Four ways generative AI is set to transform the telecom industry

Artificial intelligence (AI) has become arguably the world’s hottest tech topic over the past year. Though AI-powered algorithms using data in prediction and to make more informed decisions is nothing new, what has really disrupted industries around the globe are critical technologies such as generative AI. This impact has been most notably demonstrated through the explosive success of OpenAI’s text-based AI chat tool, ChatGPT.

So, what is generative AI? What makes it different from the AI we’ve been using in the past? And how can generative AI transform the deployment, management, operation and improvement of our telecom networks – and businesses?

Generative AI is a type of AI that creates entirely new content, based on the characteristics or patterns of the data it was trained from. In the past, we’ve seen a lot of predictive AI use – that is, AI algorithms that, when trained on a series of previous data, will tell you what they expect to come next – a forecast or prediction of the likely outcomes based on a set of values. Generative AI tends to be more innovative, having the capability to create unique new content seemingly from scratch.

Imagine a famous author who only ever wrote short children's stories throughout their career. In the past, AI could only analyze the content and make predictions on characteristics, for example, the author, the topic or the sentiment of the text. With generative AI, however, characteristics of that writer's style can be learned and captured – their use of language, their grammar use or the way they structure a story, for example. These can then be used to create whole new children’s stories in their style or even applied in a whole new context never seen before, such as an adult fantasy novel, a poem, a script for an animated film or even for product advertising.

How will generative AI help transform telecom?

It's easy to see why this sort of technology is transforming communication and content generation – but what role does generative AI have to play when it comes to the telecom industry?

There are four main avenues through which generative AI can (and likely will) deliver significant value for communication service providers (CSPs), end users and other players in the telecommunications sector.

Generative AI: delivering value in telecom - The four main avenues through which generative AI can deliver value in telecom.

1. Human-readable content

The first type of application – and the one you are most likely familiar with – relates to the generation, summarization, presentation or translation of text, images, audio and video content. These types of applications can be implemented across almost all aspects of a business including marketing and sales, customer service, operations, legal, reporting and analytics, career development, and even in the software development lifecycle, for example through code generation or completion.

Often these applications utilize the ability to generate human-readable content such as texts, software code, images, music or videos from a set input of characteristics, or produce text descriptions from a non-text input (such as an image or other data characteristics). In telecom, this can range from the creation of service level agreements (SLAs), product documentation for troubleshooting, upgrades or installation, or drafting to 3GPP standards. The reverse can also apply – for example, the generation of network characteristics from text-based documentation, such as customer SLAs. Intuitive dialogue-based interfaces like that of ChatGPT can also be introduced for expert systems – making user interactions and access to relevant information easier and more efficient.

2. Machine-readable content

These types of applications are those where sources of data such as mobile network data, raw-format logs or network configuration parameters and structures (for either virtual or physical network elements) are used to generate content such as coverage maps, incident identification or detection, search optimization, recommended configurations or even resource allocation. It’s also worth noting that these outputs can usually also then be translated into human-readable content, if necessary. We will go into some of these in further detail when we outline more specific examples of applications in mobile networks.

Generative AI can also be used to synthesize additional data to augment an existing data set for training or other purposes – particularly when data is scarce or expensive to collect. In networks, this might be because existing data is scarce due to a low volume of connected devices at a particular time, a technical malfunction, or the network being overloaded. You can imagine this along the lines of how video-conferencing or virtual reality (VR) software will smooth over the visuals when a connection is lost or poor, displaying artificially generated frames to fill in the gaps and ensure you don’t notice the disruption – or get motion sickness.

3. Semantic communication

Another form of non-human-readable content, semantic communication refers to the process of encoding information in a more compact, compressed format that represents or describes the raw information sufficiently so that it can be decoded or synthesized again at the receiver end for users to understand. By avoiding the necessity of transmitting the raw data in full, this type of communication can increase transmission efficiency and save a significant amount of bandwidth. Several areas in mobile networks have already been identified that can benefit from compressed content in this way.

Generative AI models could be used at both ends of this process – both in the generation of these multi-dimensional symbols prior to transmission, and for synthesizing the transmitted content again at the receiver end. However, these processes are demanding in terms of computational cost, as well as storage cost to keep the encoder or decoder in memory. This can particularly impact Radio Access Network (RAN) applications, where resources may be constrained, so ways to compress the generative AI algorithms or distribute the computation would be vital for large-scale adoption – especially at the network edge.

Semantic communication: encoding information into a compact, non-human-readable symbolic (but still sufficiently descriptive) format that can then be decoded for use after transmission.

4. Digital twins made simple

Generative AI could also play a vital role in creating, or assisting in the creation of, digital twins. Digital twins are virtual representations of a physical object, process, or system, created to simulate and model the behavior, performance and characteristics of the object or system. It can represent any scale or complexity, from an entire mobile network to a single network function or wireless protocol, and is used to test, analyze, optimize, monitor or validate with lower (or no) risk to the live network. While the advantages are significant, creating digital twins has traditionally been costly – both in time, programming resources and data collection for the creation, and computationally for their operation and maintenance.

With generative AI, you don't have to write the code for the behavior of the digital twin, you can train it on the behavior of its physical counterpart – not only saving time, but generating outputs that are more consistent with reality. Generative models can also be used to create more simplified digital twins, while still accurately representing the relevant functions and behavior, not only saving resources but also delivering responses much faster than a more complex digital twin. Essentially, this could make digital twins more accessible and affordable than ever before.

Finally, generative AI opens up possibilities for the creation of interactive virtual environments where control techniques like reinforcement learning (RL) algorithms could be trained and tested, alleviating the risk of unsafe explorations in the real network. Such an RL algorithm, once trained in the digital twin, can be migrated to the real network and deliver optimal performance with the help of domain adaptation and/or Sim2Real techniques (which are discussed in more detail in this blog on RL solutions). Generative AI can even design reward functions for RL, as demonstrated in recent research work from NVIDIA.

Applying generative AI in mobile networks

Now that we’ve discussed the different formats in which generative AI can be used in the telecom industry, let’s take a closer look at some more specific examples of how and where this technology can be applied to improve various functionalities and parts of the mobile network – in both RAN and network management.

Harnessing generative AI for improved RAN

As mentioned earlier, many of the applications in this space revolve around using generative AI for semantic communication, or to learn complex behaviors, characteristics and functions and use the outputs to create more accurate models or to generate datasets to allow more efficient training and operations. Some key uses include the below:

Wireless channel modeling: Learning implicit probability distributions of multiple-input and multiple-output (MIMO) channels, resulting in more accurate channel models, widely used for benchmarking and creation of simulation scenarios.

Spectrum sensing: Generating synthetic data to create classifiers that predict available spectrum, or to estimate the wireless channel occupancy, enabling reallocation of resources and more affordable training of predictive models.

Channel quality estimation: Estimating or efficiently compressing complex and high-dimensional Channel State Information (CSI), which can then be reconstructed at the destination.

Hybrid beamforming (HBF): Generating a low-dimensional representation of the high-dimensional search space, allowing optimum precoders to be searched for and identified, simplifying HBF optimization, decreasing CSI feedback overhead and increasing spectral efficiency.

Network traffic generation: Reducing the bandwidth and computational requirements for training of machine learning (ML) models in dynamic scenarios, such as communication for connected vehicles or unmanned aerial vehicles (UAVs, or drones), or generating more granular data while maintaining high accuracy.

Network traffic analysis and anomaly detection: Together with unsupervised deep learning, generative AI can be used for anomaly detection and does not require large amounts of labeled data for training, making it likely to detect “zero-day” attacks, while allowing scalability both in terms of data and features that can be considered.

Network selection: A potential future application area for user equipment with multi-connectivity capabilities, where a load estimation algorithm can, in turn, provide input to a network selection algorithm that is used to select networks with lighter loads.

Transforming network management

The applications we’re seeing in network management are much more diverse, and include opportunities for improving network planning, deployment and operation. Prime examples of these include:

Customer incident management: Generative AI can support this process end-to-end, enabling greater automation – from using raw log data such as alarms or program traces to help detect incidents, to drafting texts for customer support requests or trouble reports (and converting them where necessary), and even generating labeled clusters of categorized trouble reports for easy searching when similar tickets are raised, enabling faster and easier resolution.

Network planning and deployment: Here generative AI can facilitate not only network planning but also deployment and configuration of network nodes once planning is complete. Some examples include the accurate generation of estimated radio maps (even with sparse data), estimating cell load and traffic routing for various scenarios, resulting in improvements to network coverage and spectrum utilization, and the generation of network element configurations, even with parameters that are more complex to tune, by learning and utilizing structures of existing good configurations.

Network operations: Generative AI could have a significant impact in operations, firstly in fault diagnosis. Model Drive Test (MDT) coverage maps can be generated, even from sparsely available data, which can in turn be used as a valuable input for potential fault diagnosis. Resource allocation for network slicing is another strong example, with a hybrid RL and generative AI solution able to generate allocations of certain bandwidth to each network slice based on the data packets arriving. Forecasting resource utilization could also be useful in this area.

Finally, network security can be taken to the next level by leveraging generative AI to improve malware and rogue device detection. Generated new potential malware threats (optimized to avoid detection) can be used to train malware detectors, without the risk of real exposure to threat actors, or to detect rogue radio frequency (RF) transmitters.

Creating a more intelligent, connected future

The potential value generative AI can deliver for telecom networks and businesses is clear – and it’s not just the use cases already emerging we have to be excited about. While there are still challenges to be overcome, this technology and its innovative ability to learn and generate new data and content is set to transform how we deliver the services of tomorrow – from optimizing XR network datastreams and realizing the safe remote operation of autonomous vehicles and UAVs, to enabling life-saving robotic surgeries and remote diagnosis.

The only question left to ask is – how will you take advantage of the opportunities being generated?

Read more

Find out more about our ‘Benefits of Telecom AI ’ blog series, or sign up here to be notified whenever a new post is released.

Explore the opportunities of applied generative AI for enterprise in this recent blog post, and learn how this technology is being harnessed within Ericsson’s internal operations.

Learn more about Telecom AI, and how it is enabling more intelligent networks.

If you have access through a subscription or academic institution, more information on this topic is available in our full research article Generative AI in mobile networks: A survey via Springer.

Explore AI in networks