Anyone in RAN engineering will tell you that it has become that much harder to master all different radio technologies. There's just too much to learn and not enough time. Never mind that each vendor does things slightly differently.
The unspoken expectation? Do new things with less training and less resources. And don't complain if you can't get access to the best of vendor documentation.
RAN technical complexity though is what appears to fuel the view that AI/ML needs to play a bigger role ... to eventually master the network and all of its data, and provide even greater savings.
Marketing is everywhere
In Radio Networks, name-dropping AI/ML seems to happen all the time. If you don't talk it with some level of know-how, then you feel inadequate.
Vendors tell you that their new RAN software products come AI-powered with something really special inside. They don't ever tell you how it works though. Operators tell you that they're digitally transforming, their workforce newly enabled with the tools and skills to use AI/ML.
And marketing is everywhere, proclaiming how AI/ML will solve many of RAN's complexity problems, especially in the new 5G world.
But AI/ML is vast
If you take a look at this summary from Stefan Kojouharov, it becomes apparent that AI/ML is a vast complex field.
There are tons of techniques and algorithms, concepts like supervised and unsupervised learning and simply a mountain of information that we straightaway know might be months or even years of work to understand deeply. And this before we try to understand how it might be used in RAN.
With so much information, the easy option for a telco might be to hire data scientists. Come in, bring your toolbox, work with our team and do something. After all, it might be less risky and cheaper than buying an AI product from a vendor that has a brilliant-sounding promise but might not deliver ... while leaving you on the hook politically.
Scoping out AI/ML
A recent Gartner study said that up to 85% of AI/ML projects fail. It's tempting to dismiss these sort of studies as "it doesn't apply to us". But are the typical criteria for such projects? An interesting perspective comes from Ben Evans, an ex-telecom analyst now in the VC world.
What areas are narrow enough that we can tell AI/ML the rules but deep enough that looking at ALL the data in a way a human can't might bring about new insights?
What he appears to hint at is that most successful real-world use cases have certain traits.
- Narrow well-defined scope and rules
- High volume of data and lower data "width" (number of variables)
- Fairly static environment
Given RAN's dynamic fast-changing environment, that straightaway appears challenging.
The story on the ground
For the engineers on the ground, the stories are usually involve lots of struggles. Most self-train on Udemy or with Andrew Ng on Coursera. And then have to figure out how to make what they've learnt work for RAN.
One senior RAN engineer with a Tier 1 operator was frank about how, in a wave of cost-cutting, the RAN engineering analytics function was centralised under finance. "And they even took the department's 128-core server" he'd said before being explicitly told not to work on any more machine learning projects.
Another spoke about the struggle to build ML-based traffic prediction tools off his own back, because of lack of budget for training or resources, and then fighting to get servers from IT.
And then, there is the frustration of being unable to access data stored in 3rd party OSS/BSS tools - "The vendor API isn't open to us to retrieve all the RAN data that we need".
One conversation in Australia seemed to reflect the business mindset around AI/ML. "When we started exploring machine learning, our bosses hired a data scientist into the team and then wanted to see results within 6 months. But who has time to train the data scientist on radio?".
Lack of business value
And so over 3 years, we saw many AI/ML projects struggle to deliver to business expectations. Those that did ended up being custom solutions, crafted for a particular telco with all its peculiarities. Without exception, these projects were extremely time consuming.
What bothered us the most was that many of these problems did not even need machine learning ... but ... there was real pressure to use AI/ML because everyone else was using it.
And the hardest customer concerns to answer?
- What would happen to the network if we implement AI/ML recommendations?
- Can you guarantee the prediction accuracy?
- Will it affect any other features in the network?
So we asked ourselves - given the unique effort required for each problem and no guarantee of outcome, how could it possibly scale? Across different problems? Across different telcos?
Will AI/ML end up being used on a case-by-case basis only?
In order to answer that, we must first take a step into the world of RAN configuration.
The increasing complexity in RAN configuration that's driving AI/ML