Artificial Intelligence for Network Management in Open RAN Edges

Abstract

The mantra of the International Telecommunication Union’s (ITU) International Mobile
Telecommunications (IMT) 2030 Framework on ‘Ubiquity, Intelligence, Sustainability’
delineates the development directions of future networks. In achieving these ambitions,
Multi-access edge computing (MEC) and the Open Radio Access Network (RAN) serve as complementary pillars that push intelligence to the network edge and optimize connectivity for
sustainable operations. MEC extends computational capabilities to the edge of the network, reducing latency and enabling real-time, intelligent applications, thus contributing to the ubiquity
and intelligence of the network. Meanwhile, Open RAN is essential for robust communication
management between user devices and the network core, enhancing both the performance and
sustainability of future networks.
This thesis focuses on three essential use cases in Open RAN Edges, including baseband
function placement, task offloading, and network slicing, that are indispensable for enabling
end-to-end communication. Baseband function placement is responsible for establishing communication links within the RAN and dynamically adapting to changing network conditions in
real-time. Task offloading and network slicing leverage this connectivity to enable computational
support and service differentiation across a wide range of applications. However, these use cases
also reveal intricate challenges in orchestrating distributed resources under dynamic network
conditions, where latency, bandwidth, and energy constraints must be balanced alongside user
mobility and diverse QoS requirements, all while controlling operational costs and ensuring
system-wide adaptability. Conventional heuristic and optimization-based methods often fall short
in these contexts, due to limited adaptability to dynamic environments and difficulty scaling to
high-dimensional problem spaces. This has intensified the focus on AI-driven approaches, which
can learn from historical patterns and adapt to evolving network states. Accordingly, this thesis
designs various AI-based management algorithms for the three use cases, focusing on various
network performance metrics such as reducing user energy consumption and service latency,
increasing edge server resource utilization, and lowering overall operational energy consumption.
In the first use case, a Deep Reinforcement Learning (DRL)-based algorithm is proposed
to activate only the necessary servers, thereby minimizing network energy consumption. This
method is then combined with a heuristic to optimize baseband function placement and routing
provisioning, thereby ensuring that the latency requirements of incoming service requests are
consistently satisfied. Building on the same scenario, we further consider another case where
servers remain active. Beyond optimizing baseband function placement to reduce power consumption, the proposed solution, NetMind+, leverages a Graph Convolutional Neural Network
(GCN) encoder and an adaptive layer-wise incremental learning mechanism to establish a general DRL model capable of handling various static network topologies, while minimizing the
re-optimization overhead under dynamic network conditions. In the task offloading scenario,
we consider a multi-objective optimization problem that balances user energy consumption and latency according to request requirements. This service optimization is formulated as a
Mixed Integer Nonlinear Programming (MINLP) problem and divided into two sub-problems
for tractability. First, multi-agent deep reinforcement learning (MADRL) is employed to allocate
multi-server resources under a long-term optimization objective. Next, an optimization-based
approach manages offloading decisions and transmission power allocation, maximizing the utilization of DRL-allocated resources. Finally, in the network slicing use case, a MADRL algorithm
is adopted to model the resource competition among network slices and handle resource allocation.
An incremental learning approach is introduced to improve the generalization performance of
the algorithm, accommodate changes in the number of slices and reduce retraining time in new
scenarios by leveraging previously acquired knowledge. An urban-wide testing infrastructure is
built based on Open MANO and Kubernetes to validate the proposed solution.
While the aptitude of AI for evolving 5G networks is well demonstrated through these
applications, several challenges remain beyond the lack of generalization capability. Therefore,
this thesis concludes by summarizing the likely obstacles that need to be addressed to transition
AI from conceptual hype to practical application, and by outlining future research directions
along with potential solution pathways.

Date of Award	30 Sept 2025
Original language	English
Awarding Institution	University of Bristol
Supervisor	Shuangyi Yan (Supervisor) & Dimitra Simeonidou (Supervisor)