What is autonomous networking?
Imagine you have developed a video to promote network neutrality. As you upload it to a video server, the network does not establish the required connection. You want to know what the problem is, so you ask: “Hi network, what’s wrong?” The response comes as a surprise: “The content of this video conflicts with my interests; I do not want you to upload this video!” This is autonomous networking! The network has its own identity and agenda; it acts according to its own interests.
Let’s have a look at a different scenario. This summer, as wildfires burnt across California, mobile broadband helped firefighters to coordinate their action. However, their communication system became subject to bandwidth throttling, creating significant impact on the firefighters’ ability to provide emergency services. This scenario had nothing to do with autonomous networking. The network was simply enforcing an SLA policy. Here it would have been better if the network had understood the impact of its action and prioritized public safety over business economics.
These two examples shed some light on what autonomous networking might eventually become. They also take us beyond most of the current debate around the topic. Not only technical but also ethical questions will need to be addressed. What this also makes clear is that real autonomous networking is still several years or even decades away and its ultimate value is difficult to predict. The common view in the industry is that operational savings will be tremendous. However, the architecture and operation of those intelligent networks will need to be aligned with AI paradigms, which is a time-consuming effort. Fortunately, this journey has already begun.
From analytics to AI
An intelligent algorithm can analyze and correlate big data to identify patterns and create heuristic rules. This helps an operational team to take decisions such as reconfiguring a network node, suggesting strategies for network expansion, or triggering maintenance action. Any decision can be verified by reference to the underlying data set. This data analytics-based approach makes is easy for the operational team to look at the historic data and understand the reasons for any suggestion provided by intelligent analytics.
Things are different with deep learning, an AI technology that is currently gaining a lot of attention. In this case, a multi-layered neural network is trained by observational data. These training sequences adjust parameters in hidden layers within the deep-learning network, creating different levels of abstraction between the original data and the output. A trained network becomes a black box which can assess new data and make conclusions on specific characteristics. For a human user, it’s very difficult to verify the outcome of a deep learning algorithm as the relation with the original data is separated by those hidden abstraction layers. The abstracted representation of information is almost impossible for humans to comprehend. This lack of transparency and verifiability is creating a level of discomfort in applying deep learning with critical operational processes.
Gaining confidence as an essential step towards autonomous networking
As communication service providers start to benefit from AI and deep learning, they need strategies to gain confidence with this new technology. They need to trust the intelligent algorithm as they allow it to take over functions that were previously carried out by humans. As outlined above, the challenge with operations guided or controlled by deep learning lies in how quickly the operations team can verify the output of an intelligent algorithm. Strategies for gaining confidence need to be developed. Initially the output of the intelligent algorithm will be compared with previous practice, mapped against experience from experts and verified by other analytical methods. Tools will allow operators to verify intelligent reasoning rapidly, for example by simulating complete networks in software.
Those strategies will help to smooth the process of replacing human action with artificial intelligence. There will however still be a need for human interaction by a supervisor to monitor and supervise those algorithms. The data inputted into an intelligent algorithm must be correct, comparable and complete. Situations might come up that are beyond the scope of training scenarios. An independent supervisor needs to monitor and prevent failure caused by an intelligent algorithm producing a wrong output.
Rearchitecting a network for AI
In addition, networks need to be prepared to make best use of emerging AI technologies. Two different domains must be looked at. Gathering of observational data will require a move from polling network elements for information to subscription-based telemetry streaming. And the network needs to remove the restrictions of hardware-limited feature sets through software-defined networking – namely open SDN control and NFV as well as network operating systems on white boxes.
If you’re interested in this topic, please join me at my talk, Making a move to autonomous networking, at SDN NFV World Congress in The Hague, October 9. I’ll be exploring some low-hanging AI use cases and providing guidance on preparing networks for AI-powered operation.