Intelligent Reconfigurable Intelligent Surface Control Using Machine Learning for Future Wireless Communications

Reconfigurable intelligent surfaces (RIS) have emerged as a key enabler for beyond-5G and 6G wireless networks by enabling programmable propagation environments. However, realizing these gains in practice requires joint optimization of base-station precoding, RIS configuration, power allocation, and channel estimation under hardware and feedback constraints. This dissertation develops learning-based methods to address these challenges in different RIS-assisted communication scenarios. The first part of this dissertation deals with system-level optimization. A three-dimensional RIS-assisted MISO downlink is formulated as a reinforcement learning problem in which a Randomized Ensembled Double Q-learning (REDQ) agent is trained and learns how to adjust the base-station beamforming vectors and RIS phase shifts to maximize the spectral efficiency of multiple users under different RIS elements and wireless signal power levels. Simulation results indicate consistent gains in average spectral efficiency over well-known model-free, off-policy baselines, including Deep Deterministic Policy Gradient (DDPG), Twin Delayed Deep Deterministic Policy Gradient (TD3), and Soft Actor-Critic (SAC). The same idea is then extended to a wideband OFDM RIS-aided NOMA setting with more realistic assumptions: a beyond-diagonal RIS with finite phase resolution, delayed and noisy CSI feedback, and explicit checks on the feasibility of successive interference cancellation. For this scenario, an OFE–REDQ agent is introduced that jointly adapts the RIS configuration and the inter- and intra-cluster power allocation, leading to improved sum rate, faster and more stable learning, and more reliable SIC compared with competing schemes such as REDQ and DDPG. The second part of the dissertation focuses on measurement-based modelling and hardware efficiency. Using real sub-6 GHz RIS measurements collected in anechoic chamber, a one-dimensional convolutional neural network (1D-CNN) is proposed and trained to map beam-steering angles to received power using k-fold cross-validation, to outperform standard feed-forward and recurrent networks such as Long Short Term Memory (LSTM), Gated Recurrent Unit (GRU) and Fully Connected Feedforward Neural Networks (FCN). This is followed by the Row-Vector Transformer (RVT), which treats each sample as a single dense token so that all features are processed jointly; RVT achieves lower prediction errors and more stable training than 1D-CNN, LSTM, GRU, FCN, and Feature Tokenizer (FT)-Transformer across several datasets. To link prediction with actuation, a multi-task architecture based on Symmetric-PCGrad is proposed that simultaneously selects the active group of RIS elements and predicts the corresponding received power, allowing the system to meet quality-of-service targets with a reduced number of active elements. Finally, a sub-terahertz drone scenario is investigated, in which a multi-layer LSTM is trained for channel estimation in a Flying-RIS link supported by a ground base station, and is found to outperform GRU-, CNN-, and fully connected baselines. In summary, the dissertation develops and evaluates a set of learning-based tools for intelligent RIS control, covering system optimization, data-driven power prediction, energy-aware element activation, and Flying-RIS channel estimation. The proposed methods improve spectral efficiency, reduce operational overhead, and increase robustness, and they provide concrete design guidance for future RIS-enabled wireless networks.