Towards zero-touch network operations | Ericsson Research Blog

Ericsson Research Blog

Research, insights and technology reflections

Towards zero-touch network operations

Imagine if mobile network operations were limited to the effort of setting the intent for how to utilize network resources in relation to the desired business logic that your operations were built upon.

Today’s mobile network operations require considerable human involvement where actions are reactive rather than proactive, and are often repeated. A prominent example is intensive configuration through parameter settings in response to a constant flow of alarms and incident reports. Another example is the challenging effort to ensure a sustainable network expansion to support the evolution of industry and society by adding more capacity, devices and new sophisticated services.

With the rapid development of Artificial Intelligence and Machine Learning – what we call Machine Intelligence at Ericsson – there are growing possibilities for making the network more automated and intelligent. When the traditional management tasks become automated and network nodes become more intelligent, it opens up the ability to steer the network with KPIs that optimize for business performance. For every achieved technology landmark, we come closer to the intent-driven network. Ultimately, a vision we call the zero-touch vision, and Ericsson CTO Erik Ekudden has identified the realization of zero touch as one of the five system-related trends to watch for the future.

There are already solutions to run the network more efficiently based on AI and ML technologies. Examples are support for site engineers that simplifies their work and limits costs for site visits, methods to predict alarms, and reducing the time to repair, if at all needed since you may have the ability to act before the error even occurs. In short, future network management organizations will be capable of focusing more on business generating tasks, while still operating a larger network with more services.

Having applied AI and ML in various parts of networks and its operations, we have learned that if you have a good understanding of the problem area, are able to define a simple enough problem statement and have relevant data available, you can often tailor a SW agent to efficiently address the specific problem. This does not mean it can always be financially motivated or that the technology you need is mature enough. Even though recent technology breakthroughs seem promising, there are still many challenges to address before the zero-touch vision can come true. Here are some relevant areas that I think need further exploration and development:

Intelligent SW agents: This is a huge area where we have seen many new applications in recent years in our daily life thanks to AI and ML development. All from basic intelligence such as a control mechanism in robotic vacuum cleaners to interactive personal services such as Alexa from Amazon and Google Assistant. Deep learning methods together with increased capabilities in hardware and availability of data are some of the vital contributors that has made this possible.

SW agent generalization: SW agents need to be trained for a given task before deployment. The success of agent training relies on how well it deals with the distinction between training and a testing environment, in other words, avoiding overfitting to the training environment. In particular, when an agent is trained only with a virtual emulation environment, it would be very challenging to make the agent run the trained task in reality where random noise and unexpected variations are present.

Robust SW agents: Even though it is possible to design and train SW agents capable of beating human world champions of various games, there is still far more challenges to have SW agents act on real world situations. A game has a limited of set of rules that the agent needs to relate to, providing an environment with limited dimensions to act in, whereas reality offers an environment with infinite number of rules with noise. An agent that is not sufficiently prepared to act on the relevant rules breaks the process and will make the system unstable. More importantly, the agent should be able to handle the unexpected incident or attack that in reality was not observed during the training.

Collaborative Robotics: Robots will play an important role to automate processes and augment humans while managing and further developing the network. A key aspect is to design the robot for collaborative trust with the human. For instance, for a human to rely on a robot he must be able to understand why certain actions were performed by the robot, because if not, the human may decide not to rely on the robot next time. Also, the robot may need to adapt to the specific capabilities of the human it is set to serve.

Federated Learning: As a Robot or SW agent works alongside a human, it may go through techniques such as reinforcement learning to be trained over time and gradually increase its performance. Assuming a service technician onsite solves new problems together with their robot assistant, the knowledge trained by such a new experience will become digitally available. The gained knowledge can be shared to other robot assistants with techniques like federated learning to gradually increase the collective knowledge.

Transfer learning: Gained knowledge from a collective of SW agents performing one task can be shared by knowledge agents performing another related task. For example, knowledge gained by operating a network in one area can be reused for network operation in another. An example is handover parameter optimization, to prevent unnecessary or untimely handovers which may affect the operation of mobile devices attached to the network. Such parameters are configured by the operator and there exists several machine learning-based approaches to automate the process of configuration. A transfer learning process could transfer the model built for configuring handover parameters at one geographical location (for example, an urban environment like a city), to another with similar characteristics.

It is still a long way to go before we have the zero-touch network operation,s but with the accelerated development we see in Artificial Intelligence and Machine Learning, the network operations of tomorrow will be very different from how we used to operate the mobile network in the past.

Find out more about what Ericsson CTO Erik Ekudden has to say about the realization of zero touch in his Ericsson Technology Review article: Five technology trends augmenting the connected society.

Viktor Berggren

Viktor Berggren is a Principle Researcher at Ericsson Research in the area of Machine Intelligence. He has over twenty years of experience from Research and Development of mobile systems. His research interests is to explore opportunities for how the fifth-generation mobile system can be better suited for future industries with digitalization ambitions. Viktor holds a Master of Science in Industrial Engineering and Management from Linköping University.

Viktor Berggren