AI in telecom: Past, present and future

How the AI journey started

The artificial intelligence (AI) journey here at Ericsson began over 15 years ago when a group of enthusiasts at Ericsson Research started exploring big data analytics methods. Their goal was to simplify the lives of their colleagues working in operations, who had to manage countless trouble tickets and alarm data from all over the world. At that time, our first- and second-line support teams were comprised of thousands of people, as our Managed Services, which still serve over one billion subscribers today, used to be a highly labor-intensive business.

Most of the knowledge and information about the telecom domain existed in manually readable text, such as product documentation. Over time, this text has been transformed into more structured, machine-readable formats, facilitating easier processing and analysis. So, at Ericsson, we leveraged this information-rich documentation to create the first version of our telecom knowledge base, which was successfully applied to resolving trouble tickets. Since then, significant changes have occurred. Automation has eliminated numerous manual tasks, and privacy concerns have been addressed by reducing manual operations for different customers.

Digital users of internet. Picture source: DALL·E 3

After addressing the most urgent areas in need of automation, which included tasks like ticket resolution and using chatbots instead of people to provide initial support to engineers, field operators, and customers, we turned our attention to optimizations that were more closely related to the heart of the telecom system. This included the radio access network, which features cell towers, base stations, and antennas, as well as the core, responsible for routing data and managing connections. The specifics of AI applications in these parts of the telecom system involve real-time analyses and actions, often occurring within milliseconds.

To proactively address any undesirable situations in the network, such as service degradation, which requires early attention when relied upon by safety-sensitive or safety-critical enterprises, it's crucial to receive input data promptly. This necessitates high-frequency data sampling and processing of input data in real time. In this context, when we say "real time," we mean "as quickly as possible," which differs from the formal definition used in real-time systems research, where the result is expected to be delivered "just in time."

Knowledge sharing - Stockholm Library. Picture source: Paul Petersson Fersman

Among the AI applications that necessitate real-time data acquisition for predictive and optimization tasks in mobile networks, we find:

• Anomaly detection: AI algorithms identify patterns that may indicate security threats or network issues.
• Dynamic resource allocation and quality-of-service optimization: AI-driven systems allocate resources, such as bandwidth and computing power, based on needs and priorities.
• Network traffic predictions.

AI systems responsible for these functionalities are either already operational within Ericsson or currently in development.

Paradigm shifts

The world of telecommunications has undergone significant transformation due to the recent adoption of AI algorithms and AI-driven systems. Beyond the immediately noticeable benefits of AI adoption in telecommunications, such as the simultaneous reduction in energy consumption and the enhancement of network performance and reliability, the most critical paradigm shifts associated with the integration of AI include the following aspects:

1. Firstly, thanks to knowledge-sharing techniques, we can observe increased reuse within the telecom industry and across adjacent sectors. This change has become possible through global knowledge-sharing techniques. These techniques encompass: ontologies, which are formal representations of knowledge describing concepts, entities, and the relationships between them (think network's physical and logical topology); semantic interoperability, which is the ability of different telecommunications systems and devices to exchange and interpret data, facilitated by well-defined ontologies; and large language models, which are advanced software systems capable of understanding and generating text based on patterns and knowledge acquired from vast amounts of text data.
2. Secondly, AI is replacing dangerous tasks that our field support operators used to do, and in some cases still do, such as tower climbing to perform an inspection. AI systems deployed at telecom sites are nowadays capable of preventively notifying engineers about any failures or predicted service degradations. This can be done by analyzing the data from base stations or external evaluations using drones in combination with computer vision models. In many cases, this predicted service degradation can even be resolved remotely, without exposing humans to potentially dangerous tower climbing.
3. Thirdly, AI has proven to be successful in control loops, including the fastest, close-to-real-time control loops. These control loops assume the execution of an AI algorithm on dedicated hardware as close as possible to where the decision is taking place. Fast control loops—one example of which are AI algorithms that can dynamically adjust bandwidth allocation in real time based on network traffic patterns, thus ensuring efficient data transmission while minimizing congestion and latency—are found at the heart of radio access networks.
4. Lastly, the latest advancements in AI for telecommunications encompass various applications of large language models. While the advantages of using these models for customer support are evident, their application is also anticipated to significantly enhance the quality and efficiency of research and development. This includes tasks such as idea generation, concept validation, knowledge extraction from primary research and product documentation, and predictive modeling.

Mothers of mobile internet. Picture source: Photo by Mudit Agarwal on Unsplash

What’s missing?

Let’s discuss what’s needed to fully harness the potential of AI at large.

1. Firstly, the AI market is fragmented. Even within one company (and even if we are talking about a rather large company like ours), once a new technology is identified and targeted, one can find dozens of groups, strategic initiatives, and systems where the new technology has been developed and applied. This is simply human nature, and the reasons behind the proliferation of developments and applications of new technologies within a company are more likely to be found in the behavioral sciences than in computer science. Although this proliferation of initiatives may reduce the efficiency of the adoption of AI solutions, all these different investigations, developments, and deployments of AI will still facilitate the overall progress of the technology, and eventually, we will converge on one - or a couple of - winning solutions, as is the case for the hyperscalers of today’s world.
2. Secondly, this technology is not being developed according to standards. Since in telecommunications, standards mean both scale and interoperability, standards are highly appreciated by telcos. In AI, however, the pace of development is much faster than what is typically seen in the telecom industry (for example, ten years are expected between two generations of telecommunication standards, such as from 4G to 5G, while it took less than a year for OpenAI to move from the large language GPT-1 model to GPT-2), and AI developers should consequently have more flexibility to drive innovation, especially in the early stages of development.
3. Another gap — and after this, you’re free to call me Captain Obvious—is observability, which is the capability of monitoring, tracking, and understanding what is happening within a system. We have lived with this gap way too long and, to be honest, it has caused me some gray hair.
   Here, we’re facing a digital divide, but it’s not about consumers, but rather industries. You see, a “youngster company”, a new entrant on the market, would be a digital native, with the opportunity to design its processes and products entirely from the ground up. This means that the company’s data collection and development platforms will likely be able to read any data from the products and services deployed in the field and use that data to improve its offerings.
   If, on the other hand, we look at a 100-plus-year-old company like ours, then processes, mindset, and culture will need to be gradually redesigned according to digital principles, putting AI at the center and not "on top" of solutions. 5G partially featured AI, while 6G is being born as an AI-native system.
4. Furthermore, AI trustworthiness needs to be taken seriously, and algorithms ensuring adherence to non-bias, non-discrimination, and explainability principles need to be in place. 
5. Lastly, let’s not shy away from the fact that AI is software, and without proper implementation its outputs are useless. Well-executed implementation is as crucial for the entire system as thorough observability.

Takeaways

1. AI isn't just the icing on the cake; it needs to be integrated into a product or service to fully use its capabilities, whether that means taking proactive measures or reacting to and compensating for network anomalies and subpar performance almost instantly.
2. In telecommunications, AI must focus on what truly matters. We have an abundance of data, much of which may not be interesting or actionable. However, when something critical happening in the system is on the horizon, we must be able to monitor it frequently, often in real time.
3. We must give AI time to prove its worth. In industries with over a century of history, like telecommunications, there are finely tuned algorithms that are hard to challenge, even though questioning them is necessary for embracing new and superior technologies that promise enhanced network performance. Transitioning from a conventional rule-based system to AI-based algorithms may initially lead to performance degradation, similar to introducing a talented trainee to a workplace filled with experienced professionals. Giving AI the opportunity to prove itself is essential, and, like new trainees, AI algorithms are inherently adaptive.
4. Facilitating AI-to-AI communication is crucial. Each industrial domain possesses its unique enterprise data, soon to be analyzed by its AI systems. Often, data and outcomes from different domains need to be connected or combined. For instance, in telecommunications, intelligent transport systems rely on connectivity, which, in turn, relies on computing infrastructure. Effective communication and collaboration between different domains, each with unique strategies and objectives, is essential.
5. Ensuring the trustworthiness of AI is paramount. From the outset, it's vital to guarantee that software and algorithms, even when well-intentioned, won't inadvertently disrupt telecommunications networks. In telco network management, trustworthiness guarantees that AI-driven optimization processes consistently improve network performance while preventing disruptions or service degradation caused by software and algorithmic errors.

Despite the exciting advancements in AI for telecommunications, we are still at the beginning of the AI journey in this field. Our company, with its talented AI scientists and practitioners, is leading the way in innovating AI-driven telecommunications of today and tomorrow.