To make the most of your techUK website experience, please login or register for your free account here.

29 Oct 2025

by Professor Stefan Vlaski, Professor Kin Leung

Network AI in Practice: Distributed Learning, Edge FL and RL for Next-Gen Networks

Recent advancements in artificial intelligence (AI), particularly in machine learning (ML) techniques, have contributed significantly to the development and optimisation of communication technologies.

These advancements have been applied to various aspects of communication networks, including network control and protocols, teletraffic management, network slicing and edge networks at the network layer. In addition, AI has also been utilised in areas such as MIMO antenna, beamforming, radio resource allocation, and sensing and communications at the physical layer.

Members of the TITAN and HASC teams have been actively involved in developing innovative AI algorithms for future network designs and access. The following highlights some of these AI techniques that will be useful in future communication networks. 

Distributed learning in future intelligent networks

The design and operation of communication networks involve reconciling many potentially competing objectives. These appear across all levels of the network, from resource allocation at the physical layer to traffic management and slicing at the network level.

Classical approaches for reconciling these objectives employ centralised, offline strategies. This means that design occurs once, based on prior assumptions about the state of nature and at a central location in the network. The resulting strategy is then broadcast and implemented for a specified period. This design philosophy allows for strict control over the network’s operation, but limits its ability to adapt and self-organise, introducing bottlenecks that become limiting as the network’s size and its requirements grow.

In contrast, we advocate for design paradigms that are distributed and online, enabling the network to intelligently respond and adapt to changes in the environment, without relying on central coordination. This results in networks that are more resilient, scalable and efficient.

On the other hand, relying on data-driven and decentralised strategies relinquishes control and opens a plethora of research questions. For example, how do we ensure that the resulting strategy is fair and does not neglect certain parts of the network or user base in the interest of the “greater good”? How do we enable the operator to efficiently optimise heterogeneous local objectives? How should information flow over the network to allow fast and efficient decentralised decision-making and operation?

These questions are fundamental to the online and distributed design and operation of communication networks; if successfully addressed, they will enable future networks of networks that are adaptive, resilient and fair.

Federated and split learning for the network edge

Federated learning (FL) is a popular form of distributed learning where data is collected and processed by clients at multiple locations without sharing the raw data. After rounds of model training, the clients forward their updated model parameters to a central controller for aggregation. The controller then broadcasts the aggregated model to all clients for further model training using their local data. This process continues until convergence.

A significant challenge for FL over the network edge is the limited availability of network resources, particularly wireless link capacity and resources on mobile devices. To address this issue, techniques have been developed to optimise the use of such limited resources for the best FL outcome.

One approach to reducing demands for network resources by FL is to remove unimportant parameters from the model. This model pruning technique has been demonstrated to significantly reduce network resource consumption and accelerate convergence without compromising performance.

As the model size increases, learning over the network edge becomes increasingly challenging. One promising solution is to split a large model into two parts, with one part processed by powerful servers deep in the network and the other learned by the resource-limited devices. Recent results have confirmed the effectiveness of such a split FL approach.

Despite these advances, various design issues in FL still require further attention. For example, as the model size continues to grow, how can FL handle the ever-increasing demands for network resources?  How can FL robustly handle data heterogeneity while accommodating personalised learning? And how can FL be decentralised for scalability?

Reinforcement learning for next-generation base-station designs

Reinforcement learning (RL) is a versatile technique which has been successfully applied in many application domains, including communication systems and networks. 

Despite extensive research, numerous open RL issues remain, including multi-agent RL (MARL) and RL with exceptional temporal and spatial dynamics. Distributed MARL is known to be challenging to design for optimal performance, unlike the centralised agent. In communication networks, there are many agents responsible for making action decisions. One notable example is base stations, each of which is responsible for allocating network resources to serve its users. Meanwhile, resource-allocation actions made by the base stations influence one another through co-channel interference.

Physical-layer wireless technologies often require the use of channel status information (CSI), which is typically made available through industrial standardised protocols. However, new technologies may require additional CSI, which can be challenging to obtain through the support of new standards, due to business competition among participating companies in the standards bodies.

To address this issue, HASC and TITAN researchers have developed novel RL techniques for configuring beamforming antennas to transmit from base stations to serve users without requiring CSI. These methods have been shown to achieve nearly optimal results and are being extended to other wireless resource allocation problems with complex temporal and spatial dynamics.

Tales Gaspar

Programme Manager, UK SPF and Satellite, techUK

Usman Ikhlaq

Programme Manager - Artificial Intelligence, techUK

techUK - Seizing the AI Opportunity

The UK is a global leader in AI innovation, development and adoption.

AI has the potential to boost UK GDP by £550 billion by 2035, making adoption an urgent economic priority. techUK and our members are committed to working with the Government to turn the AI Opportunities Action Plan into reality. Together we can ensure the UK seizes the opportunities presented by AI technology and continues to be a world leader in AI development.

Get involved: techUK runs a busy calendar of activities including events, reports, and insights to demonstrate some of the most significant AI opportunities for the UK. Our AI Hub is where you will find details of all upcoming activity. We also send a monthly AI newsletter which you can subscribe to here.

Upcoming AI Events

Latest news and insights

24 Oct 2025

National AI Policy Summit 2025: Julian David OBE Discusses AI Innovation Progress

At the UK National AI Policy, Infrastructure and Skills Summit 2025, our CEO Julian David OBE shared insights on the UK’s AI progress — from major investments and skills initiatives to the ongoing challenges of ensuring innovation remains ethical, inclusive, and responsible.

techUK news and views

21 Oct 2025

AI Growth Labs could accelerate AI innovation and deployment.

Today, the Department for Science, Innovation and Technology (DSIT) announced the launch of the AI Growth Lab, a new proposed framework for a cross-economy sandbox initiative designed to “accelerate innovation and cut bureaucracy in a safe environment.

techUK news and views