Autonomous Networks by Autonomous Agents

Supervisor: Dr Paul Harvey

School: Engineering

Industry Partner: Leon Wong, Rakuten Mobile, Japan

Description:

Telecommunication networks are the connectivity fabric that allows us to work, play, study, and call for help when we need it. This means that their high-quality continuous operation is essential for all of us.  

In the last 100 years, the way that we use networks has changed a lot – to the point where our toothbrushes, toilets, cars, tvs, and home assistants use them as much as people. All this added pressure means that running these networks is now beyond the human capability to handle thousands of events requiring sub-second response times.  Also, given that telco spends about $273 billion annually to employees its staff, hiring more people is not a practical option without increasing subscription costs.  

In light of this, network operators want their networks to become autonomous networks: networks which can optimise, configure, heal, and adapt by themselves with little or no human involvement.  

In other news, large language models (LLM), like ChatGPT1, have shown that they are capable of producing YouTube scripts, answering exam questions, writing academic papers, and producing software. In this way, they are a meaningful step towards autonomously solving arbitrary problems through the medium of text. Most recently, open-source Autonomous LLM Agents have appeared, with the promise of providing “Goal-driven self-executing software”. 

The goal of this project is to research if an LLM can be used as a first step towards a code implementation of the UN ITU-T’s Autonomous Network Architecture Framework2. Specifically, can Autonomous LLM Agents generate working code for an “evolution controller”: a software application responsible for adapting other software in response to events observed in the network. This will be tested against two resource allocation problems using a simulator, with extension to real hardware if time permits.  Additionally, this project benefits from industry input from a global leader in the field of autonomous networks.  

Concrete outcomes are: 

  • An open-source tool implementing the code generation. 
  • A collection of example generations for the resource allocation problems. 
  • Description of the feasibility of the approach and the design of the tool.