A recent ITU workshop examined the role of machine learning (ML) in ICT and 5G. While there’s been no real breakthrough app for ML yet, predictive QoS looks the most promising.
“Machine learning techniques, so far, are more or less related to prediction, classification and decision-making in the networking area,” says Qiang Cheng of the Artificial Intelligence Industry Association of the China Academy of ICT.
Cheng was highlighting the current state of play in machine learning’s contribution to ICT networking at a recent ITU workshop in Geneva. The workshop brought together experts in machine learning and ICT networking to discuss the challenges and opportunities on the agenda of the new ITU Focus Group on Machine Learning for 5G, which met for the first time from January 30 to February 2.
Looking to the future, Huawei, KT, ZTE and Deutsche Telekom see great potential for machine learning to assist the design, operation and optimization of 5G networks. Machine learning is expected to assist the ICT industry in meeting the challenges brought on by 5G and IoT – shifts representative of considerable increases in network complexity and the diversity of device requirements.
The workshop’s discussions highlighted that machine learning applications in communications networking are still very much at their nascent stage of development.
Machine learning in networking has yet to find the breakthrough applications or “high-potential success stories” required to expect near-term gains in network performance over and above what is possible today, said Jakob Hoydis of Nokia Bell Labs.
Hoydis went on to offer a qualification: “Predictive QoS, however, is something that I wouldn’t know how to do without machine learning – that’s a very promising application.”
Volkswagen’s vision for the future of automated driving is highly dependent on the capabilities expected of 5G systems. Volkswagen’s interest in the machine learning focus group hinges to large extent on the potential of machine learning to support adaptive QoS.
“We see network slicing as one of the most valuable 5G features,” explained Volkswagen’s Ahmad El Assaad, “but automotive network slices will be highly dynamic … we don’t need everything all the time.” Assaad was illustrating how machine learning could support predictive QoS, helping vehicles and 5G systems to collaborate in anticipating and adjusting to changes in vehicles’ required level of QoS.
Challenges surrounding the availability and quality of the data required to fuel ML algorithms were a recurring theme throughout the workshop’s discussions.
The lack of a unified data format remains a significant challenge to the development and training of ML algorithms. The challenge is compounded by highly distributed data, limits on available computation resources, and restrictions on bandwidth and latency.
The potential to build machine learning algorithms on datasets as well as models incorporating expert knowledge is a direction of innovation certain to be explored by the focus group.
“There is no machine learning without data,” highlights the focus group’s chairman, Slowomir Stanczak of Fraunhofer Heinrich Hertz Institut. “We have to clarify which data needs to be collected, if we need to process the data, and questions about the quality of the data – this is one of the goals of the focus group.”