No. 5, 2020

AI and Machine Learning in 5G

Lessons from the ITU Challenge Stay current // // Stay informed

ITU News has moved to a new platform. Discover MyITU your gateway to relevant ITU content, based on your interests.

Keep informed with the latest ITU News.

To receive the new weekly ITU Newsletter,

Subscribe ITU News ITU News Regular articles Magazine

Join ITU’s online communities on your favorite channel Editorial ITU News MAGAZINE No. 05, 2020 1

AI and machine learning in 5G — the ITU Challenge 2020

By Houlin Zhao, ITU Secretary‑General

J In February this year, the we announced the approval International Telecommunication by our 193 Member States of Union (ITU) set the first ITU AI/ML in an ITU Radiocommunication 5G Challenge in motion — a global Sector (ITU–R) Recommendation: competition that will culminate in “Detailed specifications of the an online prize-winning event on radio interfaces of IMT‑2020.” 15–17 December, 2020. IMT‑2020 specifications for the fifth Through the Challenge, ITU generation of mobile communica‑ Through the encourages and supports the grow‑ tions (5G) will be the backbone of ing community driving the integra‑ tomorrow’s digital economy, lead‑ Challenge, ITU tion of artificial intelligence (AI) and ing industry and society into the encourages machine learning (ML) in networks automated and intelligent world and supports and at the same time enhances the and promising to improve people’s community driving ITU standardiza‑ lives on a scale never seen before. the growing tion work for AI/ML. community In this edition of the ITU Magazine driving the The ITU Challenge enables the you will learn all about the ITU AI/ integration collaborative culture necessary for ML in 5G Challenge and also find success in emerging and future ample insight articles from industry of artificial networks such as 5G and creates and academia. intelligence and new opportunities for industry and machine learning academia to influence the evolution The Grand Challenge Finale will in networks. of ITU standards. feature keynotes by Professor Vincent Poor of Princeton University, As the UN specialized agency United States, Chih‑Lin I of the Houlin Zhao for ICTs, ITU plays a central China Mobile Research Institute, role in ensuring that these net‑ and Wojciech Samek of Fraunhofer works are rolled out widely HHI, Germany. It will also launch the and follow the highest qual‑ Challenge 2021. Enjoy! ity standards. Most recently, Contents ITU News MAGAZINE No. 05, 2020 2

AI and Machine Learning in 5G

Lessons from the ITU Challenge

Editorial

1 AI and machine learning in 5G — the ITU Challenge 2020 By Houlin Zhao, ITU Secretary‑General Cover photo: Shutterstock

5 ITU thanks the sponsors of the 2020 AI/Machine Learning in 5G Challenge

ITU AI/ML in 5G Challenge

6 Building community and trust on the ITU platform ITU News spoke with Chaesub Lee, Director of the ITU Telecommunication Standardization Bureau, to learn more about the context for the ITU Challenge on AI and Machine Learning ISSN 1020–4148 in 5G and its connection with the strategic priorities of ITU. itunews.itu.int Six issues per year Copyright: © ITU 2020 9 Message from the organizers By Thomas Basikolo, AI/ML Consultant Editorial Coordinator & Copywriter: Nicole Harper Art Editor: Christine Vanoli Editorial Assistant: Angela Smith 12 Follow the ITU AI/ML in 5G Challenge 13 Problem statements Editorial office: Tel.: +41 22 730 5723/5683 14 The Grand Challenge Finale — Tuesday, 15 December 2020 E‑mail: [email protected] 15 The Grand Challenge Finale — Wednesday, 16 December 2020 Mailing address: 16 The Grand Challenge Finale — Thursday, 17 December 2020 International Telecommunication Union Place des Nations 17 Winning prizes and certificates CH–1211 Geneva 20 (Switzerland)

18 A guide to AI/ML challenges for the next-generation CTx Disclaimer: Opinions expressed in this publication are those By Vishnu Ram OV, Independent Research Consultant of the authors and do not engage ITU. The des‑ ignations employed and presentation of mate‑ 23 A standards round-up on autonomous networks rial in this publication, including maps, do not imply the expression of any opinion whatsoever By Xiaojia Song, Researcher, Xi Cao, Senior Researcher, Lingli on the part of ITU concerning the legal status of Deng, Technical Manager, Li Yu, Chief Researcher, and Junlan any country, territory, city or area, or concerning Feng, Chief Scientist, China Mobile Research Institute the delimitations of its frontiers or boundaries. The mention of specific companies or of certain 30 ITU AI/Machine Learning in 5G Challenge webinars products does not imply that they are endorsed or recommended by ITU in preference to others of a similar nature that are not mentioned.

All photos are by ITU unless specified otherwise. Contents ITU News MAGAZINE No. 05, 2020 3

Insights from industry

32 Capability evaluation and AI accumulation in future networks By Jun Liao, Artificial Intelligence Director, Tengfei Liu, Yameng Li, and Jiaxin Wei, Artificial Intelligence Engineers, China Unicom Research Institute

35 Accelerating deep learning inference with Adlik open-source toolkit By Liya Yuan, Open Source and Standardization Engineer, ZTE

38 Challenges and opportunities for communication service providers in applying AI/ML By Salih Ergüt, 5G R&D Senior Expert, Turkcell

42 Autonomous networks: Adapting to the unknown By Paul Harvey, Research Lead, and Prakaiwan Vajrabhaya, Research Outreach and Promotion Lead, Innovation Studio, Rakuten Mobile

46 Quality of Experience testing in mobile networks By Arnd Sibila, Technology Marketing Manager, Mobile Network Testing, Rohde & Schwarz

50 A network operator’s view of the role of AI in future radio access networks By Chih‑Lin I, Chief Scientist, and Qi Sun, Senior Researcher, Wireless Technologies, China Mobile Research Institute

55 AI and open interfaces: Key enablers for campus networks By Günther Bräutigam, Managing Director, Airpuls; Renato L.G. Cavalcante, Research Fellow, and Martin Kasparick, Research Associate, Fraunhofer HHI; Alexander Keller, Director of Research, NVIDIA; and Slawomir Stanczak, Head of Wireless Communications and Networks Department, Fraunhofer HHI, Germany

58 Quotes from hosts of the ITU AI/ML in 5G Challenge problem statements 61 Quotes from the ITU AI/ML in 5G Challenge participants Contents ITU News MAGAZINE No. 05, 2020 4

Insights from academia

62 AI/machine learning for ultra-reliable low-latency communication By Andrey Koucheryavy, Chaired Professor, Telecommunication Networks and Data Transmission Department, The Bonch-Bruevich Saint Petersburg State University of Telecommunications (SPbSUT), Chief Researcher, NIIR, and Chairman, ITU–T SG11; Ammar Muthanna, Deputy Head, Science, Telecommunication Networks and Data Transmission Department, SPbSUT, and Head, SDN Laboratory; Artem Volkov, Researcher and PhD Student, Telecommunication Networks and Data Transmission Department, SPbSUT, Russia

66 AI/ML integration for autonomous networking — a future direction for next‑generation telecommunications By Akihiro Nakao, Professor, The University of Tokyo, Japan

70 Realistic simulations in Raymobtime to design the physical layer of AI‑based wireless systems By Aldebaro Klautau, Professor, Federal University of Pará, Brazil; and Nuria González‑Prelcic, Associate Professor, North Carolina State University, United States

74 Building reliability and trust with network simulators and standards By Francesc Wilhelmi, Post-Doctoral Researcher, Centre Tecnològic de Telecomunicacions de Catalunya (CTTC), Spain

78 Research projects in Nigeria advancing education and speech recognition By James Agajo, Associate Professor and Head of WINEST Research Group, Department of Computer Engineering, Abdullahi Sani Shuaibu and Blessed Guda, Students, Federal University of Technology, Minna, Nigeria

82 Why we need new partnerships for new data By Ignacio Rodriguez Larrad, PostDoc, Wireless Communication Networks, Aalborg University, Denmark

86 Machine learning function orchestration for future generation communication networks By Shagufta Henna, Lecturer in Computing, Letterkenny Institute of Technology, Ireland

88 Sponsorship opportunities for 2021 ITU thanks the sponsors of the 2020 AI/Machine Learning in 5G Challenge

Gold Sponsor TRA, United Arab Emirates

TRA

Bronze Sponsors Cisco Systems and ZTE

CISCO ZTE ITU AI/ML in 5G Challenge ITU News MAGAZINE No. 05, 2020 6

Building community and trust on the ITU platform

ITU News spoke with Chaesub Lee, Director of the ITU Telecommunication Standardization Bureau, to learn more about the context for the ITU Challenge on AI and Machine Learning in 5G and its connection with the strategic priorities of ITU.

This edition shares experiences How does the ITU Challenge from the ITU Challenge. How factor into the strategic priorities would you describe the aims of of ITU? the ITU Challenge? ^ Building community and The ITU ^ The ITU Challenge provided a trust is at the heart of all that Challenge platform for participants to apply we do at ITU. We are a global provided a ITU’s Machine Learning Toolkit membership of 193 Member platform for in solving practical problem States and over 900 companies, statements. The ITU Challenge universities, and international participants allowed participants to connect and regional organizations. ITU to apply ITU’s with new partners in the ITU standards are developed in a Machine community — and new tools and community, building the mutual data resources — to achieve goals understanding that enables the Learning set out by problem statements community to advance together. Toolkit. contributed by industry and ITU standards are significant feats academia in Brazil, China, India, of international collaboration. Ireland, Japan, Russia, Spain, They represent voluntary commit‑ Chaesub Lee Turkey and the United States. It ments to common approaches to Director, ITU Telecommunication offered participants an opportu‑ technology development, appli‑ Standardization Bureau nity to showcase their talent, test cation and supporting business their concepts on real data and relationships. The value of ITU real-world problems, and com‑ standardization, just like the value pete for global recognition. of the ITU Challenge, lies in the community that it creates. ITU AI/ML in 5G Challenge ITU News MAGAZINE No. 05, 2020 7

How do ITU standards connect Standard “toolsets”, built to be with the ITU Challenge and how 5G represents major adapted to evolving user require‑ might this connection evolve? advances in networking ments and a broad scope of use to meet the needs of cases, are also found in ITU stand‑ ^ New ITU standards for AI/ML a very diverse set of ards in fields such as multimedia, provide toolsets to enable AI/ML applications, across security, blockchain and quantum integration in 5G and future net‑ industry sectors. information technology. works as these networks evolve. The ITU Telecommunication The ICT industry evolves very Standardization Sector (ITU–T) Chaesub Lee rapidly. How have recent Y.3172 architecture — derived years’ evolutions impacted ITU from the study of use cases pub‑ standardization? lished in ITU–T Y.Supplement55 — introduced the basic toolsets ^ ITU’s standardization arm in relation to the underlying (ITU–T) has seen a very strong network: ML Pipeline for model increase in new members in the optimization and serving; ML Why are AI/ML and supporting past four years, topping over Sandbox to trial models before standards important to 5G and 50 last year. We are addressing deployment; and ML Function future networks? exciting new subjects, but the role Orchestrator (MLFO) to control of the ITU platform has remained AI/ML integration. ITU–T Y.3173 ^ Companies in the network‑ unchanged for over 150 years — (intelligence evaluation), ITU–T ing business are introducing AI/ we build community and trust to Y.3174 (data handling) and ITU–T ML as part of their innovations enable information and communi‑ Y.3176 (marketplace integration) to optimize network operations cation technology (ICT) advances all build on the ITU–T Y.3172 and increase energy and cost on a global scale. The ITU stand‑ architecture. The ITU Challenge efficiency. 5G represents major ardization platform — for many aimed to demonstrate and advances in networking to meet years central to building mutual validate these ITU standards and the needs of a very diverse set understanding within the ICT sec‑ create new opportunities for of applications, across industry tor — is now helping the ICT sector industry and academia to influ‑ sectors. Networks are growing to build mutual understanding ence their evolution. in sophistication and complexity. with its many new partners. We AI/ML will be key in managing see new partners collectively this complexity. The ITU–T Y.317x advancing ITU standardization standards provide versatile tool‑ work in fields such as smart cities, sets to support AI/ML integration energy, health care, finance, auto‑ in tune with network evolution. motive, and AI/ML. ITU AI/ML in 5G Challenge ITU News MAGAZINE No. 05, 2020 8

How has ITU approached this Where is the influence of AI/ need to support a more diverse ML most pronounced in ITU The concept of a truly set of ICT applications? standardization work and autonomous network — what are the opportunities enabled by the Level 5 ^ Although ITU’s role in build‑ to participate? intelligence described ing community and trust remains by ITU–T Y.3173 — has unchanged, we have entered ^ AI/ML is playing a key part in sparked considerable a new era of standardization ITU standardization work in fields discussion in ITU. in need of new approaches to such as network orchestration and continue building this community management, multimedia coding, and trust. We have spent many service quality assessment, digital Chaesub Lee years bringing ICT decision-mak‑ health, environmental efficiency, ers together with decision-makers and autonomous driving. And the in other sectors. This inclusive concept of a truly autonomous dialogue has helped us to create network — enabled by the Level 5 the conditions necessary to intelligence described by ITU–T deliver influential standards in Y.3173 — has sparked considera‑ fields of innovation given life by ble discussion in ITU. We wel‑ new partnerships; fields such come you to join us. as digital health, digital finance, intelligent transport systems and ITU continues to grow in inclu‑ AI/ML. Here we see the value of sivity. This year we introduced a open platforms such as ITU focus reduced membership fee option groups or the AI for Good Global for start-ups and SMEs. Academia Summit. These open platforms have benefitted from reduced help to build community and fees since 2011. Companies of trust. They help to clarify the con‑ all sizes in “low income” develop‑ tributions expected of different ing countries also benefit from stakeholders, including the contri‑ reduced fees. bution of ITU standardization. ITU AI/ML in 5G Challenge ITU News MAGAZINE No. 05, 2020 9

ITU AI/ML in 5G Challenge Applying machine learning in communication networks [email protected]

Message from the organizers

By Thomas Basikolo, AI/ML Consultant

J The ITU AI/ML in 5G Challenge rallied like-minded students and professionals from around the globe to study the practical application of artificial intelligence (AI) and machine learning (ML) in emerging The ITU AI/ML in 5G and future networks. The Challenge was a first for ITU, but with many Challenge rallied like- valuable lessons learnt, it looks to be the first of many. minded students and professionals from The Challenge welcomed over 1300 participants from 62 coun‑ around the globe. tries, forming 911 teams, and we are looking forward to the Grand Challenge Finale, 15–17 December online, where outstanding teams Thomas Basikolo will compete for a share in a prize fund totalling 20 000 CHF and a range of other prizes offering global recognition.

Partnerships made the ITU Challenge possible, and partnerships were also the name of the game. ITU AI/ML in 5G Challenge ITU News MAGAZINE No. 05, 2020 10

and our Gold sponsor, the The problem statements of this The Challenge Telecommunications Regulatory first ITU Challenge offered a welcomed over Authority (TRA) of the United variety of opportunities to apply 1300 participants Arab Emirates; and Bronze spon‑ the ITU–T Y.317x techniques, and from 62 countries, sors Cisco and ZTE. one problem statement demon‑ forming 911 teams. strated ML function orchestra‑ tor capabilities via reference Mapping solutions to ITU implementations. Thomas Basikolo standards In future editions of the ITU New ITU standards for AI/ML Challenge, we aim to provide provide toolsets that, when a reference implementation of integrated, form an end-to-end an end-to-end ML pipeline as pipeline for AI/ML integration defined by ITU–T Y.3172. Such in networks. The ITU Challenge reference implementations could The ITU Challenge enabled aimed to demonstrate and vali‑ include notebooks for ML cod‑ participants to connect with date these ITU standards. In map‑ ing and integration; tools for new partners in industry and ping solutions to ITU standards, data processing and manage‑ academia — and new tools and the ITU Challenge contributes to ment; and tools for ML model data resources — to solve real- the growth of the community able selection, training, optimization world problems with AI/ML, to support the iterative evolution and verification. showcase their talent and share of these ITU standards. new experiences. Twenty-three We also aim to enable access to problem statements were contrib‑ The ITU–T Y.3172 architecture ITU–standard toolsets for initia‑ uted by industry and academia — derived from the study of tives such as plugfests and hack‑ in Brazil, China, India, Ireland, use cases published in ITU–T athons and to set the stage for Japan, Russia, Spain, Turkey and Y.Supplement55 — introduced the collaboration in open-source pro‑ the United States, and these basic toolsets in relation to the jects and standardization work. “regional hosts” offered resources underlying network: ML Pipeline and expert guidance to sup‑ for model optimization and serv‑ port participants in addressing ing; ML Sandbox to trial models A learning experience their challenges. before deployment; and ML for all Function Orchestrator (MLFO) to We would like to thank the control AI/ML integration. ITU–T Data availability is a key challenge community that gave life to Y.3173 (intelligence evaluation), to be navigated when bringing the Challenge, our partici‑ ITU–T Y.3174 (data handling) together a global community to pants and regional hosts; our and ITU–T Y.3176 (marketplace innovate with AI/ML. promotion partners LF AI & integration) all build on the ITU–T Data, NGMN and SGInnovate; Y.3172 architecture. ITU AI/ML in 5G Challenge ITU News MAGAZINE No. 05, 2020 11

Fifteen problem statements In our work to offer partici‑ were open to all participants. pants a level playing field, ITU Preparations for the Eight were limited to partici‑ and our partners developed ITU Challenge 2.0 are pation under conditions set by tailored workflows delivering in motion, driven by a their hosts. And fourteen remain participants a unique, customized core team of challenge “under development” without the Challenge experience. management board necessary tools or data resources members, judges, for this first ITU Challenge. We ITU engaged participants in tech‑ promotion partners hope to see new partners coming nical roundtables and webinars and sponsors. together to address these four‑ to provide expert guidance in teen problem statements in future addressing problem statements editions of the ITU Challenge. and the value of new ITU stand‑ Thomas Basikolo ards in support. Together with The data sharing guidelines of the our regional hosts, we reached ITU Challenge incorporate a wide out in local languages, connected range of perspectives from indus‑ participants with mentors and try and academia on access to maintained interactive discussions real network data, synthetic data on our Slack channel. and open data. The guidelines problem statements, and new describe measures to enable data tools and data resources. We are sharing in view of different classifi‑ Up to the challenge creating new opportunities for cations of datasets, pre-process‑ in 2021? industry and academia to solve ing steps (including anonymizing) problems together, and new and secure hosting of data. Preparations for the ITU opportunities to influence the Challenge 2.0 are in motion, direction of ITU standards devel‑ We also saw the best outcomes driven by a core team of chal‑ opment and application. Contact achieved in close collaboration. lenge management board mem‑ us to participate in the prob‑ The Challenge highlighted that bers, judges, promotion partners lem-solving, judge some of the problem statements are best and sponsors. interesting submissions, promote positioned for success when sup‑ the challenge, sponsor a prize, or ported not only by the necessary We will continue to encourage mentor a few students. tools and data resources, but also new partnerships in AI/ML and by close collaboration between establish guiding principles for We thank you for your support participants and regional hosts. the sharing of tools and data and look forward to seeing you resources necessary to enact soon in Challenge 2.0. Our priority was to create com‑ these partnerships. We are wel‑ munity value in the field of AI/ML. coming new partners and new Follow the ITU AI/ML in 5G Challenge

26 partners (telecom operators, vendors, See Challenge and academia) hosted 23 problem website statements 1300+ participants from 60+ countries from 6 regions 45% industry — 55% academia 26 webinars Don’t miss the Grand Challenge 4 technical tracks: Networks, Enablers, Finale winner Verticals, Social Good announcements 20 000 CHF in cash prizes 15–17 Certificates —5 categories December 2020 online here

Timeline

Global Global Grand call for round Best teams Challenge challenge begins advance Finale — entries online

Problem statement selection Keynotes Dataset release Winners’ presentations Registration Warm-up Global Prize awards phase round ends

February 2020 to June 2020 July 2020 November 2020 15–17 December 2020 Problem statements

Title Host entity

ML5G-PHY-beam-selection: Machine learning applied to the physical layer of Federal University of Pará (UFPA), Brazil millimeter-wave MIMO systems

Improving the capacity of IEEE 802.11 WLANs through machine learning Pompeu Fabra University (UPF), Spain

Graph Neural Networking Challenge 2020 Barcelona Neural Networking Center (BNN‑UPC), Spain

Compression of deep learning models ZTE

5G+AI (smart transportation) Jawaharlal Nehru University (JNU), India

Improving experience and enhancing immersiveness of video conferencing Dview and collaboration

5G+ML/AI (dynamic spectrum access) Indian Institute of Technology Delhi (IITD)

Privacy preserving AI/ML in 5G networks for healthcare applications Centre for Development of Telematics (C‑DOT)

Shared experience using 5G+AI (3D augmented + virtual reality) Hike, India

Demonstration of Machine Learning Function Orchestrator (MLFO) capabilities Letterkenny Institute of Technology (LYIT), Ireland via reference implementations

ML5G-PHY-channel estimation: Machine learning applied to the physical layer North Carolina State University, United States of millimeter-wave MIMO systems

NEC, RISING Committee, Telecommunication Network state estimation by analysing raw video data Technology Committee (TTC)

Analysis on route information failure in IP core networks by NFV-based KDDI, RISING Committee, Telecommunication test environment Technology Committee (TTC)

Using weather info for radio link failure (RLF) prediction Turkcell

Traffic recognition and Long-term traffic forecasting based on AI algorithms St. Petersburg State University of Telecommunications and metadata for 5G/IMT‑2020 and beyond (SPbSUT)

5G+AI+AR China Unicom (Zhejiang Division)

Fault localization of loop network devices based on MEC platform China Unicom (Guangdong Division)

Configuration knowledge graph construction of loop network devices based China Unicom (Guangdong Division) on MEC architecture

Alarm and prevention for public health emergency based on telecom data China Unicom (Beijing Division)

Energy-saving prediction of base station cells in mobile communication network China Unicom (Shanghai Division)

Core network KPI index anomaly detection China Unicom (Shanghai Division)

Network topology optimization China Mobile

Out of service (OoS) alarm prediction of 4/5G network base station China Mobile The Grand Challenge Finale — Tuesday, 15 December 2020

Time Challenge Title Team Members Affiliation (CET)

12:15 5G+AI+AR Jiawang Liu Jiaping Jiang CITC and China Unicom

Analysis on route information failure in IP core 12:30 Fei Xia Aerman Tuerxun Jiaxing Lu Ping Du The University of Tokyo networks by NFV-based test environment

Analysis on route information failure in IP core Nara Institute of Science and 12:45 Takanori Hara Kentaro Fujita networks by NFV-based test environment Technology, Japan

Analysis on route information failure in IP core Ryoma Kondo Takashi Ubukata Kentaro Matsuura 13:00 The University of Tokyo networks by NFV-based test environment Hirofumi Ohzeki

Fault localization of network devices based on MEC Guochuang Software 13:15 Zhang Qi Lin Xueqin Platform Co. Ltd

Han Zengfu Wang Zhiguo Zhang Yiwei 13:30 Network topology optimization China Mobile Shandong Wu Desheng Li Sicong

Gang Zhouwei Rao Qianyin Feng Zezhong 13:45 Network topology optimization China Mobile Guizhou Xi Lin Guo Lin

14:00 Break Break Break

Energy-Saving Prediction of Base Station Cells in 14:15 Wei Jiang Shiyi Zhu Xu Xu AsiaInfo Technologies Ltd Mobile Communication Network

Out of Service (OoS) alarm prediction of 4/5G 14:30 Zhou Chao Zheng Tianyu Jiang Meijun Nankai University network base station

Demonstration of Machine Learning Function 14:45 Orchestrator (MLFO) capabilities via reference Abhishek Dandekar Technical University Berlin implementations

ML5G-PHY-beam-selection: Machine learning Mahdi Boloursaz Mashhadi Tze‑Yang Tung 15:00 applied to the physical layer of millimeter-wave Imperial College London Mikolaj Jankowski Szymon Kobus MIMO systems

ML5G-PHY-beam-selection: Machine learning Batool Salehihikouei Debashri Roy Northeastern University, 15:15 applied to the physical layer of millimeter-wave Guillem Reus Muns Zifeng Wang Tong Jian Brazil MIMO systems

ML5G-PHY-beam-selection: Machine learning 15:30 applied to the physical layer of millimeter-wave Zecchin Matteo Eurecom, Brazil MIMO systems

Improving the capacity of IEEE 802.11 WLANs Pompeu Fabra University, 15:45 Ramon Vallès through machine learning Spain

Improving the capacity of IEEE 802.11 WLANs Paola Soto David Goez Miguel Camelo University of Antwerp, 16:00 through machine learning Natalia Gaviria Belgium

Mohammad Abid Ayman M. Aloshan Improving the capacity of IEEE 802.11 WLANs 16:15 Faisal Alomar Mohammad Alfaifi Saudi Telecom through machine learning Abdulrahman Algunayyah Khaled M. Sahari

Note: The above teams have been selected to make presentations at the Grand Challenge Finale (Finale Conference). (Each team has 8 minutes for its presentation, followed by a 7‑minute Q&A with the judges and the audience).

Don’t miss the Final Conference! Take a look at the list of best teams. Register here. The Grand Challenge Finale — Wednesday, 16 December 2020

Time Challenge Title Team Members Affiliation (CET) Network state estimation by analysing raw Osaka Prefecture University, 12:00 Yuusuke Hashimoto Yuya Seki Daishi Kondo video data Japan The Kyoto College of Network state estimation by analysing raw 12:15 Yimeng Sun Badr Mochizuki Graduate Studies for video data Informatics, Japan National Institute of Network state estimation by analysing raw Fuyuki Higa Gen Utidomari Ryuma Kinjyo 12:30 Technology, Okinawa video data Nao Uehara College, Japan Institute of Computing 12:45 Compression of deep learning models Yuwei Wang Sheng Sun Technology Chinese Academy of Sciences Satheesh Kumar Perepu Saravanan Mohan 13:00 Compression of deep learning models Ericsson Research India Vidya G Thrivikram G L Sethuraman T V Atheer K. Alsaif Nora M. Almuhanna 13:15 5G+AI (smart transportation) Saudi Telecom Company Abdulrahman Alromaih Abdullah O. Alwashmi Privacy preserving AI/ML in 5G networks for Mohammad Malekzadeh Mehmet Emre Ozfatura 13:30 Imperial College London healthcare applications Kunal Katarya Mital Nitish Shared experience using 5G+AI Easyrewardz Software 13:45 Nitish Kumar Singh (3D augmented + virtual reality) Services 14:00 Break Break Break Loïck Bonniot Christoph Neumann 14:15 Graph Neural Networking Challenge 2020 InterDigital; Inria/Irisa François Schnitzler François Taiani Nick Vincent Hainke Stefan Venz 14:30 Graph Neural Networking Challenge 2020 Fraunhofer HHI, Germany Johannes Wegener Henrike Wissing Martin Happ Christian Maier Jia Lei Du Salzburg Research 14:45 Graph Neural Networking Challenge 2020 Matthias Herlich Forschungsgesellschaft Using weather info for radio link failure (RLF) 15:00 Dheeraj Kotagiri Anan Sawabe Takanora Iwai NEC Corporation prediction Using weather info for radio link failure (RLF) Juan Samuel Pérez Amín Deschamps Santo Domingo Institute of 15:15 prediction Willmer Quiñones Yobany Díaz Technology (INTEC) Traffic recognition and long-term traffic forecasting Ufa State Aviation Technical Ainaz Hamidulin Viktor Adadurov Denis Garaev 15:30 based on AI algorithms and metadata for 5G/ University (USATU) Artem Andriesvky IMT‑2020 and beyond University, Russia ML5G-PHY-Channel Estimation: Machine Learning 15:45 Applied to the Physical Layer of Millimeter-Wave Dolores Garcia Joan Palacios Joerg Widmer IMDEA Networks MIMO Systems ML5G-PHY-Channel Estimation: Machine learning Emil Björnson Pontus Giselsson Linköping University and 16:00 applied to the physical layer of millimeter-wave Mustafa Cenk Yetis Özlem Tugfe Demir Lund University, Sweden MIMO systems Chandra Murthy Christo Kurisummoottil Thomas ML5G-PHY-Channel Estimation: Machine learning Eurecom, France, Indian Marios Kountouris Rakesh Mundlamuri 16:15 applied to the physical layer of millimeter-wave Institute of Science, India Sai Subramanyam Thoota MIMO systems Communications, Canada Sameera Bharadwaja H

Note: The above teams have been selected to make presentations at the Grand Challenge Finale (Finale Conference). (Each team has 8 minutes for its presentation, followed by a 7‑minute Q&A with the judges and the audience).

Don’t miss the Final Conference! Take a look at the list of best teams. Register here. The Grand Challenge Finale — Thursday, 17 December 2020

Time Programme (CET)

11:30–12:00 Join session to test connection 12:00–12:30 Opening ceremony Welcome remarks Houlin Zhao, ITU Secretary-General Chaesub Lee, Director, ITU Telecommunication Standardization Bureau United Arab Emirates Telecommunications Regulatory Authority Overview of the 2020 Challenge Thomas Basikolo, ITU 12:30–12:55 Keynote — Recent advances in federated learning for communications Wojciech Samek, Head of Machine Learning Group, Fraunhofer HHI 12:55–13:40 Special Session: Vision for the future — AI/ML in 5G roadmap Regulator perspective Telecommunications Regulatory Authority, United Arab Emirates Industry perspective Cisco Industry perspective Wei Meng, Director of Standard and Open Source Planning, ZTE Corporation ​13:40–14:05 ​Keynote — The unfinished journey of network AI Chih-Lin I, Chief Scientist, Wireless Technologies, China Mobile Research Institute 14:05–14:30 Keynote — Learning at the wireless edge H. Vincent Poor, Professor of Electrical Engineering, Princeton University, United States 14:30–15:15 Winners’ presentations 15:15–15:30 Award announcements Prizes and certificates

15:30–15:35 Call for papers to special issue of ITU Journal on Future and Evolving Technologies (ITU J-FET): “AI/ML Solutions in 5G and Future Networks” Ian Akyildiz, Editor-in-Chief, Georgia Institute of Technology, United States 15:35–15:45 2021 outlook for Challenge 2.0 Vishnu Ram, Independent Researcher 15:45–16:00 Closing ceremony Closing remarks Hosts of the ITU AI/ML in 5G Challenge 2020 Chaesub Lee, Director, ITU Telecommunication Standardization Bureau

Don’t miss the Final Conference! Register here. Winning prizes and certificates

Teams from various problem statements will compete for the ITU AI/ML in 5G Challenge Champion title, and several awards will be presented to winning solutions at the Grand Challenge Finale taking place 15–17 December 2020.

Winners’ certificate: Awarded to winning teams in the following categories:

1st prize: 2nd prize: 3rd prize: ITU AI/ML in 5G ITU AI/ML in 5G ITU AI/ML in 5G Challenge Gold Challenge Silver Challenge Bronze Champion Champion Champion

Cash prize: 5000 CHF Cash prize: 3000 CHF Cash prize: 2000 CHF

Three Runners up will receive 1000 CHF each.

Judges Prize certificates: Awarded to winners of each problem statement as recommended by the host (excluding those under Winners certificate). Each winner receives 300 CHF.

Honorable mention certificate.

Encouragement/Community award certificate: Awarded to teams that were active during the mentoring programme and successfully submitted a solution.

Certificate of completion: Awarded to teams that completed the challenge by submitting a solution. ITU AI/ML in 5G Challenge ITU News MAGAZINE No. 05, 2020 18 Shutterstock

A guide to AI/ML challenges for the next-generation CTx

By Vishnu Ram OV, Independent Research Consultant

J The new CTx* at FutureXG repository that CTx was banking analysed the reports on on was being pulled in a zillion the screen. directions. And the buzz around autonomous networks meant (x+1)G specification delayed. xG that every part of the network And the buzz around deployments yet to be justified. was working on its own brand of autonomous networks Research & development lost in a autonomy. meant that every part maze of acronyms, old and new. of the network was New architecture diagrams every Will CTx survive this challenge? working on its own few weeks. New use cases to brand of autonomy. support in every market. Applying New ITU standards describe con‑ and integrating AI/machine cepts to enable AI/ML integration learning (ML) in the network was in 5G and future networks as nothing smooth. The open source these networks evolve.

*Any resemblances or similarities with real-life CTOs are purely futuristic. Disclaimer: This article contains some fictional information which may be referred to as forward-looking statements.

Disclaimer: Any resemblances or similarities with real-life CTOs are purely futuristic. This article contains some “fictional information” which may be referred to as “forward-looking statements”. ITU AI/ML in 5G Challenge ITU News MAGAZINE No. 05, 2020 19

The ITU–T Y.3172 architec‑ evaluation), ITU–T Y.3174 (data in the underlying network. Using ture, derived from the study of handling) and ITU–T Y.3176 (mar‑ the concepts described by the use cases published in ITU–T ketplace integration) all build on ITU–T Y.317x standards, even as Y.Supplement55, introduced the ITU–T Y.3172 architecture. the underlying network architec‑ basic toolsets including the ML ture changes from generation to Pipeline, ML Sandbox and ML Together these ITU standards the next, it will remain possible to Function Orchestrator (MLFO) in provide powerful toolsets — specify AI/ML integration using relation with the underlying net‑ standard toolsets — for operators the common terminology pro‑ work. ITU–T Y.3173 (intelligence to monitor and adapt to changes vided by ITU.

High level architecture for integration of AI/ML in networks (ITU–T Y.3172)

Management ML sandbox subsystem subsystem 6 Other SRC C PP M P D SINK management 2 1 and Simulated ML underlay networks orchestration functions 3

ML pipeline subsystem 7

M P D SINK Level-N MLFO 8 … 5 SRC C PP Level-2

9

ML intent SRC Level-1

ML = Machine learning MLFO = Machine learning function orchestrator 4 SRC = Source of data C = Collector ML underlay networks PP = Pre-producer M = Model P = Policy NF 1 … NF n NF 1 … NF n D = Distributor SINK = Target of ML output Underlay Network 1 Underlay Network 2 NF = Network function ITU AI/ML in 5G Challenge ITU News MAGAZINE No. 05, 2020 20

Real network data adds to The ITU–T Y.317x concept of ML New alerts pop up on the accuracy of the models. Pipeline and ML Sandbox man‑ the screen from the “[ML‑usecase-1xx::status::ready]” aged by MLFO enables operators MLFO monitor. What? CTx inputs to the message-box. to decouple the underlying net‑ A network-update alert! work from the AI/ML integration. The MLFO described by ITU–T Y.3172 is a logical node that man‑ At reference point 7, the ITU–T ages and orchestrates the nodes Y.3172 architecture allows the in an ML pipeline. ITU–T Y.3173 tracking of changes in the under‑ (intelligence evaluation) describes lying network and the application a key architecture scenario for the of optimizations and configura‑ evaluation of network intelligence tions in the ML pipeline by the levels by the MLFO. ITU–T Y.3174 MLFO. The architecture scenario The details of a new use case (data handling) describes the described by ITU–T Y.3173 (intel‑ arrive in the message-box. CTx sequence diagrams correspond‑ ligence evaluation) also includes runs it through the Intent-parser ing to the instantiation of various monitoring the intelligence level tool. Interesting, but how to components of the ITU–T Y.317x of each node of an ML pipeline implement it? CTx finds an ITU toolsets, based on the incoming by the MLFO. webinar on MLFO orchestration ML Intent from the operator. for managed AI/ML integration. The draft ITU standard A few API calls later, CTx is ready In combination with MLFO, ML Y.ML‑IMT2020‑MODEL-SERV aims with a tentative ML pipeline. Sandbox provides a managed to provide an architectural frame‑ environment for operators to work supporting the efficient CTx kicks off simulations in the train, test and validate ML models optimization of ML models for ML Sandbox while waiting for before they are deployed in the heterogeneous hardware envi‑ approval to access real network live network. The data handling ronments, flexible deployment of data. Digital “twins” rev into mechanism defined in ITU–T ML models for different use-case action; data is generated based Y.3174 further allows the addi‑ scenarios, and effective interfaces on previous patterns and models tion of new sources of data and in the ML pipeline when a serving are trained in the ML Sandbox, other scenarios. model is deployed. all while the Approval Authority takes its time. CTx sends the New alerts pop up on the screen results from the ML Sandbox from the MLFO monitor. What? trial models. This has the desired A network-update alert! As usual, effect. The approval arrives in the an unscheduled virtualized message-box. network function upgrade by the vendor. Do we need to rework the whole ML pipeline? ITU AI/ML in 5G Challenge ITU News MAGAZINE No. 05, 2020 21

CTx parses a new message in ITU–T Y.3176 supports the admin‑ Done! A new ML pipeline in place the message-box “[ML‑usecase- istration of different types of ML in the ML Sandbox, tested and 1xx::Evaluate::partner.edu::‑ marketplaces, internal or external, verified for the new use case. model.url]”. Pioneering algorithm and ML marketplace federation. “[status::ready]” CTx inputs to work at a partner university had The APIs defined in ITU–T Y.3176 the message-box. “[status::ap‑ produced a wrapped model enable marketplaces to find and proved]” the Approval Authority suitable for the use case. But the select ML models in other mar‑ responds. CTx schedules an Approval Authority needs an eval‑ ketplaces and pull from federated update of the network. uation of the model. Hopefully marketplaces. And they ena‑ the external ML marketplace com‑ ble marketplaces to exchange Meanwhile, unknown to CTx, a plies to ITU–T Y.3176! CTx pulls updated ML models and interact CTx-software-update package the model from the marketplace. with ML Sandboxes. had arrived in the message-box. It was time for evolution and a ML marketplace integration can new CTx-agent to take over. help network operators to follow the ML innovation curve.

The ML model metadata, ML marketplace requirements and Architecture for ML marketplace integration the architecture reference points in network (ITU–T Y.3176) defined in ITU–T Y.3176 (mar‑ ketplace integration) enable the External ML marketplace efficient exchange and deploy‑ Management ment of ML models using stand‑ subsystem 12 (Optional) ard interfaces. Not only can this Other method help solve networking management 15 and Internal ML marketplace problems using ML techniques, orchestration but also has the potential to share functions 13 and monetize ML techniques. 7 6 MLFO ML sandbox subsystem 14

3 5 ML intent ML pipeline subsystem

4

ML underlay networks ML intent input to MLFO Reused reference points (ITU–T Y.3172) ML = Machine learning New reference points MLFO = Machine learning function orchestrator ITU AI/ML in 5G Challenge ITU News MAGAZINE No. 05, 2020 22

About ITU AI/ML in 5G Challenge On top of the ability to ITU AI/ML in 5G Challenge provided a platform for adapt and improve network participants to apply the ITU–T Y.317x techniques in management and control, an solving practical problem statements. A varied selection autonomous network could of topics including beam selection, WLAN capacity self-evolve through online analysis, network state analysis, network slicing and experimentation, enabling better traffic forecasting, radio link failure prediction, the optimization of deep learning models, and MLFO compositions of controllers reference implementations, were offered in the and controller hierarchies. Challenge. Different types of data, including data from real networks, were provided in some cases for developing solutions to these problems. Vishnu Ram OV

The concept of a truly autono‑ and improve network manage‑ Level 5 intelligence. The disaggre‑ mous network — enabled by the ment and control, an autonomous gation of network components, Level 5 intelligence described by network could self-evolve through rapid DevOps and better and ITU–T Y.3173 — sparked consider‑ online experimentation, enabling better AI/ML models meant more able discussion in the ITU–T Focus better compositions of controllers work in the AI/ML integration. Group on “machine learning for and controller hierarchies. future networks including 5G” CTx.v2 searches the context and this discussion continues in CTx.v2 scanned the environment. for solutions. the ITU’s standardization expert group for “future networks and ML Pipelines, Sandboxes and ML Perhaps time for another cloud”, ITU–T Study Group 13. marketplaces are in place and ITU AI/ML in 5G Challenge? MLFO reports are green, but CTx.v2 logs into the Geneva Autonomous networks would issues remain. Divergent data for‑ Sandbox of ITU and triggers display the “self” properties: mats impacting latency between “[AI‑ML‑Challenge::v2::init]”, but the ability to monitor, operate, the ML pipelines in the network. that’s another day, another story recover, heal, protect, optimize A multitude of open-source tool‑ (for CTx.v3). and reconfigure themselves. kits to integrate. More challenges On top of the ability to adapt in the demand mapping towards ITU AI/ML in 5G Challenge ITU News MAGAZINE No. 05, 2020 23 Shutterstock

A standards round-up on autonomous networks

By Xiaojia Song, Researcher, Xi Cao, Senior Researcher, Lingli Deng, Technical Manager, Li Yu, Chief Researcher, and Junlan Feng, Chief Scientist, China Mobile Research Institute

J Mobile networks are evolving into the intelligence era with multiple application scenarios, features, services and operation requirements. Technologies including artificial intelligence (AI) are expected to enable autonomous networks in areas such as network planning, deployment, Mobile networks are operation, optimization, service deployment, and quality assurance. evolving into the intelligence era with Most of the standards development organizations (SDOs), e.g. the ITU multiple application Telecommunication Standardization Sector (ITU–T), 3GPP, ETSI, and scenarios, features, CCSA, are actively developing standards for autonomous networks. services and operation requirements. Industry bodies such as GSMA, TM Forum, and the Global TD-LTE Initiative (GTI) are working to promote autonomous networks. GSMA stated that the automatic network operation capability will become the indispensable 4th dimension of the 5G era together with enhanced mobile broadband (eMBB), massive machine type communications (mMTC) and ultra-reliable low-latency communications (URLLC), and become one of the most important driving forces for 5G service innova‑ tion and development. ITU AI/ML in 5G Challenge ITU News MAGAZINE No. 05, 2020 24

There are discussions among Since industrial convergence For example, a rule-based policy SDOs about the level of autono‑ is the key for reducing the cost engine could be one of the mous capabilities in networks (see for any single vendor or single common functional modules to framework approach in Table 1). network operator, building an support both timed control tasks open collaboration platform (see in Level 1, imperative closed The study of autonomous net‑ Figure 1) for cohesively develop‑ loops in Level 2, and adding work levels (ANL) can provide ing both a reference implementa‑ intent-to-rule translation modules reference and guidance to tion for case-agnostic functional in Levels 3 and 4. operators, vendors and other architecture and standardized participants of the telecommuni‑ external or internal interfaces The main activities on autono‑ cation industry for autonomous would be the easy way for com‑ mous networks in the SDOs and networks, standardization works munication service providers industry bodies are presented and roadmap planning. (CSPs) to kick off and stay in the in Figure 2 and briefly intro‑ converged direction towards duced below. network autonomy.

Table 1 — Framework approach for classification of autonomous network intelligence level (source: ITU–T Y.3173)

Dimensions

Network intelligence level Action implemen- Data Analysis Decision Demand tation collection mapping

L0 Manual network operation Human Human Human Human Human

Human and Human and L1 Assisted network operation Human Human Human System System

Human and Human and L2 Preliminary intelligence System Human Human System System

Human and Human and L3 Intermediate intelligence System System Human System System

Human and L4 Advanced intelligence System System System System System

L5 Full intelligence System System System System System

NOTE 1 — For each network intelligence level, the decision process has to support intervention by human being, i.e., decisions and execution instructions provided by a human being have the highest authority. NOTE 2 — This table may be used to only determine the network intelligence level for each dimension (and not the overall network intelligence level). ITU AI/ML in 5G Challenge ITU News MAGAZINE No. 05, 2020 25

Figure 1 — Open industrial collaboration towards autonomous networks

Open implementation SLA intent Open standards

RT control Intent OPS intent RT data Analysis Knowledge Analysis Knowledge RT assistance Training Data lake Rule Offline control RT collection RT control Training Knowledge Offline data

Case-agnostic Data-lake Rule Offline assistance Functional components RT collection RT control Case specific RT = Real-time data/control/assistance SLA = Service Level Agreement Managed objects OPS = Operations

Figure 2 — Main activities of SDOs and industry bodies

ITU–T Y.3172 ITU–T Y.3173 ITU–T ITU–T Y.3174 ITU–T Y.3176 ITU–T Y.ML‑IMT2020‑RAFR

IDM, SON, MDT, COSLA 3GPP ANL MDA, NWDA

Telecommunication network planning application based on AI CCSA Grading method for intelligent capability Technical requirements for ANL

AN Whitepaper 1.0 AN Whitepaper 2.0 TM Forum IG 1193 Vision & Roadmap v1.0 Autonomous Network catalyst Business requirements & architecture v1.1

Global AI Challenge GSMA AI in Network use cases in China

2018 2019 2020 2021 SDO General Management Architecture Data Use cases Industry parties ITU AI/ML in 5G Challenge ITU News MAGAZINE No. 05, 2020 26

Table 2 — ITU–T standardization activities on AI/ML and Activities in ITU AI-based autonomous networks

In ITU–T, Study Group 13 focuses on Reference number Title future networks and network aspects Supplement 55 to Machine learning in future networks including IMT‑2020: of mobile telecommunications. The Y.3170 Series use cases Architectural framework for machine learning in future Focus Group on Machine Learning ITU–T Y.3172 networks including IMT‑2020 for Future Networks including 5G Framework for evaluating intelligence level of future networks (FG-ML5G), active from January 2018 ITU–T Y.3173 including IMT‑2020 until July 2020, had been set up to Framework for data handling to enable machine learning in ITU–T Y.3174 study interfaces, network architec‑ future networks including IMT‑2020 tures, protocols, algorithms and data Machine learning marketplace integration in future networks ITU–T Y.3176 formats. Of FG ML5G’s ten technical including IMT‑2020 specifications, four have already been Requirements, architecture and design for machine learning FG-ML5G spec turned into ITU Recommendations function orchestrator

(standards), one into a Supplement, Machine Learning Sandbox for future networks including FG-ML5G spec and the other five are in the process IMT‑2020 requirements and architecture framework of being turned into ITU standards. Machine learning based end-to-end network slice FG-ML5G spec Recommendations on the AI-based management and orchestration Vertical-assisted Network Slicing Based on a Cognitive autonomous network, e.g. ITU–T FG-ML5G spec Framework Y.ML‑IMT2020‑RAFR, are currently in Architecture framework for AI-based network automation of draft stage (see Table 2). Draft ITU–T resource adaptation and failure recovery for future networks Y.ML‑IMT2020‑RAFR including IMT‑2020

Activities in 3GPP

Autonomous networks came in sight of 3GPP in the 4G era. The topics Table 3 — Standardization activities in 3GPP RAN mainly focused on Self-Organizing Networks (SON) and Minimization of TS/TR Title Drive Tests (MDT). In the 5G era, 3GPP Study on RAN-centric data collection and utilization for LTE and 3GPP TR 37.816 undertakes standardization efforts to NR enable autonomous networks: 3GPP TS 38.314 New Radio (NR); Layer 2 measurements

3GPP TS 38.300 NR; Overall description; Stage 2 3GPP RAN: RAN data collection 3GPP TS 37.320 Minimization of Drive Tests (MDT); Overall description; Stage 2 (TR 37.816), SON/MDT (TS 38.314, 3GPP TS 38.306 NR; User Equipment (UE) radio access capabilities TS 38.300, TS 37.320, TS 38.306, TS 38.331, etc.) (see Table 3). 3GPP TS 38.331 NR; Radio Resource Control (RRC); Protocol specification TS=Technical Specification, TR=Technical Report. ITU AI/ML in 5G Challenge ITU News MAGAZINE No. 05, 2020 27

3GPP SA2: Network Data Analytics Table 4 — Standardization activities in 3GPP SA2 (NWDA) (TR 23.791, TR 23.288, TR 23.700‑91) (see Table 4). TS/TR Title 3GPP TR 23.791 Study of Enablers for Network Automation for 5G 3GPP SA5: Management Data Architecture enhancements for 5G System to support network 3GPP TS 23.288 Analytics (MDA) (TR 28.809), data analytics services

autonomous networks levels 3GPP TR 23.700‑91 Study on Enablers for Network Automation for 5G — Phase 2 (TR 28.810, TS 28.100, intent TS=Technical Specification, TR=Technical Report. driven management (TR 28.812, TS 28.312), closed loop SLS assurance (TR 28.805, TR 28.535, TR 28.536, etc.), SON (TR 28.861, Table 5 — Standardization activities in 3GPP SA5 TS 28.313) and MDT (TS 28.313, TS 32.42X series) (see Table 5). TS/TR Title 3GPP TR 28.809 Study on enhancement of Management Data Analytics (MDA)

Study on concept, requirements and solutions for levels of 3GPP TR 28.810 Activities in ETSI autonomous network 3GPP TS 28.100 Management and orchestration; Levels of autonomous network ETSI is actively studying autonomous Telecommunication management; Study on scenarios for Intent 3GPP TR 28.812 networks and has several groups driven management services for mobile networks

working on the following rele‑ 3GPP TS 28.312 Intent driven management services for mobile networks vant topics: Telecommunication management; Study on management 3GPP TR 28.805 aspects of communication services ENI (Experiential Management and orchestration; Management services for 3GPP TS 28.535 Networked Intelligence). communication service assurance; Requirements

Management and orchestration; Management services for 3GPP TS 28.536 NFV (Network communication service assurance; Stage 2 and Stage 3 Functions Virtualization). 3GPP TR 28.861 Study on the Self-Organizing Networks (SON) for 5G networks

3GPP TS 28.313 Self-Organizing Networks (SON) for 5G networks OSM (Open Source MANO 3GPP TS 32.42X series (Management and Orchestration). Telecommunication management; Subscriber and equipment 3GPP TS 32.421 trace; Trace concepts and requirements MEC (Multi-access Telecommunication management; Subscriber and equipment Edge Computing). 3GPP TS 32.422 trace; Trace control and configuration management

Telecommunication management; Subscriber and equipment 3GPP TS 32.423 F5G (Fifth Generation trace; Trace data definition and management Fixed Network). Telecommunication management; Performance Management 3GPP TS 32.425 (PM); Performance measurements Evolved Universal Terrestrial Radio Access Network (E-UTRAN)

Telecommunication management; Performance Management 3GPP TS 32.426 (PM); Performance measurements Evolved Packet Core (EPC) network

TS=Technical Specification, TR=Technical Report. ITU AI/ML in 5G Challenge ITU News MAGAZINE No. 05, 2020 28

In November 2019, ETSI pub‑ Activities in industry SDOs should continue lished the report “Experiential their relevant Networked Intelligence (ENI): ENI Industry bodies such as GSMA, standardization work Definition of Categories for AI TM Forum and GTI are exploring and play a leading Application to Networks” (ETSI GR and promoting the collaboration role in enabling ENI 007), which defines various of autonomous network topics autonomous categories for the level of applica‑ among SDOs, operators, vendors networks. tion of AI techniques to the man‑ and other industry participants. agement of the network, going from basic limited aspects to the In GSMA, AI and Automation is full use of AI techniques for per‑ one of the topics of the “Future forming network management. Network”. In June 2019, the first GSMA Global AI Challenge was ENI is developing a general-pur‑ held, investigating three specific pose architecture for enhanced areas: connectivity in rural areas, network intelligence, and a map‑ mobile energy efficiency and TC INT AFI (Technical ping on NFV policy management enhanced services in urban areas. Committee Core (TC) is under discussion as NFV started Network and Interoperability policy modelling work for auto‑ At its workshop in June, at the AI Testing (INT) working group mating NFV management and in Network Seminar of the Mobile Autonomic Management and VNF CI/CD. World Congress Shanghai 2019, Control Intelligence for Self- GSMA called on the entire indus‑ Managed Fixed and Mobile try to focus on and contribute to Integrated Networks). Activities in CCSA the key applications of AI in the mobile networks, and jointly build ZSM (Zero-touch network & As one of the most influen‑ the 5G era for intelligent autono‑ Service Management). tial SDOs in the field of com‑ mous networks. In October 2019, munication in China, China GSMA published AI in Network TC INT AFI is studying Generic Communications Standards Use Cases in China”. Autonomic Networking Association (CCSA) began stand‑ Architecture (GANA), and ZSM is ards work on autonomous net‑ TM Forum has held several work‑ discussing closed-loop automa‑ works from 2010, and the items shops on autonomous networks tion in the ZSM framework, opti‑ are mainly set up in Technical since 2019, and the Autonomous mized for data-driven machine Committee TC 1, TC5 and TC7, Networks Project (ANP) was learning and AI algorithms. including use cases, architecture, established in August 2019. data handling, levels of auton‑ omous network, management requirements, etc. ITU AI/ML in 5G Challenge ITU News MAGAZINE No. 05, 2020 29

Three whitepapers have been Activities in open The importance published: Autonomous source communities of standards for Networks Whitepaper 1.0, autonomous networks IG1193 Vision and Roadmap v1.0, The End User Advisory Group and IG1218 Business require‑ (EUAG) of Linux Foundation As technologies and networks ment and architecture v1.0. In Networking (LFN) is drafting a evolve, autonomous networks will 2020, Autonomous Networks survey to solicit CSPs’ deployment be an important enabler in future. Whitepaper 2.0, business require‑ status, requirements and strate‑ In order to implement network ments and architecture v1.1, gies of network autonomy, with autonomy, a staged path is envi‑ technical architecture, demo of the hope to abstract a converged sioned, where the key differentia‑ Catalyst projects, user stories/use vision and serve as input to corre‑ tor to ensure convergence in later cases, etc. are underway. sponding technical groups. stages is to build a common archi‑ tecture across layers, standardize The Global TD-LTE Initiative has Since its earliest releases, the cross-layer interfaces and drive established a project dedicated Open Network Automation SDO alignment in earlier stages. to intelligent networks. Under Platform (ONAP) has been build‑ the 5G eMBB program, it will ing and enhancing use cases for SDOs should continue their specify use cases and require‑ radio, core and transport auto‑ relevant standardization work and ments for intelligent networks in mation on top of its policy-driven play a leading role in enabling the fields of levels, architecture, close-loop framework. Intent- autonomous networks. network elements and network based edge-to-edge (E2E) slicing management. automation is on its way. One ETSI official proof-of-concept shows that ONAP would provide the basis for constructing a reference stack for autonomous networks. ITU AI/Machine Learning in 5G Challenge webinars

19/06/2020 26/06/2020 03/07/2020 Graph Neural Networking Beam Selection — Machine Channel Estimation — Machine Challenge 2020 Learning Applied to the Physical Learning Applied to the Physical José Suárez-Varela Layer of Millimeter — Wave MIMO Layer of Millimeter — Wave MIMO Researcher, Barcelona Neural Networking Systems Systems Center, Universitat Politècnica de Catalunya Aldebaro Klautau Nuria González Prelcic (BNN-UPC), Spain Professor, Federal University of Pará (UFPA), Associate Professor, North Carolina State Brazil University, United States

10/07/2020 17/07/2020 22/07/2020 ITU AI/ML in 5G Challenge: ITU AI/ML in 5G Challenge: Radio Link Failure Prediction Improving the Capacity of IEEE DNN Inference Optimization Challenge 802.11 WLANs through Machine Challenge Salih Ergüt Learning Liya Yua 5G R&D Senior Expert, Turkcell Francesc Wilhelmi Open Source & Standardization Engineer, ZTE Researcher, Universitat Pompeu Fabra, Spain

24/07/2020 27/07/2020 29/07/2020 5G + AI + Immersive + Assistive AI Techniques for Privacy- Machine Learning for Wireless Services in Telecommunications Preserving Remote Medical LANs + Japan Challenge Brejesh Lall Diagnosis + Spectrum and Introduction Professor, Indian Institute of Technology, Network Resource Sharing in 5G Akihiro Nakao Delhi Networks Professor, University of Tokyo Brejesh Lall Koji Yamamoto Professor, Indian Institute of Technology, Associate Professor, Kyoto University Delhi Tomohiro Otani Executive Director, KDDI Research, Inc. Takanori Iwai Research Manager, NEC Corporation 31/07/2020 07/08/2020 LYIT/ITU–T AI Challenge: ITU AI/ML in 5G Challenge: Demonstration of Machine Lecture on Machine Learning and Learning Function Orchestrator Participation in Japan Challenge 07/08/2020 (MLFO) Via Reference Akihiro Nakao An Overview of the Implementations Professor, University of Tokyo ITU‑ML5G‑PS‑012 “ML5G‑PHY Shagufta Henna Koji Yamamoto [Beam Selection]” Lecturer, Letterkenny Institute of Technology Associate Professor, Kyoto University Aldebaro Klautau (LYIT), Ireland Tomohiro Otani Federal University of Pará (UFPA), Brazil Executive Director, KDDI Research, Inc. Takanori Iwai Research Manager, NEC Corporation 17/08/2020 19/08/2020 21/08/2020 Applying Knowledge Graph and ITU/AI/ML in 5 Challenge Open A Universal Compression Digital Twin Technologies to House and Roundtable No. 2 Algorithm for Deep Neural Smart Optical Network Prerana Mukherjee Networks Anran Xu Assistant Professor, School of Engineering, Wojciech Samek Researcher, China Information and Jawaharial Nehru University, Delhi, India Head of the Machine Learning Group, Communication Technologies Group Fraunhofer Heinrich Hertz Institute, Germany Corporation (CICT)

26/08/2020 31/08/2020 01/09/2020 ITU AI/ML in 5G Challenge: Traffic Recognition and Long- Milvus: An Open Source Vector China Mobile Network Topology Term Traffic Forecasting Based on Similarity Search Engine Optimization Competition AI Algorithms and Metadata for Jun Gu Question Analysis 5G/IMT‑2020 and Beyond Partner, Zilliz Wang Xing Artem Volkov Researcher, China Mobile Research Institute Researcher Ammar Muthanna Associate Professor, St. Petersburg State University of Telecommunications, Russia

04/09/2020 16/11/20 How to Bring AI into 5G Radio 28/09/2020 Harnessing Deep Learning Access Network for Mobile Service Traffic Wireless 2.0: Towards a Smart Qi Sun Decomposition to Support Radio Environment Empowered Network Slicing Senior Researcher, China Mobile Research by Reconfigurable Intelligent Institute Metasurfaces and Artificial Alexis Duque Intelligence Research Associate, Net AI Marco Di Renzo Professor, CNRS & Paris-Saclay University, France

18/11/2020 02/12/2020 The Road Towards an AI‑Native Toward Effective Network Traffic Air Interface for 6G 27/11/2020 Analytics of Mobile Apps via Jakob Hoydis Leveraging AI & Machine Deep Learning Head, Research Department on Radio Learning to Optimize Today’s 5G Domenico Ciuonzo Systems and AI, Nokia Bell Labs Radio Access Network Systems Assistant Professor, DIETI, University of and to Build the Foundation of Naples, Federico II, Italy Tomorrow’s 6G Wireless Systems Tim O’Shea Co-Founder/CTO, DeepSig

04/12/2020 08/12/2020 Scaling CNN Inference for Towards Open, Programmable, Extreme Throughput and Virtualized 5G Networks Michaela Blott Michele Polese Distinguished Engineer, Xilinx Associate Research Scientist, Northeastern University, United States Insights from industry ITU News MAGAZINE No. 05, 2020 32 Shutterstock

Capability evaluation and AI accumulation in future networks

By Jun Liao, Artificial Intelligence Director, Tengfei Liu, Yameng Li, and Jiaxin Wei, Artificial Intelligence Engineers, China Unicom Research Institute

J The rapid development of information and communica‑ 5G networks has brought many tion technology industry and new challenges — networking a predominant trend in future Network intelligence becomes more complex, the network development. has become the focus services are more diverse, and the of the information connections are increasing on a China Unicom believes that future and communication large scale. network intellectualization will technology industry ensure safe and reliable network and a predominant Using traditional ways of network services and provide customers trend in future network operation and maintenance will with a fast and ultimate experi‑ development. make meeting new requirements ence. For network operation and for network development difficult. maintenance (O&M), it has the As an important problem-solving capabilities of self-configuration, Jun Liao, Tengfei Liu, method, network intelligence self-monitoring, self-healing, and Yameng Li, and Jiaxin Wei has become the focus of the self-optimization. Insights from industry ITU News MAGAZINE No. 05, 2020 33

can flexibly integrate the existing ML model evaluation: This is The test bed will services and resources to respond conducted to solve the problem evaluate the intelligence quickly — improving the overall of estimating performance of the level of networks in a efficiency and operation ability. ML model, especially for the ML quantitative manner model used in telecommunica‑ and will help realize the China Unicom is in the pro‑ tion networks, including: accu‑ rapid implementation of cess of building a test bed racy, security, robustness, etc. The network intelligences. for evaluating network intelli‑ applicability of different frame‑ gence capabilities. Based on works and algorithms in different the ITU Telecommunication scenarios will be evaluated. Jun Liao, Tengfei Liu, Standardization Sector’s (ITU–T’s) The testing of the operating Yameng Li, and Jiaxin Wei Y.3173 Recommendation: efficiency and flexibility of various Framework for evaluating intelli- deep learning frameworks (such gence levels of future networks as tensorflow, paddle paddle) including IMT‑2020, it will pro‑ is supported, and the results of vide professional evaluation comparative analysis are pro‑ methods and services which vided. The efficiency and accu‑ will include computing power, racy of different AI algorithms in By using advanced automation a machine learning (ML) model, the same application scenario and intelligent technology com‑ and data and network intelligence are compared. prehensively, we can reconstruct capability. the existing network architecture, Data handling: Network data O&M mode, enable services The test bed will evaluate the is collected and pre-processed innovation, and create extreme intelligence level of networks in a to form a network data set, which user experience. quantitative manner and will help can be used to train ML models. realize the rapid implementation The fields covered by the dataset of network intelligence. The test include network data and image Test bed for content includes the follow‑ data. Network data is mainly evaluating network ing items: text data, involving five aspects: intelligence capabilities wireless, core network, transmis‑ Computing power evaluation: sion network, bearer network, Evaluating and sharing network For different types of accelerated and access network. Image data intelligence capabilities are key AI chips, e.g. training chips or is mainly based on target detec‑ to ensuring the effectiveness of serving chips from different com‑ tion and semantic segmentation network intelligence. To achieve panies are evaluated in terms of and annotation. rapid innovation in the future, throughput capacity, delay, power we need to build a platform with consumption, etc. An objective capability accumulation and open and accurate report will be pro‑ data. When demands change, we vided by the test bed. Insights from industry ITU News MAGAZINE No. 05, 2020 34

About CubeAI evaluation dimension; 3) analys‑ Telecommunication operators ★ ing the evaluation object; 4) scor‑ integrate their own or third-party CubeAI is an open source AI ing the evaluation dimension; and ML marketplace into the future platform completely independently 5) obtaining the evaluation result. network and push the application developed by the China Unicom Research Institute. It currently At the same time, the test results of AI on the telco network and includes sub-platforms and of stability, ease of use, accuracy, promote the level of network functional modules such as AI online and throughput of each applica‑ intelligence. training, automated model release tion are given. and deployment, and visualization of AI capabilities. While designing the ML appli‑ cation, network operators need Its core role is to break through Machine learning interoperable mechanisms for the barriers between AI model marketplace integration identification of the ML mar‑ development and actual production applications, accelerate the process in future networks ketplace which may be used as of AI innovation and application, a source for ML models. Lack and promote the rapid iteration and Another important exploration of standard mechanisms to evolution of the entire life cycle of is ML marketplace integration exchange ML models and related AI applications from design and development to deployment and in future networks, including metadata between ML market‑ operation. IMT‑2020 (commonly known places and the network operator’s as 5G), which refers to network ML deployment environments ML capability accumulation. limits interoperability. Nowadays, ML models may be hosted in varied types of ML The ITU–T Y.3176 marketplaces e.g. The Linux Recommendation: Machine Foundation’s Acumos AI, learning marketplace integra‑ Network intelligence capabil- China Unicom’s CubeAI, AWS tion in future networks including ity evaluation: Network intelli‑ Marketplace, and Huawei’s IMT‑2020, which was edited by gence capability is evaluated by Network AI Engine. China Unicom, China Mobile and the test bed from various dimen‑ ZTE, provides the architecture sions: demand mapping, data col‑ Sometimes, the latest advances and reference points for integrat‑ lection, analysis, decision-making in predictive analysis or algo‑ ing ML marketplaces in future and action implementation. Each rithms have no dependency on networks. Also, the interaction application is evaluated accord‑ network architecture evolution process of model search, model ing to the ITU–T Y.3173 standard in ML underlay networks. Cloud- selection and push, model dis‑ and is divided into L0~L5 levels. hosted ML marketplaces may covery, model training and model The test process includes five attract developers of innovative deployment are provided. steps: 1) determining the eval‑ ML mechanisms and algo‑ uation object; 2) dividing the rithms to host their solutions. Insights from industry ITU News MAGAZINE No. 05, 2020 35 ZTE

Accelerating deep learning inference with Adlik open-source toolkit

By Liya Yuan, Open Source and Standardization Engineer, ZTE

J Machine learning, and deep learning in particular (ML/DL), have gained great popularity in many areas, including machine translation, Models may perform computer vision and natural language processing. well in training, but trained models With ML/DL frameworks such as Tensorflow, Pytorch and Caffe, face new challenges we can build and train ML/DL models to learn knowledge from data, in a production and ultimately use their inference ability to generate business value environment. in a production environment.

Models may perform well in training, but trained models face new Liya Yuan challenges in a production environment. Insights from industry ITU News MAGAZINE No. 05, 2020 36

In production environments, mod‑ Challenges in deploying models as containers els can be deployed on various ML/DL models or integrating models hardware platforms (e.g. central into applications on processing unit (CPU), graphics The result is that there are still embedded hardware. processing unit (GPU), field-pro‑ some challenges to be addressed grammable gate array (FPGA)) when deploying ML/DL models Models need to be with different performance in production environments, even optimized to meet different requirements for computing cost, when the models converge well performance requirements memory footprints and inference during training stages: in different scenarios. latency in different scenarios. There is a steep learning curve for users to decide the Adlik open-source project inference framework best suited to a specific hardware. These challenges are the focus of the open-source project Adlik Users require their own initiated by ZTE and now incu‑ solutions to deploy ML/ bated by the Linux Foundation AI DL models, e.g. deploying Foundation (LF AI).

Architecture of Adlik

Model Optimizer & Compiler: Boost computing efficiency, reduce power consumption and latency

Model Training Graph Optimizer Model Compiler

8bit

big model Structural small model Graph Layer Model Export fp32 Pruning Quantization Compression int8 Conversion Fusion

Image-based Image-based File-based Binary-file File-based Engine + Model Engine Model Engine Model

Adlik Inference Engine Adlik Inference Engine Adlik Inference Engine Service Portal Service Portal Binary static/dynamic AI-based File Programs Micro Micro load Services Services Inference Runtime Kubernetes Docker Kubernetes Docker Operation System & Device Drivers

x86 ARM GPU FPGA Mali GPU Cluster Storage Cluster Mgt Nodes GPU Storage Node Mgt Nodes Cloud Deployment Edge Deployment On-device Deployment

Adlik Engine: Support three kinds of deployment environment Insights from industry ITU News MAGAZINE No. 05, 2020 37

Adlik is an end-to-end optimiz‑ Tackling DNN inference ing framework for deep learn‑ optimization — ITU AI/ML There are many ing models with the goal of in 5G Challenge technologies to be accelerating the deep learning explored to optimize inference process in cloud, edge, That’s why we invited the com‑ ML/DL models for and embedded environments. petitors of the ITU AI/ML in 5G better execution. It is composed of a model optim‑ Challenge to tackle our prob‑ izer, model compiler and infer‑ lem statement on “Deep Neural ence engine. Network (DNN) inference opti‑ Liya Yuan mization” — our challenge to The model optimizer optimizes construct a general model-optimi‑ DL models trained with different zation algorithm that can help to frameworks for better inference achieve model acceleration. performance. The model com‑ piler then compiles the models There are many technologies to into a format that the inference be explored to optimize ML/DL System-communication optimi‑ engine supports. The inference models for better execution, in zation methods such as Deep engine then loads the compiled fields including model-targetted Neural Network partition can models to provide inference abil‑ optimization, system-commu‑ accelerate model inference ity on cloud, edge or embedded nication optimization and hard‑ by optimizing communication devices. ware-specific optimization. between different computing nodes or layers. And hard‑ The model optimizer plays a key The Adlik model optimizer cur‑ ware-specific optimization with role in Adlik, especially when ML/ rently supports model-targetted toolkits such as TensorRT can DL models need to be deployed optimization methods including optimize inference operations in environments with strict con‑ model pruning and quantiza‑ based on the characteristics of straints in computing cost, mem‑ tion, and our team has focused the involved hardware. ory footprints or inference latency its recent work on knowledge — environments such as 5G edge distillation. This model-targetted The second release of Adlik will computing scenarios. optimization focuses mainly on come out soon. We welcome the compression of ML/DL mod‑ you to try it and contribute to it Adlik is an open-source project els, with other promising methods on Github. and we are eager to incorporate including kernel sparseness and as many good solutions as we can low-rank decomposition. in the Adlik model optimizer. Insights from industry ITU News MAGAZINE No. 05, 2020 38 Turkcell

Challenges and opportunities for communication service providers in applying AI/ML

By Salih Ergüt, 5G R&D Senior Expert, Turkcell

J Artificial intelligence (AI) has The science fiction of yester‑ 5G deployments are slowly transformed many industries. day is indeed fast becoming picking up all around the With new technical developments today’s reality. world and many standard‑ in computer vision, where AI ization organizations have trains computers to interpret However, AI applications have not already started to work on 6G; and understand the visual world, lived up to their potential for com‑ the ITU Telecommunication natural language processing munication networks, considering Standardization Sector (ITU–T) (NLP), which involves the the vast amount of data available Focus Group on “Technologies interaction between data science and the difficulty in managing for Network 2030” (FG NET‑2030) and human language, time series infrastructure, that is increasing and NGMN’s 6G Task Force, to forecasting to predict the future, in complexity. name a few. etc., it is hard to keep up with the progress in this field. Insights from industry ITU News MAGAZINE No. 05, 2020 39

Like many CSPs, Turkcell is com‑ To ensure the availability and It is hard to imagine mitted to using AI responsibly quality of services, customer how we can deploy and ethically and announced experience centric network and operate future its commitment publicly as “AI management is becoming networks without Intelligence Principles”. increasingly important for CSPs. AI assistance. In the 5G-PERFECTA project we Service-based architecture is get‑ are developing radio key per‑ ting more traction and end-to-end formance indicator (KPI)- based Salih Ergüt network slicing is the key technol‑ models that predict the quality of ogy for CSPs to provide services experience (QoE) of TV streaming with different quality-of-service services. (QoS) requirements on their 5G infrastructure. In the RELIANCE Such models allow near real-time project of which Turkcell is a part‑ customer experience measure‑ ner, while we are assessing the ments for this service and results It is hard to imagine how we can benefits and overhead of slicing can be extended to model other deploy and operate future net‑ under several use cases — ranging internal and external services works without AI assistance. from video conferencing to smart through transfer learning. trains to smart buildings — we are also investigating the prediction Service providers of future performance metrics, The challenges of building — investigating which helps devise and optimize AI applications and experimenting resource allocation strategies. The challenges of building It is common for communication As opposed to focusing on peri‑ AI applications in an opera‑ service providers (CSP) to imple‑ odic QoS reports and statistical tor network in a multi-vendor ment non-network related use distributions, providing a QoS and multi-technology setup cases using AI. Churn prediction, forecast instead is becoming are outlined in an ITU–T Focus customer segmentation, commu‑ more valuable for mission-crit‑ Group on “Machine Learning for nity heatmaps, up-sale/cross-sale, ical services. For example, an Future Networks Including 5G” and fraud prediction are popular autonomous car promising a high (FG‑ML5G) contribution. examples. Many operators have reliability score of 99.9999…% is started deploying and experi‑ no longer sufficient, as it needs Implementing a real-time appli‑ menting with AI techniques in net‑ to be warned about a coverage cation is difficult on a centralized working, though it is not an easy problem in advance, to be able to database architecture as it takes task due to limited resources in take precautions. 3GPP TR 22.886 a lot of processing power and this field. Also, developing unbi‑ defines use cases on informa‑ introduces delays. ased and accountable models tion sharing for partial and fully is necessary. automated driving and platoon‑ ing scenarios. Insights from industry ITU News MAGAZINE No. 05, 2020 40

About the ITU Focus Group ML5G

The ITU Standardization Sector (ITU–T) Focus Group on Machine Learning for Future Networks including 5G was established by The value of an AI‑ready ITU–T Study Group 13 in 2017. The Focus Group drafted ten technical specifications for machine learning (ML) for future networks, including architectural design interfaces, network architectures, protocols, algorithms and data formats. FG ML5G was active from January 2018 until July 2020. An AI‑ready architectural design that supports interoperability, data handling mechanisms and Learn more here. tools to evaluate network maturity in this journey is very valuable for CSPs when transforming their net‑ work into an intelligent network.

The ITU–T Focus The ITU–T Focus Group on Group on ’Machine “Machine Learning for Future Learning for Future Networks including 5G” contrib‑ When implementing a real-time Networks including uted significantly to this cause special event handling module 5G’ contributed through its specifications during we had to use non-standard significantly. its lifetime. backdoors to network equipment, to access the real-time statistics, Examples include: “Architectural and to control the nodes through Salih Ergüt framework for machine learn‑ remote commands. Non-standard ing in future networks includ‑ interfaces mandate close assis‑ ing IMT‑2020” (ITU–T Y.3172), tance from the vendor to manipu‑ “Framework for data handling late the proprietary interfaces and to enable machine learning some of those actions may cause in future networks including security vulnerabilities. IMT‑2020” (ITU–T Y.3174) and even a lack of those KPIs which “Framework for evaluating intel‑ Since these interfaces differ from makes AI application devel‑ ligence level of future networks one vendor to another, AI applica‑ opment for networks difficult. including IMT‑2020: use cases” tions need to be customized and For example, timing advance (TA) (ITU–T Y.3173). sometimes redesigned for each is reported directly as a KPI in vendor. The problem with KPIs the equipment from one ven‑ is not limited to real-time access dor, while trace logs need to be hurdles, but also the vendor-spe‑ parsed to extract TA for another cific definition of certain KPIs or vendor’s equipment. Insights from industry ITU News MAGAZINE No. 05, 2020 41

Simpler models for mission- Providing the Federated learning — critical industries best prediction easing the burden of big data processing There are still some gaps Another shortcoming of most AI between AI research and exper‑ algorithms is providing the best Finally, federated learning, which imenting AI in a live operator prediction based on the previ‑ allows jointly training a model network. Deep learning in mobile ously observed data. When they from multiple datasets, is another and wireless networking: A encounter a new scenario, or the future research direction that will survey, “IEEE Communications underlying data statistics have ease the burden of collecting Surveys & Tutorials”, vol. 21, changed, rather than reporting a and processing huge amounts of no. 3, pp. 2224–2287, 2019, by lack of confidence in their deci‑ data from all nodes into a central C. Zhang, P. Patras and H. Hamed, sion, algorithms still provide their data warehouse. In addition to gives an excellent survey of deep best guess. In telecom networks, increasing loads on the system, learning techniques for mobile it is essential to adopt human- centralized approaches also suffer and wireless networks. in-the-loop AI mechanisms for from extended delays and need robust and resilient operation. to be handled with great care to For mission-critical and highly prevent any privacy concerns. regularized industries such as Transfer learning, which provides Federated learning also allows telecommunications and health learning with fewer examples by multiple parties to develop a bet‑ care, trust and validation of the utilizing knowledge from a pre‑ ter model without compromising autonomous decisions are of viously solved related problem, data privacy. utmost importance. Achieving is an active research area with high predictive accuracy is not positive implications for networks. To conclude, it is hard to imagine sufficient for such systems; the Since the AI agent would be able network operations without results should be obtained from to adapt to changes to its envi‑ AI‑assistance for 5G and beyond proper problem representation ronment, a model of a well-known networks. There are some chal‑ rather than flawed input data. service could be extended to a lenges but also big opportunities similar service whose inner-work‑ for CSPs to adopt AI into their net‑ Therefore, simpler models such ings were unknown. works. as linear regressions or decision trees are preferred over com‑ plex models in these industries due to their interpretability. However, recent research has shown that complex models can also be designed to pro‑ vide explainability. Insights from industry ITU News MAGAZINE No. 05, 2020 42 Rakuten Mobile Rakuten

Autonomous networks: Adapting to the unknown

By Paul Harvey, Research Lead, and Prakaiwan Vajrabhaya, Research Outreach and Promotion Lead, Innovation Studio, Rakuten Mobile

J Modern pedagogy has autonomous; then, is it perhaps simplify their infrastructures, pre‑ shifted away from traditional time we trained networks to be senting a mechanism to manage rote learning towards critical autonomous too? their network. The issue here is thinking. This pendulum swing the control of this mechanism, of mindset creates people who which is still mainly performed can autonomously problem- Virtualization by humans or well-defined solve when a yet-to-be-seen automated processes. Future situation is encountered, which Before we dive into autonomy, it’s networks, however, are full of is essential for the workplace of important to ground ourselves challenges, many of which are tomorrow. Like these workplaces, in its enabler: virtualization. never-before-seen. Therefore, for telecommunication networks Virtualization is the process by future networks to function prop‑ are difficult, ever-changing which hardware in the network is erly, we must embrace an auton‑ environments, consisting of abstracted by software, decou‑ omous mindset, in the same way new technologies and services, pling an application from the that teachers now train students as well as, complex traffic hardware upon which it operates. to think critically and solve novel patterns. This begs the question: This new abstraction enables problems on the spot. if we are training talents to be telco operators to unify and Insights from industry ITU News MAGAZINE No. 05, 2020 43

applications, or even performing Exploring new For future networks intrusion detection. technologies autonomously to function properly, we must embrace an Indeed, various machine learn‑ Imagine giving a young child a autonomous mindset. ing technologies have been marker and then walking away. shown as effective ways to It’s likely that by the time you automatically address many return they will have used it on Paul Harvey, use cases within the network, every surface they could see: Prakaiwan Vajrabhaya such as those identified by paper, face, wall, or even the the ITU Telecommunication dog. Looking at ourselves, we, as Standardization Sector (ITU–T) humans, instinctively experiment Focus Group on Machine with the world around us, receiv‑ Learning for Future Networks ing feedback (praise or scolding, including 5G (FG‑ML5G), ITU–T for instance) that guides but does Y.3170‑series — Machine learn‑ not dictate future actions. Automation is ing in future networks including not autonomy IMT‑2020: Use cases. The same is true with future technologies. Autonomous net‑ First and foremost: automation is Although automation brings great works must learn to incorporate not autonomy. benefits, it also presents an inher‑ new technologies without being ent challenge due to its pre-de‑ explicitly told their purpose and/ Automation is operating fined nature. Consider when or learn to apply existing tech‑ within well-defined there is an unforeseen change, for nologies to address changing parameters or with pre- example, in the problem domain, problem domains. In this way, we defined constraints. the technology domain, or a new learn how to use new tools and application class, human engi‑ their effects upon the world. Autonomy is the neers must intervene and make independence to reflect modifications. about and adapt behaviour How to write — to go beyond well- Autonomy, on the other hand, down creativity defined parameters or pre- requires little or no human defined constraints. intervention as it is intended — How does trial and error exper‑ by design — to be self-adaptive imentation actually work? Given Automation is a powerful tool. and go beyond its well-de‑ the large number of potential It is targeted at a particular fined parameters or pre-de‑ technology combinations and problem or set of problems; for fined constraints. configurations, this requires an example: applying deep learning efficient mechanism to search the techniques to find anomalies, space to find potentially useful identifying specific web traffic or selections. Again, we look at our‑ selves for inspiration. Insights from industry ITU News MAGAZINE No. 05, 2020 44

Evolution depends on the Knowledge is power Is that everything? semi-arbitrary recombination and modification of small building Despite promising results, not This is, without a doubt, not a sim‑ blocks (genomes) and the appli‑ everything will (or should) be a ple task. Just as decisions without cation of reward “through survival semi-arbitrary exploration of pos‑ actions don’t solve problems, the of the fittest” or increased mating sible technology combinations. ability to autonomously adapt to chances for stronger offspring. Humans accumulate knowledge the unknown requires many differ‑ Evolution-based approaches have diachronically and apply this ent considerations including the been shown as effective mech‑ gained knowledge to shape creation of small building blocks, anisms to search large spaces the decision-making process. the construction of different use and find (optimal) solutions to Autonomous networks should be case ontologies and taxonomies, problems. Antenna design in able to benefit from this collec‑ the specification of languages to NASA (Hornby, G., Globus, A., tive knowledge, as well as gain describe these elements, or the Linden, D. and Lohn, J., 2006. their own. construction of simulation and Automated antenna design with canary testing environments. And evolutionary algorithms. In Space For this, we incorporate ontology the list goes on and on. 2006 (p. 7242)), and autonomous and taxonomy to represent the “rediscovery” of machine learning relationships that exist between The challenge of achieving an technologies at Google (Real, E., the relevant entities found within autonomous network is one that Liang, C., So, D.R. and Le, Q.V., the network. This representation can only be met by a concerted 2020. AutoML‑Zero: Evolving of human knowledge, combined effort. Not only to address the Machine Learning Algorithms with collected information from above, but also to ensure that From Scratch. arXiv preprint the network and ongoing knowl‑ autonomy can be realized as an arXiv:2003.03384), are some edge gained through trial and interoperable platform between examples of the concept having error experimentation, ena‑ operators, for the practical benefit been implemented. bles a codifiable way to guide of all. This approach is already autonomy. This decreases the being pursued in standardiza‑ Autonomous networks will be likelihood of certain evolution‑ tion organizations for the new no different, where biological ary combinations being chosen, wave of edge computing plat‑ genomes are replaced by the shrinks the space to be searched, forms arriving with the 5G era modular software blocks that are and makes the entire process of telecommunications. now standard practice in all soft‑ faster. Existing efforts, such as TM ware disciplines. As such, evolu‑ Forum’s Telecom Application Map In this regard, ITU and its mem‑ tion is a codifiable mechanism to (TAM), serves as a starting point bers have the opportunity to drive the creativity that is required for this process. take a leading role in bringing in addressing the unknown chal‑ together the necessary commu‑ lenges of future networks. nity and bridging the gap. Insights from industry ITU News MAGAZINE No. 05, 2020 45

In this way, Autonomy: Freedom from network. We are replacing hand- autonomous networks the mundane crafted automatic approaches present the opportunity with evolutionary-driven, exper‑ to free telco engineers Just as our children are being imentally verified, machine- from the mundane, prepared for the unknown crafted approaches. In this way, to focus on the workplace of the future, we too autonomous networks present extraordinary. are preparing our networks the opportunity to free telco to autonomously adapt to the engineers from the mundane — to unknown challenges of the future focus on the extraordinary. Paul Harvey, Prakaiwan Vajrabhaya

High-level architecture for evolutionary-driven autonomous network adaptation

Autonomy Engine Knowledge Center Adaptation Factory

Controller Module Knowledge Center Storing Controllers and Knowledge Evolution Experimentation Knowledge of Different Types Engine Engine

Autonomous Control Plane Controller

Controller Controller

Controller Controller Controller Controller Adaptation Factory

Controller Creating and Validating Controller Controller Controller Controller Controller Controller New Controllers

VNF Services VNF Services VNF Services Transport Transport Autonomous Control Plane vRAN Virtualization Network for Virtualization Network for Virtualization Infrastructure Fronthaul Infrastructure Backhaul Infrastructure Network Operation and III III III Internet Management On Premise Far edge Near edge Private Public Cloud Cloud vRAN = Virtualized radio access network VNF = Virtual network function

Source: ©2020 Mobile Rakuten Innovation Studio. Insights from industry ITU News MAGAZINE No. 05, 2020 46 Rohde & Schwarz Rohde

Quality of Experience testing in mobile networks

By Arnd Sibila, Technology Marketing Manager, Mobile Network Testing, Rohde & Schwarz

J Mobile network operators need to test the stability and performance of their networks in order to ensure good service. Due to the enormous amounts of data involved, this is hardly possible with manual methods. This is where artificial intelligence (AI) comes to the rescue. Mobile network operators need to test the stability and Network testing in the 5G era performance of their networks in order to With the advent of the fifth generation of mobile communications, ensure good service. network testers are confronted with a novel situation. Many aspects of 5G — diverse frequency bands, network operators’ different rollout pro‑ Arnd Sibila grammes, the breadth of applications such as the Internet of things (IoT), conventional mobile communications, traffic networking, etc. — lead to highly differentiated networks and test data. Insights from industry ITU News MAGAZINE No. 05, 2020 47

Analysing this data in the usual is a model that contains the Smart — a new aggregated form quickly leads knowledge of all the training generation to distorted results and incorrect data needed to accomplish our platform for mobile interpretations. AI offers a good specific objective function. network testing solution to this dilemma. While algorithm-based methods reflect At the inference stage, we take Watch this video to get a glimpse of the new paradigm specific theories, AI methods, the learned model and apply it to in mobile network testing such as pattern recognition, are new data to generate a prediction that allows to reduce able to evaluate data sets without or insight otherwise hidden in the complexity for Quality of preconceptions, and discover structure of the data. This stage Experience (QoE) centric and targeted network quality and relationships that would remain requires less computing power performance improvements. hidden to human analysts. than the training stage, thus mak‑ ing it suitable for standard server central processing units (CPUs) Big data needs or even edge computing, which machine learning avoids sending sensitive data to a server. The term “machine learning” is a bit more specific than the After the intensive training phase, term “artificial intelligence”. With the model is able to correctly machine learning (ML) the goal interpret new measurement data is to automatically derive general almost spontaneously. rules from a large volume of data. In our ML applications on drive testing data, a typical deep learn‑ Opting for machine- ing pipeline covers two stages: learning methods training and inference. We use ML methods for appli‑ At the training stage, we collect cations such as simplifying the promising for testing mobile the training data from numerous optimization of mobile networks networks where particularly large tests performed under different or improving the assessment of amounts of data are generated, configurations and environ‑ qualitative differences between so that manual analysis and ments. This data goes through providers. The Data Intelligence rule formulation are no longer a computing-intensive training Lab established in 2018 tackles practical. ML makes it possible process, normally using graphics these issues and supports Rohde to use the information hidden in processing units (GPUs) owing to & Schwarz research and devel‑ large data sets, for example to their ability to perform numerous opment (R&D) departments with derive new assessment metrics. simple operations in parallel. data-based analysis methods. An example is the call stability The output of the training stage These approaches are especially score (CSS). Insights from industry ITU News MAGAZINE No. 05, 2020 48

Rohde & Schwarz and ITU standards

Rohde & Schwarz is an ITU Sector Member active in the work of the ITU Standardization Sector (ITU–T) Study Group 12 (SG12: Performance, Quality of Service and Quality of Experience) for driving AI and machine learning. A suddenly dropped phone call is an AI and machine learning are today widely used in developing models annoying experience. to assess the quality of speech, audio and video, for example in ITU standards for the quality assessment of audiovisual streaming, in particular ITU P.1203 (progressive-download and adaptive-bitrate AV) and ITU P.1204 (video streaming services up to 4K). Arnd Sibila

New ITU quality-assessment standards address intelligent network analytics and diagnostics (ITU E.475) and the creation and validation of machine learning based models to assess media quality (ITU P.565).

Based on the experience of developing these ITU standards, Study Group 12 is giving further guidelines on how to apply AI and machine learning in ITU standardization in an upcoming ITU Technical Report and Supplement. duration and classified based on quality. The diagnostic also Rohde & Schwarz expects a significant increase in studies and includes unstable calls that were recommendations using AI and machine learning techniques and successfully completed but the supports these activities in ITU very actively. data proves they were not far away from being dropped. In conventional CDR statistics, those unstable calls would be assessed positively as successful calls, distorting the network quality assessment. Call stability score: A new Consequently, drive test cam‑ assessment metric for paigns are long and expensive. The CSS value is based on infor‑ reliable communications mation gathered from millions Therefore, we use a method to of test calls and incorporated A suddenly dropped phone call replace the binary call status in the model during the learn‑ is an annoying experience. That (either successfully completed or ing process. The assessment is is why mobile network operators dropped) with a finely graduated conclusive right from the first have been testing voice quality analog value. This is done by call. The network call quality is and connection stability for many creating a statistical AI‑generated registered more accurately and years. A popular statistic is the model that links the transmission with less test effort. The model call drop rate (CDR). But since conditions with the call status. assesses the data based on the the number of dropped calls learned rules, and outputs a is very low in mature networks, The CSS derived from the number between 0 and 1. The it is necessary to make a large model allows the reliability of higher the number, the lower the number of calls in order to obtain the mobile connection to be likelihood of a drop occurring in a statistically significant value. measured over the entire call the observed interval. Insights from industry ITU News MAGAZINE No. 05, 2020 49

The CSS measurement is part of do not conform to the regular dis‑ The visualization of the feature the SmartAnalytics analysis plat‑ tribution of the dataset to which in SmartAnalytics from Rohde & form from Rohde & Schwarz (see they belong. Finding anomalous Schwarz enables users to quickly Screenshot 1). samples, also known as distribu‑ see which phases of a test deviate tion outliers, provides valuable from the model. The overall effect insight that often correlates with of time-based anomaly detection Time-based defects or errors in the data is a more efficient methodology anomaly detection collection process (e.g. faulty or for drive tests and optimization. misconfigured equipment). Another AI‑driven function in this Anomaly detection, in particular software suite is anomaly detec‑ Detecting anomalies in data trans‑ time-based anomaly detection, tion using unsupervised learning fers based on variable-length time is applicable to various test cases. — the neural network is trained to series benefits mobile network learn information that is hidden in operators by detecting deviations the data without labels. and identifying problematic areas instantaneously that are other‑ Anomaly detection is a much-ap‑ wise masked by key performance preciated tool by data scientists. indicator (KPI) averages. It aims to find data samples that

Screenshot 1: Call stability score visualization in Screenshot 2: Time-based anomaly detection visualization in SmartAnalytics from Rohde & Schwarz SmartAnalytics from Rohde & Schwarz Insights from industry ITU News MAGAZINE No. 05, 2020 50 Shutterstock

A network operator’s view of the role of AI in future radio access networks

By Chih‑Lin I, Chief Scientist, and Qi Sun, Senior Researcher, Wireless Technologies, China Mobile Research Institute

J 5G is under commercialization all around the world and the speed of its development is amazing. China, for example, is expected to have We need to rethink more than 600 000 5G base stations at the end of this year. the traditional telecommunications 5G, however, is an extremely complex system, which cannot be accom‑ paradigm and embrace plished by simply building hundreds of thousands of base stations. new technologies. There are still many challenges on how to achieve commercial success. For example, 5G power consumption, high costs, and 5G’s high flex‑ ibility lead to difficulties in operation and maintenance optimization. Chih-Lin I, Qi Sun How to quickly and effectively serve the diverse vertical industrial appli‑ cation demands is another challenge. Insights from industry ITU News MAGAZINE No. 05, 2020 51

To tackle these issues, we need to These predictions will Four use case studies from rethink the traditional telecommu‑ undoubtedly empower the operator’s perspective nications paradigm and embrace proactive network new technologies, such as cloud, management and control, The following four typical use big data analytics and machine leading to significantly cases are detailed below, from learning (ML). Deep integration improved network resource the operator’s perspective and of information technologies, data allocation and energy based on the commercialized and technologies and communication efficiency while ensuring a test-trial experience. technologies are envisioned for customized user experience. 5G and beyond. Advanced network Energy saving Intelligence is expected to be optimization and decision- introduced into wireless net‑ making: Driven by the To cope with the energy chal‑ works for all the domains and at collected data from the real lenges caused by mobile network all levels, from local, to edge, to network, data analytics and expansion, the multiple radio the cloud. Data analytics, ML and ML can help to efficiently access technology (RAT) coopera‑ artificial intelligence (AI) have solve massive problems in tion energy-saving system (MCES) been identified as key drivers of the 5G network. These are was developed by China Mobile the intelligence evolution and usually hard to model or to improve the energy efficiency revolution in wireless networks. entail great computational of mobile networks. MCES inter‑ complexity due to extremely acts with the radio access network high dimensions or non- in real-time, and can support Network capabilities deterministic polynomial- 2G/3G/4G RAN equipment from with data analytics and time (NP)‑hardness. multiple vendors. Specifically, the machine learning MCES system has three major Focusing on the radio access technical features: Data analytics and ML will network (RAN), the use cases can Network-level energy saving. empower networks with the fol‑ be roughly categorized into four Energy-saving cell discovery lowing capabilities: types: 1) Intelligent network man‑ function based on big data. agement and orchestration; 2) Cell turn-off/on at a Reliable prediction: The Intelligent mobile edge comput‑ lower time-granularity. vast multi-dimensional data ing; 3) Intelligent radio resource collected from the network management; and 4) Intelligent MCES has been deployed in enables network traffic as radio transmission technology. 18 provinces across China, well as the prediction of including 970 000 cells. In 2019, network anomalies, service Use cases may also further extend the total energy saved amounted pattern/type, user trajectory/ to optimize the radio frequency to over 40 million KWh. Now the position, service quality of area, i.e., AI-assisted digital MCES is also evolving to incorpo‑ experience, radio fingerprint, pre-distortion. rate a 5G system to enable coor‑ interference, etc. dinated energy saving in both 4G and 5G networks. Insights from industry ITU News MAGAZINE No. 05, 2020 52

Topology of multi-RAT cooperation energy-saving system (MCES)

XDR = Extended Data Record ng-eNB = Next generation eNodeB gNB = gNodeB BTS = Base Transceiver Station OMC = Operation and management center ng-eNB gNB eNodeB BTS OSS = Operation support system GSM-BSC = Global System for Mobile Terminal Communications-Base Station Control

GSM-BSC

XDR Big Data Platform Multi-RAT Cooperation Energy Saving System 5G 4G 2G OMC/OSS OMC/OSS OMC/OSS

Automatic anomaly analysis

Anomaly detection (AD) and analysis has always been an Workflow of the automatic anomaly analysis important part of the operation and management (OAM) sys‑ AD Anomaly tem. By introducing a dynamic result cause KPI/KQI/ Anomaly Root cause Output/AD machine-learning-based AD MR data detection analysis report algorithm, the number of manual rules can be reduced and more Automatic Post Tickets labelling evaluation data accurate root cause information Online can be revealed. Meanwhile, KPI = Key performance indicator update KQI = Key quality indicator the AI‑based root cause analysis MR = Measurement report (RCA) algorithm directly replaces AD = Anomaly detection part of the human effort involved in the anomaly analysis process. Insights from industry ITU News MAGAZINE No. 05, 2020 53

Quality of Radio fingerprint-based Virtual grid of radio fingerprint experience optimization traffic steering

The 5G business model is chang‑ Traffic steering, also termed ing from “volume” to “value”. User mobile load-balancing, is a widely Serving Cell quality of experience (QoE) is used network solution to distrib‑ Intra-frequency Intra-frequency playing a key role in the commer‑ ute the traffic load among cells, Neighbour Cell1 Neighbour Cell2 cialization of 5G, thus network or to transfer traffic in order to optimization targets are shifting improve network performance. from key performance indicator This solution aims to enhance the Grid i (KPI) to key quality indicator (KQI) performance of traffic steering by Index ServingCell Id, RSRP/RSRP Segment ID related QoE. constructing a radio fingerprint IntraFreq_NeighbourCell_1 ID, RSRP/RSRP Segment ID which divides the cell into grids IntraFreq_NeighbourCell_2 ID, RSRP/RSRP Segment ID The Radio Intelligent Controller is by the serving cells and neigh‑ Attribute positioned as a data-driven plat‑ bouring cells radio signal levels, Coverage quality of inter-freq neighbour cells, e.g. RSRP, RSRQ, RSSI, CQI, MCS, beam ID. form to provide customized RAN in order to locate the user equip‑ capabilities and exposure, espe‑ ment’s (UE’s) grid and perceive RSRP = Reference signal receiving power RSRQ = Reference signal receiving quality cially for vertical industries and UE’s coverage information. This CQI = Channel quality indicator over-the-top (OTT) services. High- can largely reduce the number of MCS = Modulation and coding scheme definition (HD) video streaming, UE inter-frequency measurements cloud virtual reality (VR), and and speed up traffic steering. cloud gaming are expected to be among the most popular services China Mobile and partners also of the 5G era. carried out trial tests of radio Also, the measurement reconfig‑ fingerprint-based traffic steer‑ uration from base station and the In 2019, China Mobile conducted ing in the commercial network. measurement report signalling trials in the Shanghai 5G network Based on the optimization of overhead from UE are reduced and the following features have radio fingerprint-based traffic by 54 per cent and 83 per cent, been verified: 1) AI/ML‑based steering, the test results showed respectively. Moreover, with the QoE prediction and guarantee for that compared with traditional radio fingerprint-based load cloud VR; and 2) Radio bandwidth load balancing, the duration balancing, the average Internet estimation for cloud VR adaptive of high loading is reduced by protocol (IP) delay of the tested coding selection. 13 per cent. cells is reduced by 20 per cent. Insights from industry ITU News MAGAZINE No. 05, 2020 54

Standards progress Future work The industry is working Various studies on network Although great progress has actively towards the intelligence standardization, from been achieved, 5G specifica‑ open and smart radio management to control planes, tions barely offer mobile oper‑ access network to fully including use cases and require‑ ators any guidance on how to ensure the commercial ments, AI functional framework, make their 5G networks truly AI/ success of 5G. procedures and architecture have ML-enabled. Lots of further work been actively explored in stand‑ is envisioned on better support‑ ardization organizations includ‑ ing AI/ML‑enabled networks, Chih-Lin I, Qi Sun ing the ITU Telecommunication and to incorporate fundamen‑ Standardization Sector (ITU–T), tal concepts of AI and ML as 3GPP, O‑RAN and ETSI. the core fabric of the network, which includes: 3GPP has already introduced the data analysis and AI/ML related Customized fine granularity function of the core network data collection. and management plane with the RAN capability exposure for service-based core network and customized network and management architecture. While service optimization. for the radio access network, RAN programmability, service- considering the distributed archi‑ based architecture in RAN to tecture and more stringent timing enable AI/ML. and reliability characteristics, it Decoupling the faces more challenges when we communication and data bring in embedded AI. analytics/AI/ML to enable efficient innovation. The industry is working actively Wireless open data sets towards the open and smart radio to accelerate the pace of access network to fully ensure the wireless AI algorithm and commercial success of 5G. application innovation. Insights from industry ITU News MAGAZINE No. 05, 2020 55 Shutterstock

AI and open interfaces: Key enablers for campus networks

By Günther Bräutigam, Managing Director, Airpuls; Renato L.G. Cavalcante, Research Fellow, and Martin Kasparick, Research Associate, Fraunhofer HHI; Alexander Keller, Director of Research, NVIDIA; and Slawomir Stanczak, Head of Wireless Communications and Networks Department, Fraunhofer HHI, Germany

J Modern communication is the foundation of successful digitization. Thanks to the 5G standard, completely new applications are emerging Modern across industry verticals, and there is a strong demand for new wireless communication is technologies that address industry-specific problems in campus the foundation networks, also known as private networks. of successful digitization. To address this demand, for example, the Federal Government in Germany has laid the foundations for the development of campus networks by allowing their operation in the 3.7–3.8 GHz band, which is Günther Bräutigam, particularly suitable for covering large areas and supporting high mobil‑ Renato L.G. Cavalcante, ity scenarios. Martin Kasparick, Alexander Keller, and Slawomir Stanczak Insights from industry ITU News MAGAZINE No. 05, 2020 56

This frequency band is expected The technologies are based on This lower barrier to entry also to be complemented with the mil‑ disaggregation, virtualization, expands the market by creating limeter-wave band, which offers openness, and AI. In this new the conditions for niche technol‑ many advantages in terms of approach, the RAN may be disag‑ ogies from other markets to enter transmission bandwidth, security gregated into modules perform‑ campus networks. For example, against tapping, and robustness ing network functions in software. existing data mining algorithms against jamming, to name a few. Open interfaces between such may be customized for wireless virtualized RAN modules will be systems, such as the analysis Similar to conventional mobile crucial in enabling interoperability of RAN data streams to detect networks, campus networks have among vendors. attacks on radio interfaces. a core network containing the central elements for network con‑ While the interfaces are standard‑ If based on flexible multi-ven‑ trol, and a radio access network ized, standards do not address dor solutions, campus networks (RAN) managing the wireless the implementation of the net‑ would meet the needs of complex connections from base stations to work functions, enabling ven‑ industry-specific environments. mobile terminals. dors to differentiate to their own advantage and the advantage of Many vendors offer software for their customers. Moving beyond 5G the core network. Only a few networks with AI vendors have been supplying This high degree of softwarization RAN technologies. The vertically of RAN functions, in combina‑ As we move beyond 5G networks, integrated solutions hardly allow tion with the use of commercial a holistic approach to manage for interoperability. Due to the off-the-shelf (COTS) software and this complexity requires AI as an high barrier to entry, industry-spe‑ hardware leads to positive econo‑ integral part of the overall sys‑ cific technologies targeting niche mies of scale, and reduced costs. tem design. markets are not being developed. This lack of competition impedes The above evolution from When integrating new hardware innovation in RAN technology. closed to open and program‑ and software components into a mable systems has two impor‑ network, the integration can be tant consequences. tested in a “digital twin” environ‑ Fostering innovation in ment where simulation is acceler‑ campus networks It reduces the investment ated by AI. required to develop technolo‑ To foster innovation in campus gies for radio access networks, Novel AI tools are needed in networks, a new business model enabling small and medium-sized support of network functions such has been advocated. It is com‑ companies to develop indus‑ as scheduling, beam manage‑ prised of system design, optimiza‑ try-specific solutions for cam‑ ment, interference coordination, tion, and the integration of open pus networks. localization, symbol detection, and secure wireless technologies. and channel estimation, to name just a few prime examples. Insights from industry ITU News MAGAZINE No. 05, 2020 57

These tools may be fundamen‑ These issues may be addressed tally different from existing tools by hybrid model and data-driven To meet the strict applied in fields such as speech methods, where the data is used requirements of recognition and computer vision. to mitigate the uncertainty of the mobile applications models and the coarse models in campus networks, However, the success of open help to reduce the amount of the industry needs to and programmable systems will training data required by the devote a great effort mainly be decided on the lower learning tools. Crucially, some of to the development layers of the communication these operations must be per‑ and application of stack, where the environment is formed within timeframes ranging novel AI methods. highly dynamic and uncertain. from microseconds to millisec‑ onds, making it increasingly important to develop scalable Günther Bräutigam, The challenges algorithms that can be highly Renato L.G. Cavalcante, parallelized on commercial off- Martin Kasparick, Alexander Keller, and Acquiring datasets for training the-shelf (COTS) hardware. Slawomir Stanczak data-driven machine learning algorithms becomes a challenge as a result. By the time sufficient Call to AI action data has been collected, the environment may have changed To meet the strict requirements significantly enough to render the of mobile applications in campus training data out of date. networks, the industry needs to devote a great effort to the devel‑ Pure model-based methods are opment and application of novel now widely applied but may also AI methods. face serious challenges. As cur‑ rent wireless models often ignore This will create opportunities for beam-squinting effects and rely countries to develop their digital on further approximations such sovereignty and for companies to as waves in the far field region, provide solutions for niche mar‑ they may become too crude with kets that are waiting to be served. increasing operating frequency, rates, and number of antennas. Quotes from hosts of the ITU AI/ML in 5G Challenge problem statements

The ITU AI/ML in First of use cases in 2030? 5G Challenge has provided a great opportunity to bring together Andrey Koucheryavy standardization activities with academia Chair and Professor of by allowing students and professionals to solve Telecommunication Networks and Data important problems in communications. In particular, Transmission Dept., SPbSUT, Chief the challenge has led to obtaining very interesting Researcher NIIR, Russia, and results that may open the door to revolutionize Chairman of ITU–T SG11 the way communications are understood. For the Dynamic Channel Bonding problem, the application of Deep Learning models accounts for a significant breakthrough in the field.

Francesc Wilhelmi PhD Student, Wireless Networking To ensure network coverage density even in sparsely Research Group, UPF, Spain populated areas while guaranteeing high security and satisfying quality of service, an integration of intelligence edge computing and blockchain technology will play an important role in future 2030 networks.

Ammar Muthanna Deputy Head in Science, Telecommunication Networks and Data Transmission Dept., SPbSUT, and Head, SDN Laboratory, Russia

We are grateful to the “ITU AI/ML in 5G Challenge” for giving us the opportunity to connect with AI researchers around the globe in solving an operational challenge for Turkcell. Our challenge topic, “Radio link failures”, has critical implications for an operator network. Modelling link failures based on weather forecasts is a daunting task due to the unreliable nature of forecasts and rare occurrences of radio failures. Challenge preparation and communication with contestants have already helped us better formulate the problem, resolve inconsistences in the data, and consider The main important task for AI in alternative approaches. We are expecting to receive high- 5G is traffic recognition “on the fly” quality solution proposals for this challenging task and we without making delays for new services traffic are looking forward to applying the learnings from this management such as the Tactile Internet, medical challenge to implement mechanisms that take precautions networks, and autonomous vehicles. before radio link failures occur, to improve customer experience. Artem Volkov Researcher and PhD Student, Telecommunication Networks Salih Ergut and Data Transmission Dept., SPbSUT, Russia 5G R&D Senior Expert, Turkcell, Turkey

Hosting a competition within the AI/ML in 5G Challenge has been We are very honoured to be a part of the ITU AI/ a completely worthwhile experience. ITU ML in 5G Challenge — a great opportunity to organizers have made it very easy to prepare and accelerate the pace of applying AI/ML algorithms run the whole competition, and have done a great job in telecommunication networks. We proposed in attracting participants. In our problem statement, a problem statement “DNN model inference the participation numbers were far beyond what we could optimization” in the Enabler Track. The topic is expect at the beginning. As members of academia, we see this quite important when considering deploying ML Challenge as a pathway to disseminate the research we do. models in the network, especially where there are This kind of competition is very interesting for us to connect strict requirements for inference performance. with a different public from both the academia and industry. We have received some great submissions so far, Moreover, we were impressed by the proposed solutions, which we believe will definitely enlighten our some of them pushing the state of the art. We are looking work in the Adlik open source project. forward to the next edition in 2021. Liya Yuan José Suárez-Varela Open Source and Standardization Engineer, ZTE, China PostDoctoral Researcher, BNN‑UPC, Spain

The ITU AI/ML in 5G Challenge The Challenge provides a stage to showcase the helps to depart from the “small- potential of applying machine learning to enable data regime” and properly assess AI/ML network intelligence. It attracts more researchers algorithms such as deep learning, with large and engineers to put effort into the research and datasets and reproducible experiments. development of the intelligent network and has gathered the operators, suppliers, researchers, college students Aldebaro Klautau and related communities together to build a more flexible, efficient, green, and resilient network. By integrating AI into Professor of Electrical and Computer Engineering, UFPA, Brazil the network services, the network is being evolved to support better social, commercial and technology services. The Artificial Intelligence Industry Alliance will continuously support ITU in organizing challenges to drive the convergence of AI technologies and network infrastructure and services.

Qiang Cheng Artificial Intelligence Industry Alliance (AIIA), China

We are very honored to be part of this very exciting ITU AI/ML in 5G Challenge and to provide a problem set. AI/ML adoption is still at the very early stage in the telecommunication industry and innovation will definitely be brought by very unique ITU communities consisting of different countries, organizations, and backgrounds. Let’s keep innovative and collaborative in the 5G era and beyond.

Tomohiro Otani Executive Director, KDDI Research, Japan

Running the ITU challenge on MIMO channel estimation at millimeter wave has been a great way to kick off my new research program at North Carolina State University. It has made me realize that putting together the AI and wireless communities, so they can speak the same language, is a huge challenge. I have enjoyed exposing the participants to one of The Ministry of Internal Affairs and my favourite problems in millimeter wave MIMO Communications in Japan recognizes communications. that research and development on network autonomy using AI/ML is extremely important Nuria Gonzalez Prelcic for the further development of 5G and the early Associate Professor, NCSU, United States realization of B5G (Beyond 5G). By using AI/ML and competing to solve various 5G network problems, engineers will greatly contribute to the development of information and communications, such as the improvement of AI/ML technology and human resource development. The global ITU AI/ML in 5G Challenge is a very meaningful activity.

Ministry of Internal Affairs and Communications (MIC), Japan

We posit that one of the promising research and development directions for Beyond 5G/6G telecommunications is enabling “autonomic intelligence” in networking, as outlined by the “Beyond 5G Strategic Board” held by the Japanese Government’s Ministry of Internal Affairs and Communications. We observe that the ITU AI/ML in 5G Challenge is in perfect alignment with the strategy and has been quickly and globally encouraging young researchers to participate in competitions. It is an indication that ITU–T’s activity is proving to be an effective means for putting forth the strategic direction of R&D for enabling super- intelligence in networking. In Japan, RISING (Cross-Field Research Association of Super- Intelligent Networking) has led the regional competition of ITU’s AI/ML in 5G Challenge, fully supported by MIC, 5GMF, TTC, and industry partners, KDDI and NEC. I really appreciate ITU’s global promotion of applying AI/ML in telecommunications and I am looking forward to a successful outcome of the activities.

Akihiro Nakao Professor, UTokyo, Japan Quotes from the ITU AI/ML in 5G Challenge participants

Participating in this challenge was a golden chance to test and improve our ML skills in the context of newly introduced technology such as 5G. During My participation in the ITU AI/ this competition, we have managed to overcome ML in 5G Challenge has allowed some challenges and meet new people from me to gain hands-on experience around the world. of relevant topics to build and design future wireless technologies. Specifically, it has been exciting to apply different ML methods to the Khalid Al-Bagami particularities of future spectrum access mechanisms Telecom Engineer, Ericsson, Saudi Arabia that allow even better performance of what we have today.

Paola Soto-Arenas PhD Researcher, University of Antwerp, Belgium

The challenge has given us the opportunity to rethink classical channel estimation I feel ecstatic to say that approaches from an ML my participation in the ITU AI/ perspective. I am excited ML in 5G Challenge has been a great to see the results of other learning experience. It’s been a good participants. opportunity to work on the exploratory problems in wireless communication and

trying to find solutions to them using AI/ At UC3M we Dolores Garcia ML approaches. trained an NN solution PhD Student, to forecast 802.11 WLANs IMDEA, Spain throughput. The organizers (UPF) Megha Gururaj Kulkarni provided both datasets, and support Student, PES University, India during the challenge. It has been a very fulfilling experience.

Jorge Martín Pérez PhD Student, UC3M, Spain I’m a health tech enthusiast and my UNDP R42 group’s work is on an AI in 5G mobile chatbot providing COVID-19 information. My work spans Africa and Europe, and I aim to reach more regions to influence global markets and society, young and old. ITU has been fantastic in bringing together such dynamic participants to take part in innovations for 2020 challenges.

Mahlet Shimellis UNDP, Ethiopia/United States Insights from academia ITU News MAGAZINE No. 05, 2020 62 Shutterstock

AI/machine learning for ultra-reliable low-latency communication

By Andrey Koucheryavy, Chaired Professor, Telecommunication Networks and Data Transmission Department, The Bonch-Bruevich Saint Petersburg State University of Telecommunications (SPbSUT), Chief Researcher, NIIR, and Chairman, ITU–T SG11; Ammar Muthanna, Deputy Head, Science, Telecommunication Networks and Data Transmission Department, SPbSUT, and Head, SDN Laboratory; Artem Volkov, Researcher and PhD Student, Telecommunication Networks and Data Transmission Department, SPbSUT, Russia

J 5G networks are designed to The International Mobile Massive machine-type com‑ integrate all the achievements of Telecommunications (IMT) Vision munications, the concept mobile and fixed communication in ITU Recommendation ITU–R of the Internet of Things networks to provide ultra-high M.2083‑0, describes three 5G (ITU Recommendation data speeds, enabling a range usage scenarios: 1) enhanced ITU–T Y.2060/Y.4000), refers to of new services with new cloud mobile broadband (eMBB); 2) trillions of things connected and computing structures such as fog massive machine-type com‑ uniquely identified, and requires a and edge. munications (mMTC); and fundamental shift from traditional 3) ultra-reliable and low latency ideas about the number and communications (URLLC). volume of device databases in the network. Insights from academia ITU News MAGAZINE No. 05, 2020 63

5G usage scenarios from the ITU–R IMT‑2020 Vision Recommendation

Enhanced mobile broadband

Gigabytes in a second

3D video, ultra-high definition (UHD) screens

Smart home/building Work and play in the cloud

Augmented reality

Voice Industry automation

Mission critical applications, e.g. e-health Future IMT Smart city Self-driving car

Massive machine type Ultra-reliable and low latency communications communications

The Internet of Things, as a driver Reducing latency in 5G This represents a movement for the development of network networks and beyond towards a greater number and technologies, has put forward a variety of network and computing whole layer of new services cover‑ We can reduce latency drastically technologies, resulting in more ing all spheres of life and society. in 5G networks and beyond, complex networks and more building on innovations in soft‑ complex network management. But ultra-reliable and low latency ware-defined networking (SDN), communication creates the most network function virtualization This calls for a review of the significant challenges for the sci‑ (NFV) and edge computing. established principles in net‑ entific and technical community. work management. And working towards the ambi‑ While it could give life to “Tactile tious goal of latencies lower than And this introduces artificial Internet” applications in areas 1 millisecond to enable tactile intelligence (AI) — new capabil‑ such as telemedicine, auton‑ applications, the scientific and ities well-positioned to support omous vehicles and industrial technical community is looking to the high programmability and robotics, it also generates fog and edge computing tech‑ automated provisioning possible stringent new quality of service nologies as key to architectural with SDN-enabled orchestra‑ (QoS) requirements. approaches in future networks. tion systems. Insights from academia ITU News MAGAZINE No. 05, 2020 64

A Tactile Internet application in 2020 may have been remote- controlled robots assisting patients in hospitals ill-prepared for outbreaks of COVID‑19. Shutterstock

Andrey Koucheryavy, Ammar Muthanna, Tactile Internet Artem Volkov Here are the key characteristics of Tactile Internet applications:

` Decentralized network architecture to enable Tactile Internet services at the network edge. (Can network decentralization reduce digital inequalities?)

` Real-time tactile interaction matching human sensations requires latency lower than 1 millisecond. AI is a class of mathematical ` The under 1 millisecond target creates stringent new machine-learning algorithms and requirements for system-network solutions. models and Big Data processing algorithms. The latest processor A Tactile Internet application in 2020 may have been remote- controlled robots assisting patients in hospitals ill-prepared for technology makes it possible outbreaks of COVID‑19. to implement AI algorithms quite effectively.

Volumes of traffic are growing, and the heterogeneity of traffic is also growing. The Internet of Things and ultra-reliable and Existing QoS tools cannot per‑ Operators also need full-fledged low latency communication form at the required level. And system forecasts of infrastruc‑ introduce a wide variety of new most solutions already need to ture development that take into requirements, and this will call for have load forecasts for certain account the speed at which new much greater efficiency in deci‑ services — forecasts that take technologies are introduced, to sion-making relevant to quality of into account geographical and enable new services over the service (QoS). dynamic factors such as sub‑ Internet, as well as the related scriber movements, including at changes expected in peo‑ high speeds. ple’s lifestyles. Insights from academia ITU News MAGAZINE No. 05, 2020 65

New AI approaches the ITU AI/ML in 5G Challenge to traffic identification analyses the metadata of SDN The main breakthrough with software network streams to identify and to achieve the full defined networking predict traffic. promise of tactile, ultra-reliable and low AI can perform two tasks This method of traffic identifica‑ latency communication of key importance to QoS tion and forecasting introduces applications, is assessment: unambiguous no delays in the transmitted traffic expected with the next identification of traffic, and subse‑ at the data plane level, ensures generation of physical- quent forecasting. the portability of analytical mod‑ layer technologies. ules across SDN controllers, and The AI task of identifying traffic can expand in types of recog‑ includes the need to recognize a nized traffic. Andrey Koucheryavy, large number of traffic types with‑ Ammar Muthanna, out introducing additional delays AI algorithms — in conjunction Artem Volkov (taking into account ultra-reliable with new technologies for build‑ and low latency communication ing networks and cloud services services), and the need to expand calculations — can offer valua‑ and adjust the AI algorithm to ble support to the movement different geographical locations towards ultra-reliable and low of the network and services. latency communications.

The software defined networking But the main breakthrough (SDN) capabilities of 5G networks to achieve the full promise of make new approaches to traffic tactile, ultra-reliable and low identification possible. latency communication applica‑ tions is expected with the next The machine learning model generation of physical-layer developed by St. Petersburg technologies — namely, quan‑ University (SPbSUT) as part of tum communications. Insights from academia ITU News MAGAZINE No. 05, 2020 66 Shutterstock

AI/ML integration for autonomous networking — a future direction for next‑generation telecommunications

By Akihiro Nakao, Professor, The University of Tokyo, Japan

J Recently, commercial 5G services have been globally deployed and utilized all over the world. At the same time, research and development (R&D) strategies aimed at beyond 5G, i.e. 6G, are already in progress. Emergent strategy proposals for beyond In Japan, 5G services started in the spring of 2020, and the discussion 5G/6G stipulate new key body organized by the Ministry of Internal Affairs and Communications performance indicators. (MIC) for defining the R&D directions towards 6G was initiated prior to that, in January 2020. Akihiro Nakao In June 2020, the MIC Beyond 5G Strategic Board outlined the strategic proposal for 6G R&D. Insights from academia ITU News MAGAZINE No. 05, 2020 67

The emergent strategic proposals NTT has recently organized for beyond 5G/6G stipulate new a global forum called IOWN We recognize the key performance indicators (KPIs) (Innovative Optical and Wireless need for research that should improve target values Network). Although the concept on autonomous of existing 5G wireless technolo‑ includes many insightful build‑ networking. gies, such as large bandwidth, low ing blocks such as all photonics, latency, and a large number of data centric, low power, and low connections, by orders of mag‑ latency networking, the primary Akihiro Nakao nitude, as well as new elements aim seems to build infrastruc‑ such as ultra-low power, security, ture that can be operated auto‑ autonomy and deployability. matically by smart algorithms beyond human intelligence These last two new elements are and experience. especially interesting ones, as the goal is to realize self-operating The first RISING symposium networks and to extend telecom‑ Advanced communication included the presentation of munications to include systems research in Japan 111 posters and panel discus‑ that have been considered sions to enlighten many young difficult to deploy, such as High- In Japan’s academia, the Institute researchers and students. It Altitude Platform Stations (HAPS) of Electronics, Information and attracted great interest from and underwater, etc. Communication Engineers (IEICE) researchers in all fields of wireless is playing a central role in putting and wired communication infra‑ In fact, autonomous networking forth advanced communica‑ structure technology. using machine learning (ML) tion research. and artificial intelligence (AI) has been actively discussed recently. We recognize the need for The ITU and AI and “Autonomy” is the design and research on autonomous net‑ machine learning for 5G implementation of a communi‑ working to automate operations cation infrastructure that has the and automatically detect and Many stakeholders from Japan, functions of autonomous opera‑ predict failures in information including the University of Tokyo, tions without human intervention. and communication technology as well as industry partners have It is “zero touch” networking that infrastructures. This prompted us participated in the ITU Focus involves the construction of opti‑ to organize RISING — a cross-field Group on Machine Learning for mal network infrastructure that study group on super-intelligent Future Networks including 5G goes beyond wired and wireless. networks — of which I am Chair. (FG‑ML5G). Insights from academia ITU News MAGAZINE No. 05, 2020 68

Key Features for Beyond 5G

Synchronization of Physical & Cyber Spaces Beyond 5G THz Waves Sensing Ultra Fast & Large Capacity Ultra Low Latency Ultra Numerous Connectivity (Ultra eMBB) (Ultra URLLC) (Ultra mMTC)

ƒ Network Access: ƒ Latency: 1/10 of 5G ƒ Simultaneous Connectivity:  10x Faster than 5G ƒ Advanced Synchronization with CPS 10x more than 5G ƒ Core Network Access: ƒ High Level of Synchronization with 100x Faster than now Complementary Network

All-Photonics Network Quantum Cryptography Further Upgrade of 5G Features Ultra low power consumption Ultra security and resiliency eMBB URLLC mMTC ƒ Power Consumption: ƒ Always Ensuring Cybersecurity 1/100 lower than now 5G Without reduction measures, IT-related power ƒ Instant Recovery from Disaster/Failure consumption would reach to 36x than now, Adding New Features (1.5x of total consumption than now). Contribute to Generate Sustainable and New Values Low Power Consumption HAPS Semiconductor

Autonomy Scalability Inclusive Interface

ƒ Autonomous coordination among devices without manual ƒ Seamless Connection with Satellites and HAPS (incl. space and intervention ocean) ƒ Construction of optimized network highly integrating wired and ƒ Transforming various interfaces such as terminals and windows wireless connection into base stations ƒ Ubiquitous connections through coordination between devices Complete Virtualization

Green Words are the examples of areas where Japan has advantages or is actively engaged in. eMBB = enhanced Mobile Broadband URLLC = Ultra-Reliable and Low Latency Communications mMTC = massive Machine Type Communication Insights from academia ITU News MAGAZINE No. 05, 2020 69

The Ministry of Internal Affairs RISING came up with another Many stakeholders and Communications, KDDI, problem set regarding wire‑ from Japan, including NEC, Hitachi, and NICT, are less communication. the University of all Japanese entities that are Tokyo, as well as conducting research and devel‑ We closed the contest in mid- industry partners opment of innovative AI net‑ October 2020, and we will select have participated in work integration and have also three winners per problem set, the ITU Focus Group proposed many contributions and hopefully send the research‑ on Machine Learning to ML5G. ers to present their results at for Future Networks the global conference taking including 5G. Since July 2020, the ITU AI/ML in place online at ITU from 15 to 5G Challenge has been taking 17 December, 2020. place — a problem-solving com‑ Akihiro Nakao petition related to the application In the light of a series of research of AI/ML in 5G networking. and development activities and a great deal of engagement by Countries including China, many stakeholders in the ITU Brazil, Turkey, Ireland, India and Challenge, we posit that the inte‑ the United States have hosted gration of AI/ML in networking is contests. In our region a contest is a promising direction for defining being held by a coalition of part‑ next generation telecommunica‑ ners such as TTC and NEC/KDDI, tions, such as Beyond 5G/6G. centered on the aforementioned RISING community. We indeed intend to contribute to the research and develop‑ In the regional ITU Challenge in ment, as well as standardization, Japan, KDDI and NEC presented in the integration of AI/ML into research problem sets, respec‑ telecommunications in the com‑ tively. About 20 teams of four ing decade. persons registered to take part in working out solutions to each of the problem sets. Insights from academia ITU News MAGAZINE No. 05, 2020 70 Shutterstock

Realistic simulations in Raymobtime to design the physical layer of AI‑based wireless systems

By Aldebaro Klautau, Professor, Federal University of Pará, Brazil; and Nuria González‑Prelcic, Associate Professor, North Carolina State University, United States

J In innovation towards 5G and beyond, there is a clear trend towards learning from experience.

Simulated network environments are helping us to investigate a range In innovation of open questions of fundamental importance to AI’s contribution to towards 5G and beyond, wireless communications. there is a clear trend towards learning Complex mathematical models have provided the basis for the design from experience. of the current wireless systems. Human intelligence was the main driver of this design. But systems are moving towards wireless networks based on AI, with machine learning algorithms playing a key role. Aldebaro Klautau and Nuria González-Prelcic To derive AI‑based solutions, large and diverse datasets are needed to achieve good generalization performance, especially when deep-learn‑ ing approaches are considered. Insights from academia ITU News MAGAZINE No. 05, 2020 71

Raymobtime datasets and the ITU AI/ML in 5G that some design decisions need Challenge to be made prior to building the system. Datasets and sensible benchmarks are essential in guiding research efforts. Raymobtime focuses on the physical (PHY) layer, but datasets In this context, realistic simula‑ for many other applications are now available via the timely ITU AI/ML in 5G Challenge. tions deliver essential value.

There are currently ten Raymobtime datasets available here, totalling They do not completely substitute 290K communication channels. prototypes and measurements, A subset of these datasets are supporting contestants in the ITU but they support rapid innovation Challenge on AI/ML in 5G in two challenges: by providing efficient environ‑ ments for the assessment of ` The ML5G-PHY [beam-selection] challenge assumes a millimeter new algorithms. wave (mmWave) system in a vehicle-to-infrastructure network using an analog MIMO architecture. The machine learning (ML) model to be developed takes input features such as the ones indicated in How Raymobtime datasets Figure 2 and outputs the indices of the best pair of beams. support the ITU AI and ` The ML5G-PHY [channel estimation] challenge attacks one of the Machine Learning in most difficult problems in the 5G physical layer: acquiring channel 5G Challenge information to establish a mmWave link considering a hybrid MIMO architecture. It consists of estimating the frequency-selective channel from a small number of received training pilots. The Raymobtime methodology was conceived to provide realistic The two challenges look at key 5G challenges where ML can offer simulations of communication valuable solutions, with other examples including user selection, link adaptation, and joint positioning/communication. channels using ray-tracing, taking into account the mobility of trans‑ ceivers and radio scatterers and their evolution over time.

By adopting sensible values for the sampling intervals between distinct scenes (snapshots), Some applications of deep however, depends on the availa‑ ray-tracing leads to simulated learning in 5G networks — such as ble channel data. communication channels with anomaly detection based on per‑ consistency over time, frequency formance indicators from routers Building the wireless infrastruc‑ and space. These channels and other network components — ture and prototypes to make can enable, for instance, the rely on data that is abundant. extensive channel measurement assessment of ML‑based chan‑ campaigns to store these channel nel-tracking methods for multi-us‑ The use of machine learning (ML) data is infeasible, due to cost and er-multiple-input-multiple-output in designing the physical layer, time constraints. Not to mention (MIMO) algorithms. Insights from academia ITU News MAGAZINE No. 05, 2020 72

And Raymobtime datasets are and Ranging) sensor. The open- pedestrians and drones are not restricted to communica‑ source Blender and BlenSor also modeled in 3D, and their tion channels. packages are used to simulate positions in different scenes are cameras and LIDAR sensors, controlled by the open-source Motivated by the increasing respectively. Commercial simulator SUMO (Simulator of usage of sensing information solutions Wireless InSite from Urban Mobility). SUMO allows to aid communication systems, Remcom and Altair’s Winprop are imposing realistic traffic statistics additional software is used to used for ray-tracing. and facilitates studying seasonal compose multimodal datasets. aspects, such as fluctuations in Raymobtime also allows the number of users and data traffic Figure 1 indicates how development of data-driven algo‑ in specific regions over a day. Raymobtime pioneered joint rithms for specific sites in a city. simulations of ray-tracing and The interaction among all soft‑ 3D computer graphics software. To obtain realistic data for out‑ ware packages is orchestrated The result is “paired” information door scenarios, Raymobtime uses via Python, a widely favoured regarding the communication Cadmapper and OpenStreetMap programming language for channel and the respective visual to create 3D simulations of a AI — which can also enable the information obtained from a site’s buildings, streets and stages of feature extraction and camera and/or the point cloud other immovable objects. AI modelling to be integrated, as from a LIDAR (Light Detection Moving objects such as vehicles, depicted in Figure 1.

Figure 1 — Block diagram indicating the software involved and how Raymobtime datasets are used for developing ML‑based solutions to communication problems

Datasets Communication Realistic Scene at time t Ray Tracing Signal parameters scenarios Static Data Processing (MIMO channel, from satellite power, etc) data

LIDAR Data Output Point Cloud Parameters Mobile Input Parameters Images from Feature Positions Simulated Extraction Cameras

Positions SUMO = Simulator of urban mobility ("GPS-based") MIMO = Multiple-input multiple-output Insights from academia ITU News MAGAZINE No. 05, 2020 73

Figure 2 — Example of 3D scenario in which the number of faces was reduced to decrease the ray-tracing execution time. The The number of faces in a 3D This blending of communication vehicles are highlighted in vibrant scene can be reduced to speed systems and virtual reality will colors in the street under study up ray-tracing and create also be leveraged by research in smoother simulations, such as hybrid channel models, com‑ in Figure 2 where faces were bining ray-tracing and statistical reduced to create smoother sim‑ channel models. ulations of the buildings super‑ imposed over satellite imagery of the city scene.

Research continues in opti‑ mizing the tradeoff between Figure 3 — Evolution of realism the speed of ray-tracing and in describing 3D scenarios for the accuracy of 3D scenes for wireless simulations AI‑enabled communications.

Another challenge to improve Raymobtime is the assignment of a material to each face or object that composes a 3D scene. Research to improve Raymobtime Each material has electromag‑ netic properties that impact Computational cost can become ray-tracing, and consequently, a challenge when repeatedly the communication channel. Static ray-tracing with rectangular shapes invoking ray-tracing and other Raymobtime evolved from sim‑ simulation techniques because pler simulations with two materi‑ the time scales differ significantly. als, as shown in Figure 3.

A wireless channel may change Version 2 of Raymobtime will in tens of milliseconds even while incorporate automatic assignment objects like vehicles barely move. of electromagnetic properties to materials. It will also be based The time to execute ray-tracing on improved 3D engines used depends on the complexity of the to generate datasets in a larger Raymobtime 3D scene, which is related to its number of sites to enable studies total number of faces (simulated in the context of transfer learning. renderings of objects).

Raymobtime v2.0 Insights from academia ITU News MAGAZINE No. 05, 2020 74 Shutterstock

Building reliability and trust with network simulators and standards

By Francesc Wilhelmi, Post-Doctoral Researcher, Centre Tecnològic de Telecomunicacions de Catalunya (CTTC), Spain

J Network simulators can play a key part in building reliability and trust in machine learning (ML). But for simulators to deliver this value, innovation towards ML‑enabled 5G and 6G networks must include the Network simulators integration of interoperable testing environments. can play a key part in building Truly autonomous, secure, and reliable ML‑aware communication reliability and trust in systems can be achieved in the near future. And network simulators will machine learning. contribute to this achievement. But the successful adoption of simula‑ tors will depend on the definition and standardization of interopera‑ ble components. Francesc Wilhelmi

The integration of simulators into autonomous ML‑aware networks is now becoming possible with the softwarization and virtualization of networking functions. Insights from academia ITU News MAGAZINE No. 05, 2020 75

As for network simulators, we Why ML5G? Overcoming barriers to find a plethora of proprietary reliability and trust with the and open-source tools (e.g. ns-3, The increasing application help of simulators of machine learning OMNET++, OPNET) for character‑ in communication izing multiple types of scenarios, networks is motivated Machine learning mechanisms technologies, and network‑ by large amounts of can produce non-linear outputs, ing functionalities. unexploited data and prediction functions for example, inherent complexity leading to questions around the of novel use cases such But bringing all these tools as Vehicle to Everything trustworthiness of the outputs together forms a very significant (V2X) communications, produced by these “black boxes”. challenge — one that can only be massive Machine Type Communications (mMTC), addressed through the definition The perhaps most notorious and extended reality and implementation of standard‑ and high-quality video black boxes are deep learning ized interfaces. communications. models for problems with high dimensional spaces, where the Interoperability in future These use cases vary accuracy of a model is typically widely in mobility ML‑aware networks will enable requirements, the numbers tied to its complexity. The greater entities in different domains of implicated devices, and the complexity of the dataset, the such as network functionality, bandwidth and latency higher the number of neurons machine learning repositories, requirements. In complex and hidden layers required by 5G networks, substantial and data providers, to interact performance gains the deep learning model, thereby seamlessly in addressing optimi‑ may result from machine hampering interpretability zation challenges. learning’s ability to learn and explainability. complex patterns and adapt to multiple contexts Cross-domain interoperability will and domains. Moreover, depending on the be one of the main enablers of characteristics of the use case, fully autonomous future net‑ Neural networks, for training data may be scarce, noisy, works, and the conditions for this example, are fast and even non-stationary, thus becoming very popular in interoperability must be estab‑ compromising the reliability of signal processing thanks to lished soon. their ability to characterize ML models and questioning their unknown channel models. sustainability. Highly complex models can also require large computational resources that are not always available. Insights from academia ITU News MAGAZINE No. 05, 2020 76

Consider V2X communica‑ Validate the output of tions, where security is critical, machine learning models ITU standards are and where connected vehicles before being applied to an providing a toolkit to and mobile devices will create operating network. introduce machine complex radio frequency envi‑ learning methods ronments. Machine learning can Generate synthetic data for into 5G networks. help to address such complexity, training machine learning but shortfalls in data and com‑ models, which can be used to putational resources can lead address the lack of data, or to Francesc Wilhelmi to model misbehaviour and extend training data sets. degrade the key performance indicators informing machine Train machine learning models learning’s adoption in networking. in a simulation domain, which is particularly useful to avoid Research towards explainable AI the effects of exploration in shows promise in building trust online learning. have played an important part in in the outputs of neural networks the ITU AI/ Machine Learning in and other complex AI methods, Act as experts to assist the 5G Challenge. but in the near-term network operation of machine learning simulators can provide powerful models, thus providing tools to enhance trust in machine advice in certain situations ITU’s flexible ML‑aware learning. These simulators can (e.g., initialization, narrow architecture and proposed also be integrated into ML‑aware exploration, break ties). ML sandbox communication systems. To illustrate the potential simu‑ ITU standards are provid‑ Network simulators are a lators in future communications ing a toolkit to introduce cost-effective tool to reproduce systems, consider the example of machine learning methods into the behaviour of communica‑ transmitter-receiver implementa‑ 5G networks. tions systems, ranging from tions based on neural networks. communication protocols to Particularly relevant is ITU–T physical phenomena related to Aiming to improve the accuracy Y.3172, which defines an archi‑ signal propagation. of this kind of solution, network tectural framework for the flexible simulators can generate syn‑ integration of machine learning Network simulators can be thetic data characterizing human into networks. integrated into ML‑based net‑ behaviour and thus enrich the works to provide the follow‑ datasets used for training. This is This ITU–T Y.3172 ML‑aware archi‑ ing functionalities: motivating the design of novel tecture proposes the ML sand‑ standardized datasets and tools, box as a solution for enhancing including network simulators, and confidence in machine learn‑ the resulting synthetic datasets ing application. Insights from academia ITU News MAGAZINE No. 05, 2020 77

Proposed machine learning sandbox

SANDBOX NETWORK

Replicate

SIMULATE Manage Monitor

Test/Train

Validate Adopt

ML OPERATIONS

ML = Machine learning

An ML sandbox enables the Network simulators are an impor‑ Ongoing ITU studies are defining testing, training, and evaluation of tant part of the ML sandbox, effec‑ the requirements, architecture, machine learning models in a safe tively and flexibly reproducing and interfaces for the proposed environment before their applica‑ different networking behaviours ML sandbox. tion to operating networks. and scenarios. Insights from academia ITU News MAGAZINE No. 05, 2020 78 Federal University of Technology, Minna, Nigeria Minna, University ofTechnology, Federal

Research projects in Nigeria advancing education and speech recognition

By James Agajo, Associate Professor and Head of WINEST Research Group, Department of Computer Engineering, Abdullahi Sani Shuaibu and Blessed Guda, Students, Federal University of Technology, Minna, Nigeria

J As part of our participation in the ITU Focus Group on Machine Improving education Learning for Future Networks including 5G (FG‑ML5G), in March 2019, in Africa our WINEST (Wireless Networks and Embedded Systems Technologies) Research Group launched a study on “use cases and solutions for This study led us to propose the migrating to IMT‑2020/5G networks in emerging markets”. “AI‑Based Classroom” project, designed to improve education Our focus was to determine how machine learning could help emerg‑ for young pupils in Africa. ing markets to leapfrog technology generations to take advantage of emerging and future networks while optimizing energy consumption, Students are driving the project, network coverage, and communication overheads. under the guidance of James Agajo, Head of WINEST Research Group at the Federal University of Technology, Minna, Nigeria. Insights from academia ITU News MAGAZINE No. 05, 2020 79

An efficient speech recognition project at the 7th Regional This study led us to library is a critical prerequi‑ Workshop on “Standardization of propose the ’AI‑Based site for the development of an future networks towards build‑ Classroom’ project, AI‑based classroom. ing a better-connected Africa” designed to improve in Abuja, Nigeria, 3–4 February education for young This proved very difficult to find. 2020, convened by ITU’s stand‑ pupils in Africa. ardization expert group for future networks and cloud comput‑ Automatic speech ing, ITU Telecommunication James Agajo, Abdullahi Sani recognition for Africa Standardization Sector (ITU–T) Shuaibu, and Blessed Guda Study Group 13. We were in search of a speech recognition library which was able The expert feedback provided by to function locally, meet users’ the Abuja workshop motivated privacy concerns, and was freely our launch of a pilot project in available. In view of the extraor‑ Nigeria to develop an African dinary number of languages automatic speech recognition With AI‑based natural language spoken across Nigeria, and Africa (ASR) system. processing (NLP), classroom at large, we also needed a library conversations between pupils able to perform well in processing We are collecting speech data and teachers are processed the English language accented in and developing the ASR engine at the network edge to extract many different ways. to deliver a prototype able to keywords while maintaining guide the development of a sys‑ speaker anonymity. We evaluated many software tem ready for market deployment. libraries, but none of them These keywords are transmitted succeeded in meeting all of We have developed the to a trained classifier in the central these requirements. “Wazobia” mobile application, server, which is able to recom‑ to support the necessary data mend captivating media content, This led to the launch of a new collection, where Nigerian “voice providing students with intuitive WINEST Research Group project donors” read displayed text examples, supporting a teacher’s in February 2020 to develop a aloud and donate the recording explanations. The media content new speech recognition frame‑ — anonymously. is then shared on a digital display work able to meet the unique in the classroom. requirements of the AI‑Based “Wazobia” is an amalgam of three Classroom project. words meaning “come” in Yoruba The system is designed to (wa), Hausa (zo) and Igbo (bia), augment the efforts of elemen‑ The project evolved from the dis‑ Nigeria’s three largest linguis‑ tary school teachers rather than cussions sparked by our presenta‑ tics groups. attempting to replace them. tion of the AI‑Based Classroom Insights from academia ITU News MAGAZINE No. 05, 2020 80

Architecture of the African automatic speech recognition project AI and machine learning to help Africa manage pandemics

Central Server In future ITU AI/machine ML Model learning in 5G challenges we Training Speech Data also plan to propose a new Text Bluetooth®-enabled contact-trac‑ Speech Data Speech Data ing application supported by [Anonymous, [Unvalidated] Speech Data Unvalidated] [Validated] machine learning.

This pandemic tracing applica‑ tion (PTA) project aims to build Voice Donation Speech Validation Automatic exposure-risk prediction models Speech Recognition ML = Machine-Learning trained from anonymized user data [see table].

The speech data is stored on our AI research article as a refer‑ Data collection for pandemic tracing server as “unvalidated” by default, ence implementation. application contact pending the crowd-sourced detection validation of this data by volun‑ We are progressing with the seg‑ teers via the Wazobia mobile app. mentation and pre-processing of Straight line distance This validation results in Boolean the collected data for supervised between user equipment evaluations of the accuracy of the and semi-supervised machine Bluetooth signal strength ASR engine’s transcriptions of learning settings, but thus far the recorded speech. African ASR project only accepts User equipment model English as an input language. Operating system version To date, the project has collected over three hours of speech cor‑ We aim to introduce African Indoor/outdoor (based on ambient light) pus from over 170 voice donors. languages as inputs as the ASR project develops, and we plan Radio-frequency The African ASR system devel‑ to stimulate this key avenue of interference (wireless opment phase includes data innovation by submitting our local area network) pre-processing, training and speech corpus of African lan‑ software design. The project guages to future ITU challenges uses the Wav2letter++ ASR on AI and Machine Learning in 5G toolkit and looks to Facebook’s and beyond. Insights from academia ITU News MAGAZINE No. 05, 2020 81

Architecture of the Pandemic Tracing Application

Environmental conditions

External sensors Central Server (e.g. weather) ML Model (Training)

Model parameters Environmental conditions Model deployment A B Expert Source (WHO) Close Contact Data

Risk Prediction ML Model

Pandemic Tracing ML = Machine-Learning Application WHO = World Health Organization

The proposed PTA deployment 1. The contact tracing will be The training data will be specific scenario would incorporate generic with configurable to pandemics, as recommended data collected from users with parameters to accommodate by the World Health Organization Bluetooth® discoverable, not future pandemics. (WHO) and other health author‑ requiring all users to install the 2. It will reuse relevant features ities. The resulting exposure-risk application. However, a more of existing frameworks but prediction models will also be detailed picture of the envi‑ customize these features for specific to pandemics (trained ronment can be achieved by application in Africa. and deployed from a cen‑ incorporating data from mobile 3. It will include privacy- tral server). devices’ gyroscopes and accel‑ preserving mechanisms erometers when two Bluetooth®- by design. We are now focused on collecting connected devices both have the 4. The application will determine the required data and we plan PTA installed. the appropriate scope of to submit this data to future ITU data sharing beyond a user’s challenges on AI and Machine The development of the pro‑ device-based user-indicated Learning in 5G and beyond. posed PTA will follow these guid‑ preferences with respect ing principles: to privacy. Insights from academia ITU News MAGAZINE No. 05, 2020 82 Shutterstock

Why we need new partnerships for new data

By Ignacio Rodriguez Larrad, PostDoc, Wireless Communication Networks, Aalborg University, Denmark

J The ITU AI/Machine Learning (ML) in 5G Challenge attracted over 20 problem statements proposed by different industry and academic organizations. Participants from all over the world are competing to find the neural network-based algorithms that lead to The Grand Challenge the optimal solution. Finale in December is an exciting The available problem statements are organized into different tracks prospect. related to diverse themes such as network, security, operators, or vertical markets. Ignacio Rodriguez Larrad The number of statements is certain to increase significantly in future editions of the ITU AI/ML in 5G Challenge. The application of AI and machine learning to networks is gaining strong momentum in indus‑ try and academia, and the ITU AI/ML in 5G Challenge will engage more and more students, researchers and engineering professionals. Insights from academia ITU News MAGAZINE No. 05, 2020 83

The Grand Challenge Finale in Regulation, business and the December is an exciting prospect nature of data all factor consid‑ To achieve accurate — sure to shed new light on how erably into data availability, and functional ML/AI AI and machine learning will help data availability factors con‑ algorithms, high- us to optimize networking. siderably into the success of a quality input data global competition like the ITU is essential. AI/ML in 5G Challenge. New partnerships to expand access to data This also highlights the impor‑ Ignacio Rodriguez Larrad tance of the ITU AI/ML in 5G Access to high-quality data Challenge in encouraging remains a key challenge data availability. facing innovation in AI and machine learning. ITU is working to attract and engage new institutions who Synthetic datasets built with sim‑ are able to share problem state‑ Research networks — real-world ulations have played an impor‑ ments and data with the interna‑ operational networks managed tant part in the ITU AI/ML in 5G tional community. by public research institutions — Challenge in view of the difficul‑ will continue to play a key role in ties associated with access to data To achieve accurate functional the collection of traces for ML/AI from operating networks. And ML/AI algorithms, high-quality processing in the coming years. some of the problem statements input data is essential. were very relevant and well-for‑ Collecting data from research mulated but lacked any source of Extensive datasets from operating networks also requires some data — participants needed to find networks would be ideal inputs effort, and typically also the and provide their own datasets as to AI/ML algorithms. However, at treatment and anonymization of part of their solutions. present it is very difficult to obtain the collected data, but research such collections of data from networks come with fewer com‑ While most of the problem commercial network operators. plexities when it comes to sharing statements were open to inter‑ It requires some effort to col‑ datasets with other institutions. national participants, some were lect specific data and treat and restricted to national competi‑ anonymize it. Operators need to Research institutions from all over tions. However, one has to under‑ protect the privacy of their cus‑ the world are part of the open stand that the data belongs to tomers, and they are also hesitant data movement, with research the institution that provides it and to share data containing busi‑ communities advocating for open is therefore subject to internal/ ness-critical information about the access to experimental research national export regulations and operational status or performance data and associated scientific national general data protec‑ of their networks. findings and articles. tion regulations. Insights from academia ITU News MAGAZINE No. 05, 2020 84

But research networks are not How Aalborg University industrial-grade manufactur‑ easy to build. collaborates with industry ing and production equipment on its AAU 5G Smart from different vendors, includ‑ While it is quite common to see Production Lab ing production line modules, research networks operating robotic arms, and autonomous in unlicensed spectrum such As an example of a research mobile robots. as Wi-Fi networks or wireless network with big ML/AI potential Internet of Things (IoT) networks established at a public research In collaboration with Nokia Bell such as long range (LoRa), it is institute in collaboration with Labs and the telecommunications rare to see research networks in vendors and operators, Aalborg operator Telenor Denmark, the 4G or 5G cellular systems. University (AAU), Denmark, research lab is equipped with 4G hosts one of the most advanced (two private networks), 5G (one These networks come at a much Industry 4.0 wireless playgrounds private, standalone network), and higher cost and research institutes in Europe. Wi-Fi 6 (two networks). This makes will depend on the sponsorship it possible for AAU researchers and close collaboration of ven‑ The recently inaugurated and industrial partners to work dors and operators. AAU 5G Smart Production together in the integration and Lab is a small factory indus‑ testing of advanced Industrial trial research lab with access IoT systems for the factories of to a wide range of operational the future.

AAU 5G Smart Production Lab

UWB positioning NOTE: Multi-connectivity is also possible! Edge-cloud servers system MES = Manufacturing Execution System MES GW AMR fleet GW = Gateway controller management manager AMR = Autonomous Mobile Robot ac = IEEE 802.11ac I/O SCREEN I/O SCREEN ax = IEEE 802.11ax Wi-Fi 5 (ac) PLC IN PLC IN 2.4/5 GHz x3 GW UWB = Ultra-Wideband AMR SWITCH SWITCH GW GW GW pLTE = Private LTE Wi-Fi 6 (ax) dLTE = Dedicated LTE AAU 5G Smart Production Lab x3 5 GHz pNR = Private NR I/O SCREEN dNR = Dedicated NR Wi-Fi 6 (ax) PLC IN x3 LTE = Long-Term Evolution (4G) 5 GHz SWITCH GW NR= New Radio (5G) production module LoRaWAN 4G pLTE x3 GW 868 MHz pLTE 1 3.5 GHz core network 4G pLTE x3 I/O SCREEN Network pLTE 2 PLC IN Internet 3.5 GHz management core network GW SWITCH pNR 5G pNR x3 3.7 GHz 5G dNR x3 core network 4G pLTE Public 1.8/2.1/2.6 GHz x3 4G/5G core 3.5 GHz network

Source: Aalborg University, Denmark, 2020. Insights from academia ITU News MAGAZINE No. 05, 2020 85

Recently inaugurated AAU 5G The lab is also equipped with research possibilities for AAU Smart Production Lab a research LoRa network spon‑ and the international research sored by Cibicom Denmark, community working on AI and and an ultra-wideband position‑ machine learning applications ing system. in networking.

All wireless research networks in the lab are interconnected Let’s expand our work via a central dedicated network together management interface which Source: Aalborg University, Denmark, 2020. allows not only for the control Research networks built together and configuration of the different by research communities and networks but also the monitor‑ industry can generate extensive Equipment in the AAU 5G Smart ing and collection of network and relevant sets of multi-domain Production Lab data traces in dedicated edge- and multi-technology perfor‑ cloud servers. mance data. This data could be very useful for the design and Having the possibility of record‑ optimization of advanced wireless ing unrestricted network infor‑ solutions targeting the different mation, simultaneously with communication requirements of measurements from the differ‑ real-world operational industrial ent user devices, opens new IoT use cases.

Industry 4.0

Source: Aalborg University, Denmark, 2020. Connection Automation

IoT Big Data

System integration Cloud computing

Source: techstartups.com, 2020. Insights from academia ITU News MAGAZINE No. 05, 2020 86 Shutterstock

Machine learning function orchestration for future generation communication networks

By Shagufta Henna, Lecturer in Computing, Letterkenny Institute of Technology, Ireland

J Fifth generation and beyond network operators are interested in exploiting the potentials of machine learning (ML) Fifth generation and to solve intractable problems using a large amount of data. beyond network operators are interested in exploiting However, these network operators struggle to integrate ML the potentials of machine in their networks and often rely on data scientists to create an learning to solve intractable ML pipeline, i.e., from data collection to model deployment. problems using a large An ML pipeline, if not managed and orchestrated appropri‑ amount of data. ately, is subject to bottleneck. Further, lack of standardized ML orchestration mechanisms can result in highly complex and expensive ML pipelines. Shagufta Henna Insights from academia ITU News MAGAZINE No. 05, 2020 87

Other challenges in the context Overcoming the challenges Machine learning function orchestrator of future generation networks with machine learning include ML model updating, function orchestrator model optimization, ML pipeline ML intents chaining, ML pipeline perfor‑ The main objective of machine mance monitoring, evaluation, learning function orchestra‑ pipeline splitting, policy-based tor (MLFO), as per the draft ML pipeline deployment, and ITU Telecommunication management and coordination Standardization Sector Machine learning function orchestrator (MLFO) of multiple ML pipeline instances (ITU–T) Recommendation: across the network. “Requirements, architecture and design for machine learning func‑ These challenging character‑ tion orchestrator”, is to overcome istics call for the requirements the aforementioned challenges Management NF NF NF of big-data-enabled tools that of integration, orchestration, and plane can collect, store, and pre-pro‑ management of pipeline nodes ML underlay cess data to train ML models. and dependencies in an ML pipe‑ Currently, there are efforts to line, while reducing operational ML = Machine learning NF = Network functions address some of the above issues costs. Machine learning function by the big players including Uber, orchestrator offers a unified archi‑ Netflix, Google, Facebook, and tecture to facilitate the orchestra‑ Airbnb, with the help of custom tion of end-to-end ML workflows, ML orchestration platforms. i.e. data collection, pre-process‑ ing, training, model inference, provides flexibility, reusability, However, these solutions aim at model optimization, and model and extension of the ML pipeline in-house ML pipeline manage‑ deployment. It can monitor and to accommodate the rapid pace ment, and do not address the evaluate ML pipeline instances for and development in ML pipe‑ requirements of diverse 5G and performance optimization. line nodes. beyond use cases. In-house ML orchestration solutions require It aims to hide the underlying Future aims will be to consider its significant investments without complexities of orchestrating distributed implementation with the actual realization of their ben‑ ML pipeline nodes by providing minimum complexity and over‑ efits within an industry setting. an abstraction to the users and head. Further, it will be interesting developers with the help of high- to test the machine learning func‑ level application programming tion orchestrator-specific con‑ interfaces (APIs) as illustrated cepts in various 5G and beyond in the figure. Its architecture use cases. Sponsorship opportunities for 2021

What are the benefits of sponsoring next year’s ITU AI/Machine Learning in 5G Challenge?

Brand visibility for an entire year on the Challenge website at the successful ITU AI ML in 5G webinar series at the Grand Challenge Finale

Program opportunities

Media opportunities

Mentorship

Tailored hands-on workshops

If you are interested in sponsoring the 2021 AI/Machine Learning in 5G Challenge, please contact: [email protected] Stay current // // Stay informed Sign up to: Key ICT trends worldwide Insights from ICT thought leaders The latest on ITU events and initiatives

Every Tuesday Regular blogs Six editions per year

Join ITU’s online communities on your favorite channel

Follow podcasts Receive latest news