The potential of AI and machine learning in network management has been generating pace so the role of CW TEC was to explore automation models for AI powered networks and debate some of the applications for which AI can deliver advanced solutions. Stephen Unger (former CTO at Ofcom and Chair of the CW TEC 2018 Steering Committee, pictured below) introduced the day and you can read his key lessons learnt in a blog here.
One of the most relevant sessions of the day for 5G enthusiasts was chaired by Mary-Ann Claridge of Mandrel Systems. 'Making networks think!' featured well versed speakers including Dean Bubley, Founder of Disruptive Analysis (pictured below with Mary-Ann), Dave Salam, Director Mobile Core and Data Analytics for BT and Tero Rissa, Chief Architect, Machine Learning, ATF, Mobile Networks at Nokia.
Within this track, Futurist Dean Bubley, who describes himself as a 'hype buster', provided a colourful commentary on the broader context of automation in networks and 5G.
Image recognition plays an interesting role; perhaps in maintaining physical infrastructure, drone inspections of cell towers can be important, for example, after a natural disaster like a hurricane in order to understand what the physical state of a network is and the extent of repair required.
Dean went on to say "there are all sorts of input data sources as well. Not just internal data from the telecoms industry, but data coming in from third party sources that might help optimise things in the future. For example, the weather - what happens with propagation when it’s raining?
"There are decisions that can be made based on traffic, vehicle and people flows. We might be able to know that in a particular area or particular time of the day, different applications are used, new videos streamed, social media or enterprise activity performed.
These data sets will be siloed, and fragmented, they will be difficult to get. Perhaps locked, in some cases, into certain vendor equipment and systems. They are spread out across the large organisations with vast departments that don’t always communicate with each other, or share knowledge or data with each other. It’s going to be a big organisational challenge, even before we start to think about mergers, acquisitions and consolidation.
You can view Dean's slides here.
Next up, Dave Salam from BT discussed the importance of real time data and providing quality service to your customers.
When he joined the operational function at BT, one of the first things Dave did early on, was put in place a framework that was going to drive how they would develop the structure to change how they operate the network. He worked to move the organisation from network performance key performance indicators to key quality indicators and from there to key value indicators.
We are thinking much more about what the customers want and turning that into actionable insight... it’s actually very powerful. Getting that in real time is the challenge we have.
He continued, "What became clear is the need to have the right level of network intelligence. Thousands or millions of alarms coming through your network trying to interpret aren’t going to solve anything for you. It became clear a range of data sources is needed to come together, and it’s been said these are often very siloed."
But actually you need some way of aggregating it together, you need an overarching big data solution, even if it draws the data of the back of each of its individual platforms or capabilities, can actually start giving yourself a position where you can start learning and providing the intelligence side that you need.
For example, by comparing predicted coverage with external test data and information on e.g geographical positition, weather, nearby vegetation, over time networks will be able to better anticipate the actual coverage that the user sees. With that level of intelligence, Dave believes that a network operator will be in a position to invest more smartly in capital expenditures, upgrading services where it will deliver the highest return of customer value. BT is currently capturing around 3Bn data points a day to help build a picture of the quality of their network.
You can view Dave's slides here.
Tero Rissa, Chief Architect for Machine Learning at Nokia, concluded the track. His role at the company is to discover areas where ML can be valuably productised.
In setting the context of the day from an artificial intelligence system design perspective, within the deep learning categories that are studied, the most common is supervised machine learning. While at any conference on ML today, the majority of talks are on unsupervised learning, he advised that network operators venturing into the use of artificial intelligence should start with supervised.
It’s hard because you have to have the correct answers to show the machine how to do it. You can’t solve anything that you can’t solve yourself. If you can’t figure out in any amount of time a solution to an answer...don’t expect a machine to know.
Now is finally a good time to exploring the functionality that AI offers networks. The basic algorithms are well-established and understood, and there is a broad array of further techniques for those seeking to venture further. There are frameworks available such as Keras, TensorFlow and PyTorch that offer a huge increase in productivity. Computing capacity has reached a level where hardware with the necessary perforamance is affordable (take this with a pinch of salt).
In fact, of all the tasks required to initiate a machine learning system in a network, the machine learning calculations have been the smallest concern in the projects that Tero has encountered. Because it is now an accepted area of knowledge, technology is widely available on the market for use that requires little additional customisation to make it organisational/task specific. Far more time-consuming is the data engineering, infrastructure installation and non-ML software development.
You can view Tero's slides here.
Ultimately, the track concluded that the areas where AI can offer the most in the 5G era is:
- New service models including fully personalised customer experience informed by context
- Real-time automation for SDN, NFV and Dynamic Network Slicing
- Predictive maintenance
However, this will require the removal of system siloes both within and between organisations in order to enable a comprehensive view of network performance and what affects it.
A memorable panel moment, captured by Mary-Ann Claridge, featured competing AIs vs co-operating AIs. A discussion on where AI processing should take place: on ‘independent’ devices or within a controlling network.
Mary-Ann speculated that AI on a device would be more likely to act ‘selfishly’, each acting to the benefit of its own device, and so on-device AIs would compete, while those in the network would be more likely to work for the ‘greater good’ of the network and so would co-operate. She asked the panel to comment on the balance between competing & co-operating AI.
She continued, "Arthur (Fech.ai) said it all came down to price - that sounded to me like a cop-out, and not really relevant. Ray said standards would ensure they co-operated. Yue initially said that good performance of the network would benefit all, so even ‘selfish’ AIs would end up co-operating, then she backtracked and said that if they didn’t co-operate, regulation would be put in place to ensure good behaviour.
In human behaviour we have the concept of Enlightened Self-Interest - where each person understands that acting to help someone else will ultimately be in their own interest
Images courtesy Charles Sturman (www.sturman.co.uk)
A special thank you to Dr Ian Wassell, Senior Lecturer, University of Cambridge Computer Laboratory for hosting the conference and Fetch.ai for sponsoring.