AI Solving Pain Points in Telco Networks, but the Path to AI Remains Rocky

ABI Research’s Artificial Intelligence in Telecom Networks report (AN-2555) explained that a growing number of operators have started integrating AI capabilities into their services. These AI capabilities feature natural language processing, the ability to perform speech recognition, and the provision of automated customer services to customers. At the same time, industry associations, such as the TM Forum and ETSI, have moved to establish working groups around zero touch and automation, finding ways to introduce AI into network deployment and operation.

In the past 2 years, the telecommunications industry has witnessed the emergence of AI. This is mainly driven by several major factors. The increasing complexity in network topology has resulted in laborious network optimization and planning processes. Pattern combinations in a 5G antenna, for example, can reach more than 1,000, due to beam pattern and antenna downtilt in massive MIMO. In addition, both vendors and operators have invested in the creation of data lakes (usually based on Hadoop MapReduce or Apache Spark) over the years. Coupled with the availability of open-source machine learning frameworks and more affordable computing hardware, machine learning-based AI has begun to gain in popularity among vendors and operators.

At the moment, AI in telco networks can be based on robotic process automation (RPA) and machine learning-based AI. While the ultimate goal for the presence of AI in telecommunications networks is to bring in automation, it is crucial to highlight the differences between RPA and machine learning. RPA refers to a set of models programmed to execute narrowly-defined tasks, focusing on precision, speed, and full autonomy. SoftBank, for example, has deployed RPA in its network maintenance procedures. Other vendors, such as Tupl, have RPA solutions focusing on customer complaint resolution and network operation automation. ABI Research believes that RPA is the perfect primer for operators to introduce AI into their network operations, before they move on to using machine learning-based AI.

On the contrary, machine learning-based AI requires both training and inference to take place. Large datasets and computing power are required for training a model. Once the training is complete, the trained model shall be inferred to resolve specific issues or challenges based on the patterns and regularities shown in the dataset. This approach is much more robust, as AI is capable of learning new data patterns and being able to address new scenarios.

At the same time, the emergence of open-source machine learning frameworks, such as TensorFlow, MXNet, and Deeplearning4J (DL4J), allows vendors and service suppliers to focus less on the creation of machine learning framework, and more on using and adapting the frameworks to their own needs. WPOTECH, a self-organizing network (SON) startup, uses H2O.ai platform to develop its standalone SON solution using Spark MLlib and DL4J. A similar approach has also been taken by other network infrastructure vendors.Nokia’s AVA platform is a standalone telco cloud analytics platform that can provide predictive analysis and preventive maintenance of network elements in the telco cloud environment, while MYCOM OSI offers closed-loop assurance, automation, and an analytics solution based on the TensorFlow framework.

In recent years, operators have been trying to emulate over-the-top (OTT) players, such as Facebook, Google, and Netflix, to make their infrastructures more agile and virtualized. While the progress on NFV has slowed down in the industry, operators using a newly virtualized telco cloud find that data acquisition has become easier. However, there is still a long way to go in the journey of NFV. In order for operators to become truly agile, all current network elements need to be in a container architecture, instead of the currently preferred virtual machine architecture. The training and inference of machine learning-based AI in the future will be performed in a container architecture, which can be a challenge when the machine learning framework requires certain heterogeneous computing resources for training.

In addition, current network automation solutions remain fragmented and loosely connected to physical elements in the network, such as the antenna and cell site. As compared to the aforementioned vendors, Huawei took an integrated approach to AI in telecommunications networks via the launch of SingleRAN Pro, a single platform that offers RAN, core networks, and AI capabilities for networks ranging from 2G to 5G. Huawei uses a wide range of frameworks and libraries, including TensorFlow, MXNet, Scikit-learn, and Spark MLlib, and supports machine learning on Kubernetes via RiseML. An integrated approach to AI will enable all of the desired data to be collected from the network architecture and allow the model to be trained in the intended manner in the cloud.

Of course, this is not to say that an integrated approach is the only way to achieve ideal AI in a telecommunications network. After all, multi-tenancy is the main reason why operators pursue NFV in the first place. To address that challenge, Tier One operators with strong software development capabilities opt to construct their own data collection and solution enablement platforms. AT&T, for example, created the AT&T Network 3.0 Indigo, a platform that can support multi-vendor solutions in big data, cybersecurity, software-defined networking, and AI. Others may choose to segment their network infrastructure into different zones and deploy integrated solutions from different vendors in different zones.

At the end of the day, AI will first be found in telco networks that are fully virtualized, regardless of the size of an operator. However, AI should not be limited to network operators. ABI Research believes that operators will move on to monetize their AI solutions. During MWC 2018, Huawei launched AUTIN, an operations consulting and software-as-a-service (SaaS) solution, to help operators manage hybrid ICT environments. Other vendors have also started deploying AI platforms that span their companies and business groups. Working together with vendors, operators must start to actively incorporate AI capabilities into new use cases and business models, such as smart homes, elderly care, video systems, and public safety, in order to maximize the investments they made in AI.