Research

Compares five ways of configuring reconfigurable intelligent surfaces for 6G: analytical baselines (phase gradient reflector, focusing lens), gradient-based optimisation, and two learning-assisted designs — a hybrid Mixture-of-Experts and CNN gating. Gradient-based optimisation reached the highest average path gain (about −92 dB), while the hybrid MoE traded a little gain for localised coverage improvements above 40 dB.

Builds a low-cost IoT device that measures particulate matter, carbon monoxide, carbon dioxide and volatile organic compounds, streams readings to the cloud, and forecasts short-term air quality with machine learning. A GRU model performed best — R² above 0.93, latency under 130 ms, and over 91% accuracy on health-based air quality index categories.

Uses ray tracing inside a digital twin to measure how transmitter antenna orientation affects path gain, received signal strength and SINR, then trains k-nearest neighbours, an MLP and XGBoost to predict the best configurations. XGBoost was the most accurate, and the learned predictors found orientation refinements that conventional grid sweeps miss.

Proposes a deep Q-learning scheduler that selects routes and balances load across mobile edge servers of differing capability, so that several user requests can be served at once. Evaluated on realistic mixed-traffic scenarios, the aim is to cut end-to-end delay across cellular and edge networks.

Introduces a MAC-layer scheduling algorithm for LTE-A downlink in two variants: I-MAS for the full-buffer model and R-MAS for realistic incoming traffic. Against round robin and best-CQI baselines, both maximise throughput while sharing resources fairly, and stay robust as user speed increases.