AI will undoubtedly have a significant social and economic impact in the UK. A recent WIPO analysis shows that many sectors and industries are exploring the commercial exploitation of AI, including telecommunications which is among the top industries in the fields of AI application.
The early use of AI for telecom has been limited to, e.g., business intelligence, and chat bots for customer services. As the 5G network becomes more virtualised and softwarised, enabling AI solutions to automate service provision and optimise network operation and management, has become promising. Many early work on the use of such operational level AI in the network has shown significant advantages of both enhancing network efficiency and quality of service, and reducing network operational cost (OPEX).
However, like in many other industries, the full adoption of AI for network operation and management has some fundamental barriers. The major ones at the moment, from the technical perspective, are data, reliability, and deployability.
The issue around data is universal for AI adoption in all sectors. It is an even more significant issue for telecom network. Unlike self-contained applications of AI, e.g., facial recognition, where the data source is relatively easier to obtain (pictures) and to label (dog or cat), the data from network can take different format, with different time granularity, and from the different domain and locations of the network. Network operational data is difficult to acquire, not only due to privacy concerns and GDPR, but also the significant efforts needed for measurements in a large scale and varied infrastructure. In addition, labelling of network data needs substantial domain knowledge and experience. All these contribute to the lack of relevant and mature data sets for AI in the network at the moment.
Limited network data are available. However, it is usually gathered in a fragmented manner at the moment - for a specific problem and at a specific network domain or location - therefore may not necessarily represent the dynamicity and heterogeneity of the real network. Different data sets may be interpreted differently, resulting in partial, and to some extent, incorrect understanding of the network contexts, when fed to the AI. The industry therefore needs a unified approach to have ‘a common language’, for AI to understand and interpret the diverse data sets from the large scale 5G infrastructure.
Lack of confidence in the reliability of the AI solution is the second key factor that hinders the adoption of AI in operator’s network. There are several reasons behind this.
Firstly, there is in general a lack of benchmark for different AI solutions. Given that the development of AI models are based on exhaustive experiment of data, the fact that the isolated AI solutions are developed using fragmented data sets makes it almost impossible to tell if one AI solution is better than the other, or if it will work at all when the data sets (network contexts) change.
Secondly, unlike self-contained systems, network by nature is linked. The AI solution, developed to optimise the operation of one specific domain (e.g., radio access network) of the network, does not necessarily mean it optimises the entire network end to end (see here and here). Given the scale and interlinked nature of the network, unless such a solution is properly validated and integrated into the network, it is very difficult to predict the overall effect of AI that has brought to the network.
Thirdly, there is this “black-box” nature of AI – because we are not entirely clear what exactly is going on inside the AI modules, we are not 100% sure that AI will do what we want them to do, and this is especially a problem when it is used in critical infrastructures such as 5G. Some technical solutions exist from this perspective, for example, developing AI with traceability and transparency, but at the early stage of research. There are also concerns that with AI, devices and network functions may be given too much autonomy2, and that may bring catastrophic consequences to the network.
The key to the market of using AI for 5G and beyond network operation and management is the deployability of the AI solutions to the network. At the moment, lots of the AI solutions developed for network operation and management are isolated, and not scalable, where specific solutions are designed for specific part of the network and for specific problems/applications. The industry needs to have the strategy in place to allow AI solutions to be developed and scaled using unified data set, and properly validated, integrated, and deployed in the network. When AI in network becomes large scale, we need to consider not only how to use AI to enhance network efficiency but also how to efficiently use AI - the reusability of data and AI modules, and the synergy among them as well as with the network (see here) - all these direct us to develop scalable and deployable AI for 5G networks and beyond. This of course needs the capability of utilizing data uniformly across the network, common tools and platforms for validation and integration, etc., which will eventually bring evolutionary changes to the 5G network operation, as well as to the core research of AI.