On the Hype of AI: Let’s Be Real!

We think so often about how Artificial Intelligence (AI) is going to change the world. For example, Sundar Pichai, Google CEO, and Kai-Fu Lee, founder of Microsoft Research China, both think that AI is even more disruptive than electricity. 

AI can find its place in many different areas, creating value across a variety of use cases and application scenarios. This article focuses on adopting AI in telecom, and how new technology can help achieve autonomous networks.

Is AI intelligent?

The term itself suggests that AI is intelligent. But what is intelligence? As Tim Klapdor argues, intelligence requires understanding and meaning. We could then say that the “intelligent machine” needs to understand the meaning of the “system” where it operates.

AI is a technology that depends on humans to provide the machine with knowledge for it to become “intelligent”. The larger and more complex the “system” is, the more knowledge AI needs to be reasonably intelligent. Just like small kids, that while smart, have a limited understanding of the world and are not ready to compete for a Nobel prize yet.

We could talk about the level of intelligence as a function of the amount of knowledge obtained and the ability to draw correct conclusions from observing the world (the system) in all needed situations that may occur. Intelligence is not black or white!

AI systems can also obtain new knowledge by reinforcement learning, which we could consider a proof-point of intelligence. But reinforcement learning requires a clear reward function, that in many situations becomes incredibly complex or not even possible as we would need to capture the full state space of the system. Such reward function is intelligence “on the next level”.

Applying AI to autonomous networks

There are many areas in a telecom system which could benefit from AI, enhancing any decision made within the system for the best possible outcome. Applying AI is relevant from the lowest layers of a telecom system, like assigning resources based on a subscribers’ demans to communicate, all the way up to a communication service provider’s need for intelligent decisions to achieve autonomous networks.

The industry has coined the term autonomous network to capture the concept of a self-driving, intelligent network and maturity levels, as defined by TM Forum. A fully autonomous network is the holy grail for communication service providers, which today remains a highly complex mission. The number of all possible scenarios implies an immense amount of knowledge that AI needs to be trained at to understand the system and make sensible decisions for each scenario.

Using AI to achieve autonomous networks requires rigorous training on specific data to get AI sufficiently intelligent. Note that sufficiently here translates into the acceptable rate of wrong or improper conclusions drawn by AI. For those errors, we would have to accept not only that AI misses out to react properly to the situation at hand, but it might actually worsen it.

How does AI fit into a control-loop?

Autonomous networks can be considered a control-loop – a control and feedback model that includes sensing, analysis, decision, and action, operates in a specific network, and is defined by a set point, the desired state of the system. Let’s take a closer look at the various parts of a control-loop and how AI can be utilized at every stage.

The loop begins with sensing, which is very much about gathering data and making sure it is properly cleaned, and semantics are well defined. The problem of obtaining a clear and transparent view of our networks is an issue far from new, which impacts a lot of non-AI functionality too. Here, AI proves barely useful and can’t offer much for data gathering and cleansing.

The Analysing stage is the best suited for AI as it can deal with large amounts of sensor data subject to complex relations. Applying AI helps better understand the state of the network by considering relevant context. On many occasions the functionality possible will not close the loop. For example, anomaly detection solutions can identify an anomaly, but fail to provide any clarification on the reason why the issue surfaced. Such conclusion cannot be fed into an autonomous control loop without complimentary information.

The Deciding part is crucial to make the autonomous network a reality, but unfortunately it would generally be difficult to apply AI at this stage. The reason lies in

a) the lack of completeness of the sensing and analysing, and
b) the huge state space that the network can be in, i.e. understanding the meaning of that state-space.

The implication here is that to train a system to act appropriately in every possible situation, we must first identify all those situations. For a comprehensive telecom network, we might not even be able to define all the situations, while some of them would have never occurred, complicating the AI model training. It is possible to define a deciding function for limited parts of the network, but the state space grows very quickly when the control scope is extended.

The Acting stage is not very suitable for AI as it doesn’t contain any element of intelligence. It is simply a transformation functionality which maps the decisions to the exact information models exposed by the underlying network elements. We do not want to introduce any non-exact properties, such as those derived from AI in the final stage of the control loop.

Instead of aiming for the complete self-driving vision of autonomous networks, more realistic approaches are those where AI adds decision support. The final decision-making remains then on people to weigh in on the network situation at hand, as well as factors that couldn’t be foreseen when training the AI-based decision system. As technology evolves, the human role will shift from making a final decision to supervising the self-driving process. These initial approaches correspond to lower maturity levels of autonomous networks according to the classification from TM Forum.

It is important to think about how the set point is expressed, how we define the intended state of the network. Typical operational considerations include making trade-offs between quality and cost, for example, so it is imperative for a communications service provider to control such complex settings without exposing unnecessary complexity. Finding the correct level of abstraction in that control interface is crucial as this is an information model that is exposed to the operator. If aiming for a closed loop approach this information model needs to be matched (on the same level of abstraction) with the output from the analyzing part of the control loop.

As there is such strong hype around AI at the moment, the technology is proposed to be used in situations more suitable for a classic algorithm approach. If we can define a mathematically optimal expression that would yield a better, faster, and more resource-efficient solution, the control loop should be implemented as a mathematical algorithm. Examples here are some of the mobile network layer 1 algorithms, like filtering, interpolation, or transformations.

Multiple control loops and the need for tailored architecture

A telecom network is an extremely complex system that entails various technologies, reaches across numerous administrative domains (regions, countries), all while providers push it forward for a reasonable level of autonomy. To make it work with a sufficient degree of accuracy, multiple control loops need to be applied.

There are two patterns that are of interest when thinking about the architecture of multiple control loops, irrespective of whether they are based on AI or not.

The first is what we refer to as horizontally distributed control loops – they control different technologies, or different aspects within a technology. Ideally, these control loops are independent from each other, but in practice that is rarely the case, creating a situation where a few control loops may conflict or contradict one another. For example, one says “reduce power to save energy”, while another command to “increase power to improve quality”.

To avoid this situation, we can apply the second pattern – hierarchical control loops, where an additional control loop deals with conflicting situations to achieve a more holistic control. Such a setup requires control loops to interact vertically, demanding proper abstraction and information model definition, similar to the set point discussion in a previous section of this article.

The issue of multiple control loops calls for a proper architecture for the final solution to operate stably and deliver the intended business value. The more autonomous the network is, the more complex an architecture will be. The architecture will be specific to each communication service provider to reflect its selection of domains, organization structure, and way of working.

Do autonomous networks pay off?

TM Forum autonomous networks Level 5 implies a fully autonomous network. As we have tried to outline in this article, that might be a very hard target to achieve. Apart from the question whether self-driven networks are at all feasible, it raises the question if it can be justified from the ROI perspective.

To assess that question, we need to better understand the benefits of autonomous networks.

The first aspect that comes to most people's minds is the human efforts needed for network operation and the opportunity to reduce that effort thereby creating savings in operational costs. It is obvious that for use cases that need less human effort, or do not happen frequently the savings are limited, whilst for high-frequency and time-consuming use cases the savings are larger.

A second, and perhaps more important aspect is the fact that networks are getting increasingly complex over time. The set of features, controls, and inter-system dependencies that a communication service provider needs to manage is continuously growing. An intelligent autonomous network could encapsulate some of that complexity. An AI-powered system would reduce the operating time and yield more optimal configurations.

A third aspect is OpEx costs that may be incurred from operating the networks in non-optimal conditions, that may lead to unnecessary energy being consumed, for example, or unnecessary high CapEx as equipment is procured that wasn’t needed.

A fourth aspect (that relates to the second) is the aspect of customer churn due to networks having unnecessary low performance (i.e. leading to low-quality services). This is particularly concerning as it impacts the revenues of the communication service providers.

For any capability related to autonomous networks we need to assess the potential benefits against the investments needed to achieve it. In case these capabilities are powered by AI, we need also to correctly estimate the costs for:

• data management – labelling, filtering, etc.
• training – may need to happen regularly, driving costs
• model management – the governance related to making sure those models are upgraded sufficiently.

AI-based capabilities also come with additional security issues. Autonomous networks are smart(er) and according to Hypponen’s Law the smarter it gets, the more vulnerable it becomes. Introducing AI to a complex system exposes it to new threat vectors.

This will of course lead us to realize that some use cases just don’t have a positive ROI, which means that the vision of a full autonomous network remains a vision – albeit an important one!

What is a realistic view on unsupervised AI control?

We are at the peak of the AI trend, where many talk about all the exciting opportunities, and fewer see it more realistic. For unsupervised AI control, like in high-level autonomous networks, it is important to understand the magnitude of the challenge. So, lets share some of the more sceptical (realistic) views:

In Harvard Business Review we can read:

In our view, AI still has a long way to go in making the ultimate decisions in real-world life situations that require more holistic, subjective reasoning. It still is merely a factual engine that acts based on probabilities and scores, mostly based on historical data, with no context of the implications of the information it is delivering. 

AI Isn’t Ready to Make Unsupervised Decisions

And Forbes writes:

Humans have been shown to both over-rely (automation bias) and under-rely (algorithm aversion) on algorithmic advice and fare badly at judging the accuracy of algorithmic predictions.

AI Unlikely To Ever Work Unsupervised, At Least For The Big Stuff

So, there are reasons to be somewhat sceptic about how AI can bring along a great breakthrough in achieving autonomous networks.

But at the same time, we should be aware of Amara’s Law:

“We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.”

Technological evolution is not linear but in fact exponential. It makes it hard to imagine what becomes possible in the longer run, and what sounds like fiction today may well be perfectly possible in 5-10 years from now.

Tietoevry experience in AI for telecom

Tietoevry, a leading telecom R&D provider, has been engaged in numerous AI-based projects for telecom. Our experience started well before 2017 and encompasses projects related to visual inspections of mobile base station masts and antenna feeders, algorithms for optimizing mobile networks as well as applying it for operational tasks.

During recent years we have been involved in RAN Intelligent controller projects that combine multiple AI-based applications with the purpose of reducing the energy consumption in mobile networks.

Recently, Tietoevry has joined the AI-RAN alliance where we have engaged with some of the Alliance partners in direct projects on AI-based algorithms for RAN control. We strive to contribute to more cutting-edge projects, share our knowledge, and provide insights for the industry.

Reach out to us for a discussion on how AI could be leveraged in your use cases – but let’s be real about it!