when augmented with data-driven optimization, holds the potential to transform the architecture of 6G networks, enabling intelligent, adaptive, and efficient wireless systems.