This project investigates the application of agentic AI reasoning to the autonomous control and optimisation of wireless communication networks, leveraging the OCUDU platform in conjunction with an AI-native network controller. The student will extend the existing controller architecture to support closed-loop agentic control, wherein an AI agent autonomously issues configuration commands, observes network responses, and iteratively refines its decisions in pursuit of defined performance objectives.
A central focus of the project is the safe deployment of agentic AI in a live radio access network environment. The AI-native network controller exposes a set of configurable safety guardrails designed to constrain the action space of the agent and prevent unsafe or destabilising configurations. The student will exploit these guardrails to design and implement a structured control loop that incorporates configuration checkpointing, enabling the system to save, evaluate, and if necessary restore prior network states. This rollback capability is essential for safe experimentation in physical testbeds and forms a key contribution of the project.
The student will conduct a systematic experiments using real networking equipment in the testbed, evaluating the agent’s ability to autonomously optimise radio resource management parameters under varying traffic and channel conditions, while remaining within safety boundaries enforced by the controller. The project will produce both a working implementation and an empirical analysis of the trade-offs between autonomy, performance, and safety in agentic network control.