Sponsored by: Aria Networks
As operators look for a way to address both the pressures and the opportunities they face, they are acknowledging that nothing less than a transformation of their operating models will do.
The result of this transformation will be a business in which demands, though unforeseen, are nonetheless satisfied instantly. In which virtual networks will come and go in minutes, defined entirely in software, according to the needs of new applications or devices, in motion or at rest. And above all, one in which the telecom operator is profitable.
This calls not for a new kind of network so much as a new kind of network operations. One that can dynamically adapt the network to meet demand; to optimize costs or power consumption; anticipate future failures and self-configure to mitigate or avoid service impact.
This is perhaps the most exciting area for the application of AI techniques. But operators should be aware that it requires a different application of AI, beyond that employed in conventional analytics tools.
That is not to say that this goal is a distant vision. The foundational tools and techniques exist today. Moreover, the business realities that telecom operators face – high costs, falling margins and relentless growth in demand – make this not a research project, but an essential near-term goal for today’s operators.
Artificial Intelligence is central to this effort. And the combination of NFV, SDN and AI represents the enabling technology foundation for the future of telecom. Enlightened operators are beginning to recognise the central role that AI has to play: beyond analytics, as an intelligent actor, designing the networks and services that best satisfy the goals of the business – autonomously.
Business problem first…
The terms Artificial Intelligence and Machine Learning are umbrella terms for whole families of techniques and technologies.
So it is vital that anyone exploring AI/ML starts with a clear understanding of the business problem they are trying to solve – and not with the technology. As the saying goes, necessity is the mother of invention. Operators without a driving business need that must be addressed are unlikely to be able to make a business case for using AI. Even those that do have a clear business need must be able to unpack the nature of the problem, and only then start to ask whether and how AI and ML technologies may be a good fit.
To a TMForum readership, this may seem obvious. But it is worth restating, because of the apparent positioning of AI and ML as cure-alls for a wide range of ailments in telecom.
In addition, the availability of generic, open source components (the likes of Google TensorFlow) should only serve to reinforce the point that the value of AI increasingly isn’t in the core software technology, but in the understanding and expertise required to apply it to a particular problem.
Analysis vs action
Getting to the sort of network autonomy described above requires a clear understanding how AI supports two quite distinct problems.
Figure 1 illustrates the point. The left hand side shows two related types of business question. These represent typical use cases for Descriptive and Predictive Analytics respectively. This is the domain of big data sets and tools, pattern recognition, trending and so on. In this domain, operators hope to gain greater insight into past or future events. Hidden within the data are correlations, implicit models or historical patterns that AI techniques can be used to uncover. The outputs from these scenarios are charts, reports and dashboards. Vendors working in this space refer to “actionable intelligence”, which provides a clear contrast to the “intelligent action” which occupies the right hand side of the figure.
Figure 1. Analysis vs Action
For autonomous networking, we need the network to act. More specifically, we need the network to make a decision, and then act. That’s a very different form of problem than looking for patterns in a big data set. But it is better aligned with an overall automation agenda.
As shown at the bottom of Figure 1, the outputs from this application of AI are the designs, configurations or commands required to adapt the network to meet a particular set of constraints.
Determining the optimum design or change can be complex and time-consuming. However, this activity is at the heart of telecom operations, so automating it is key to transformational change.
Networks are expected to become more dynamic. In other words, they will change in more significant ways, and more frequently, than today. That will require faster decision-making around the choice of network resources and paths that will satisfy both a customer requirement and business constraints.
Drilling into the different contexts for automation also helps us distinguish different applications for the use of AI in telecom, at a customer, network operations and business level.
Figure 2 illustrates how, as part of automation goals, outputs from analysis are required as necessary input to automated decision-making.
Any process today that is dependent on human intelligence to execute, represents an obstacle to scalable growth. And perhaps few parts of a telecom operator’s business are more dependent on human expertise than the planning, design, and day-to-day management of the network. However, before they can apply AI to the problem, operators must be able to answer one very simple question….
What Does Good Look Like?
We all know that AI can be used for discovering patterns and correlations. But how do we train an AI engine to design networks? For this, we have to be able to define what is meant by a “good” design. By generating and scoring many different candidate designs, AI can be used to converge on a solution that offers the best fit to the given constraints.
For example, in a pure networking context, there are standard technical measures of the good-ness of a path, such as shortest path, fewest hops, or lowest end-to-end latency.
But in practice, human intelligence is relied on to factor in non-network constraints such as:
- Vendor preference (phasing in Vendor B’s kit vs Vendor A’s kit)
- Geo-political rules (such as avoiding routing traffic through a particular country)
- Planned engineering work
And there can be yet more corporate or regulatory policies that also affect the design, such as:
- Total power consumption (telcos are a major consumer of national grid power)
- Overall cost (especially if third-party services such as leased lines are involved)
- Service profit margin
The complex, multi-constraint optimization problem that this represents is both real, and an ideal application for AI: as a design tool. (Interestingly, there are now a number of comparable examples in automotive engineering for using exactly this form of AI, using it to design a chassis that provides an optimum solution to satisfy many different physical and environmental stresses. Telcos should note that the resulting designs were radically different than a human engineer would develop – but yielded a superior result.)
Virtualization, orchestration and AI
The promise of virtualization is the ability to make rapid, instantaneous change to networks. The freedom to expand required resources seamlessly, without limitation.
From the outset, the industry recognized the need for an orchestrator – a capability that would provide overarching control of virtual resources. That could turn demand into the instructions required to satisfy demand for services.
However, that vision was incomplete.
- The work on defining an industry standard for Orchestration left critical questions out of scope:
- When should I use NFV from vendor A vs NFV from vendor B?
- Which VNFs represent a critical point of service failure?
- When will it be profitable to migrate physical services to virtualized alternatives?
- How will this all work on networks of continental scale?
As a result, while ETSI MANO was a springboard towards a realization of the vision, operators also continued to develop beyond it to meet the end goal of scalable business automation. (AT&T’s ECOMP paper does not include a single reference to AI, though much of its public discussion about its virtualization effort does).
MANO provides a standardised basis on which network change can be effected. But it does not (nor did it seek to) provide a basis for determining which network change should be effected.
The core activity of a telco – service and network design – is not part of any industry-standard reference model for Orchestration. Yet automating it is critical to the success of the drive to virtualize networks. Insightful metrics alone can’t deliver that.
AI: More than Just Big Data
A UK mobile operator was recently fined £1.9 million – not for an actual outage, but for a network design configuration that would have compromised access to the 999 (911) emergency service, thereby breaching a condition of its operating license.
This wasn’t an operational failure. There were no alarms, no screens turned red in the NOC. At no point was life-saving service interrupted. However, it was a design failure – one that cost £1.9 million in fines. Had the critical failure scenario actually occurred, the cost could surely have been far higher. Could AI have been used to avoid this?
For most operators, network planning and design remains a largely manual process, run by smart, experienced engineers using spreadsheets and whiteboards. It’s complicated and difficult work. But with networks undergoing constant change, it’s a risky strategy to rely on human design alone to ensure no such single points of failure arise.
Automation is much on the agenda for telcos. But as this example demonstrates, it must come with the sort of intelligence that prevents hidden vulnerabilities from creeping into the network. That’s an even more complex computational problem – which makes it a great use case for Artificial Intelligence techniques.
The goal of greater automation requires telcos to address the question of how changes to networks are designed, not only how changes are executed. By incorporating AI technology into the design process, operators will be able to remove internal barriers to speed and scale, and enable truly intelligent, autonomous networks.