For some time, artificial intelligence, with Machine Learning techniques, has become a useful tool in many application domains, from image recognition to medical diagnostics. For some time, artificial intelligence ( AI ), with Machine Learning techniques (ML, machine learning), has become a useful tool in many application domains, from image recognition to medical diagnostics. However, AI experts often ignore how many and what applications are foreseen for this technology in developing future 6G networks. Instead, new generations of telecom engineers need a learning foundation, and AI experts need to turn to the telecom industry. Where will AI apply specifically in 6G? What problems will it have to offer a solution to?
The 6G, like previous generations, will consist of mobile devices and objects ( things ), base radio stations and a network infrastructure that will connect the devices to the cloud and the Internet. The AI will play a fundamental role in each of these segments, operating on the devices’ data and the signals exchanged between the network nodes. Let’s start from the application side and then from the data produced by the devices. The use of AI for their analysis when the data are in huge quantities ( big data analytics ) or you want to obtain information that cannot be extracted with heuristic techniques is not a novelty introduced with 6G. On this front, industrial automation, for example, will require maintenance prediction applications that will extensively use ML techniques. Connected cars, with dozens of multimedia sensors producing huge amounts of data, will rely on ML algorithms for their analysis.
And The Network?
As anticipated in the first issue of this column, AI will be at the heart of the 6G network, in the sense that AI-based algorithms will make all decisions on how to use computing and communication resources. In particular, the SON (Self Organizing Network) paradigm, known for some time in the scientific field but never applied in-depth, requires the network to optimize its configuration, acting on thousands of different parameters in real-time and autonomously. AI, particularly Multi-Agent Reinforcement Learning (MARL), plays a very important role here. The writer believes that the enormous amount of configurable parameters in 6G networks can only be addressed with techniques such as MARL, which learns from the environment (environment) and aims to optimize a reward, intervening with suitably designed agents (neural networks).
The Base Radio Stations
We come to the base radio stations, which exchange signals with mobile devices. The choice of the transmission signal format and the extraction of information from the received one will be able to use AI to produce decisions supported by the experience acquired in the learning phases. In this case, it may be useful for nearby base stations to exchange information on the environment they explore so that each can optimize their choices based on the experiences of the others. A technique extensively studied in recent years is Transfer Learning; in this case, a model trained on one neural network can be transferred to another to start executing a different task.
Finally, mobile devices and things. Think, for example, of objects that have to carry out a single task in a shared and collaborative way; it is useful that the neural networks resident on them periodically share the status of their parameters with a centralized body (which could be the radio base station) which processes the received models, integrates them and re-sends new configurations to the objects ( Federated Learning ). Depending on the complexity of the neural networks involved, these exchanges can impose the sending of tens of MBytes with very high frequency (e.g., every 10 ms). Minimizing the amount of information necessary to describe the state of a neural network is a very important goal requiring specific skills.