Artificial Intelligence
Future Telecoms
Why JOINER promises joined-up thinking for future telecoms and AI innovations
Reading time: 5 mins
At Foresight Live experts debated how AI could evolve beyond the network to create real value in telecoms
A commonly held view in telecoms is that there are networks for AI and there is AI for networks. But Dan Warren, director of communications research at Samsung, argues there ought to be a third part to this equation: AI at the application level.
Samsung has been working with the University of Bristol on the REASON project since 2023. The corporation brought a number of pre-existing AI applications for networks to the project and “looked at developing a framework within which AI could exist,” Warren told the audience at the Foresight Live event on the Convergence of Critical Technologies in AI & Telecoms. The question, in essence, was how the network needed to adapt to enable AI to be applied to it, and how having the right levers for automation might allow an AI to directly make a performance or optimisation impact on the network.
Warren worked with fellow panelist Shadi Moazzeni, a lecturer in networks at the University of Bristol, who highlighted one concrete result from REASON: the Multi-Access Technology Real-Time Intelligent Controller, which combines multiple radio access networks (RAN) and LiFi controllers, and adds AI to optimise network resources and performance.
Adding AI isn’t just a question of adding functionality and that always being a positive. Doug Pulley, co-founder and chief technical officer of RANsemi, a stealth start-up working on 5G/6G RAN baseband technologies at the edge, said adding AI to things like base stations required consideration of the power requirements: there is no point in integrating AI if it costs more in electricity bills than the benefit it produces.
For Robert Curran, a consulting analyst with telecoms consultancy Appledore Research, AI represents another major disruption to telecoms following the cloud and the softwarisation of networks. He stressed that AI itself is not necessarily new for telecoms: machine learning has existed in the industry for at least a decade. “The way it’s being applied is starting to become much more novel,” he added. “What if I can interact with the network in natural language to say: ‘you configure the network to do something else, do a different job’ – and for that not to involve an engineer and a set of screwdrivers and planners? That’s a very powerful idea.”
The telecoms sector is trying to adapt at a pace that it is not necessarily used to, Curran explained, and the supply chains aren’t quite there yet. “We are seeing, I think, a willingness to entertain some of the newer ideas, even some of the newer companies, but it will take time. I think in the RAN space, something like 95% of the spend is with five companies. That’s 5% for everybody else in RAN. So that’s going to be a tough one.”
A key opportunity for AI and telecoms is the abundance of data that networks produce. “Everything is generating metrics,” said Warren, “which you require to be able to manage the network.”
Warren gave a real-world example of using AI that didn’t initially pan out as expected. In Rochester, New York, an AI was used to analyse load and turn off parts of the network that weren’t needed in order to conserve energy – a total of 25% of electricity was eventually saved. The AI also detects if load is going to increase quickly and brings more capacity back online.
What nobody had foreseen, however, was that every time users were moved onto a different radio band by the AI so that part of the network could be turned off, the users immediately came back. “Unbeknownst to us, there was another AI which had been developed by our sister research centre in Toronto, which was trying to do load balancing,” said Warren. “Every time we pushed devices off of a radio band so that we could turn the radio band off, their algorithm looked at it and went, ‘Why are they all there? We should be load balancing across the three radio bands,’ and push them all back again.’”
It shaped that aforementioned argument that networks need a layer “where all of the intents of the AIs are understood so you can prioritise,” added Warren.
Moazzeni agreed that conflicting AIs represent a challenge that’s also receiving attention from academia. Researchers are also focusing on the data side, she added. It’s a question of both building datasets that are good enough to train AI on and safeguarding user privacy.
Curran said the telecoms community is “becoming much more alert to and aware of the dependency of AI on data”. He added that with regards to privacy, it’s important to consider how the data is managed, governed, who’s allowed to access it, and how long it will be stored. “There’s a temptation to keep everything,” he noted. “From an operational point of view, that’s a terrible approach because you almost never need everything.”
There is an opportunity around this data management, he argues, even before it gets to AI. Curran posited that pulling data into one set – particularly in a telecoms setting, which is a multi-vendor environment – will be useful “even before you do something clever with it”.
Some telecoms networks face a self-inflicted problem: over the years, they have relied on giving a big equipment vendor a large contract to solve challenges and now find that they “don’t understand how the network is engineered”, added Curran, because the contractors didn’t always build the network in an optimal way or bring the expertise back into the telecoms operator. “There’s a large executive which are trying to re-skill and skill from the inside out where they can,” Curran observed.
The concern, in telecoms companies, of adding more functionality is also around assuring current services. Curran said this is critical things like the emergency services: “Our networks are regulated to support 999. I may only need it one call in 10,000. It doesn’t matter – don’t do anything that affects that.”
Pulley noted that the question of safety also points to the question of the business model. “One of the big opportunities today is the connection of multiple technologies at the same time and specifically around safety, you can definitely make a value proposition on building sites, on driving,” he said. “The autonomous vehicles thing is a red herring,” Pulley also cautioned. “But are there things that you, as a telco, can do with other software requirements, with other video analysis companies that will make it safer to be on the roads? Absolutely. But you’ve got to piece those pieces together. Making it a safety play freaks telcos out as a concept, but there’s obviously opportunities there.”
Pulley added: “Every time you’ve got heavy machines for people, or crowds, or remote locations, or dangerous locations, there is an opportunity. There’s risk; there’s an opportunity.”
“The entire network is a compromise. You’re permanently in the decision about compromising coverage versus capacity,” Warren stated. Telecoms companies can apply AI to optimise but a challenge is that, at the management layer, “standards are flaky at best”.
“They have been the source of considerable contention because they result in vendor lock-in,” Warren said. “But if you don’t have interoperability on the management layer, you end up in a situation where one management layer controls everything and it’s impossible to churn anything out – and then you’re stuck.”
One of AI’s early promises, Warren continued, was anomaly detection: a blip in the network that’s indicative of something much bigger going wrong somewhere else later. “If you get a very limited dataset from an operator to use, to do research against, one of two things happens: there are either no anomalies in it, or there is one anomaly in it. But the dataset is so small, the anomaly is repeated because you have to keep training against the same dataset over and over again – and it’s not recognised as an anomaly; it becomes a pattern.”
The only way to solve this is through access to real-life operator network data to research against. Warren noted that there is one big dataset released by Telecom Italia that is used by everyone: “You could open academic papers, read the first paragraph and go, that’s based on the Telecom Italia dataset.”
Moazzeni added that it’s tricky to figure out how to educate the next generation of telecoms engineers. Industry requirements in a decade are “very hard to forecast from now because it’s changing in minutes”, she said. This is why it’s crucial for a university like Bristol to actively engage in big research projects with industry, like REASON, because they help everyone understand where the technology is headed, she stresses.
Ultimately, all is still to play for in this space. “The company which does well out of this is the company that gets to the point where they can de-risk this,” said Warren, adding that no such company exists yet – with the exception, perhaps, of Nvidia. “It’s all risk because it’s all future and the de-risking part is the bit which presents the opportunity.”
Thierry is a freelance journalist specialising in university research commercialisation. He has over a decade experience covering spinouts and university venture funds globally, with his research cited in publications including the UK government's Spinout Review, the Financial Times, and The Wall Street Journal.
Quantum
Reading time: 10 mins
Quantum
Reading time: 10 mins
Quantum
Reading time: 11 mins
Robotics
Reading time: 1 mins
Quantum
Reading time: 3 mins