The controller cannot optimize one layer at a time
Programmable radio access networks are often described through their interfaces and software components. Maria Tsampazi’s dissertation frames the harder question: what happens when slicing, scheduling, link adaptation, propagation control, power allocation, and spectrum sharing all become parts of the same feedback system? The public abstract presents a cross-layer program built around an open, disaggregated RAN whose control loops can react to traffic, channel, and service conditions.
The work spans several control horizons. Deep reinforcement learning is used for network slicing and scheduling, while a framework called PandORA automates the design, training, and deployment of learning-based Open RAN applications on a wireless emulator. Online reinforcement learning addresses link adaptation, where decisions must track a changing radio link rather than a fixed offline dataset. At the physical layer, the dissertation models reconfigurable intelligent surface channels and optimizes resources across spectrum bands. It also studies cellular and non-terrestrial spectrum sharing through power control and beamforming.
Cross-layer gains need system-level validation
This breadth matters because an improvement at one layer can simply move the bottleneck. A scheduler may raise throughput while increasing energy use; a reconfigured propagation path may help one service class but change interference elsewhere; a terrestrial–satellite sharing policy may depend on beam and power decisions made on different timescales. Treating these components as a coordinated system makes the interfaces between algorithms part of the research contribution.
The validation path is also notable. The abstract describes simulation, hardware-in-the-loop emulation, digital twins, and over-the-air 5G experiments. That progression is more informative than a single simulation environment because it tests whether a control idea survives increasingly realistic timing and integration constraints. The public abstract, however, does not provide a common quantitative benchmark across every contribution, so it should be read as a map of a dissertation’s validated systems work rather than as evidence that one controller dominates every alternative.
The practical lesson is architectural: AI-native wireless control needs explicit boundaries, observability, and validation at the points where layers meet. Intelligence is not added by placing a learned policy beside the RAN; it emerges only when the control loop, resource constraints, and deployment path are designed together.
Research notes
Cross-Layer Optimization and System-Level Design of Next-Generation Wireless Networks via Intelligent RAN Control
- Author: Maria Tsampazi
- Public record: arXiv
- What is established: The public abstract connects Open RAN control, deep reinforcement learning, online link adaptation, RIS-assisted resource allocation, and terrestrial–non-terrestrial spectrum sharing with multiple levels of experimental validation.
- Read with care: The abstract summarizes a dissertation-wide portfolio and does not expose one shared benchmark or a single numerical comparison across all contributions.