Call: H2020-ICT-2020-2 Project reference: 101015956 Project Name: A flagship for B5G/6G vision and intelligent fabric of technology enablers connecting human, physical, and digital worlds Hexa-X Deliverable D4.1 AI-driven communication & computation co-design: Gap analysis and blueprint Date of delivery: 31/08/2021 Version: 1.0 Start date of project: 01/01/2021 Duration: 30 months Document properties: Document Number: D4.1 Document Title: AI-driven communication & computation co-design: Gap analysis and blueprint Editor(s): Aspa Skalidi (WIN), Tamas Borsos (EHU), Quentin Lampin (ORA), Miltiadis Filippou (INT), Leonardo Gomes Baltar (INT) Authors: Aspa Skalidi (WIN), Tamas Borsos (EHU), Quentin Lampin (ORA), Miltiadis Filippou (INT), Leonardo Gomes Baltar (INT), Alessio Bechini (UPI), Andras Benczur (SZT), Giacomo Bernini (NXW), Emilio Calvanese Strinati (CEA), Panagiotis Demestichas (WIN), Pietro Ducange (UPI), Johan Haraldson (EAB), Insaf Ismath (OUL), Dani Korpi (NOF), Ignacio Labrador (ATO), Giada Landi (NXW), Guillaume Larue (ORA), Luc Le Magoarou (BCO), Dileepa Marasinghe (OUL), Francesco Marcelloni (UPI), Ricardo Marco-Alaez (ATO), Mattia Merluzzi (CEA), Jafar Mohammadi (NOG), Markus Mueck (INT), Stéphane Paquelet (BCO), Pietro Piscione (NXW), András Rácz (EHU), Nuwanthika Rajapaksha (OUL), Nandana Rajatheva (OUL), Vismika Ranasinghe (OUL), Alessandro Renda (UPI), Elif Ustundag Soykan (EBY), Emrah Tomur (EBY) Contractual Date of 31/08/2021 Delivery: Dissemination PU1 level: Status: Final Version: 1.0 File Name: Hexa-X D4.1_v0.1 Revision History Revision Date Issued by Description 18.02.2021 Hexa-X WP4 ToC 09.04.2021 Hexa-X WP4 Draft for internal review 13.05.2021 Hexa-X WP4 Draft for external review 24.06.2021 Hexa-X WP4 Draft for PMT review 26.07.2021 Hexa-X WP4 D4.1 for GA approval 30.08.2021 Hexa-X WP4 Final 1 PU = Public Abstract This report will provide the rationale leading to the incorporation of AI/ML in 6G networks and document gaps that need to be addressed to make it possible. Built upon them, associated problems will be detailed and resulting solution directions will be presented. Applications in the air interface will be considered first and, subsequently, in-network learning methods will be investigated. Keywords 6G, services, Artificial Intelligence, Machine Learning, Connecting Intelligence Disclaimer The information and views set out in this deliverable are those of the author(s) and do not necessarily reflect views of the whole Hexa-X Consortium, nor the official opinion of the European Union. Neither the European Union institutions and bodies nor any person acting on their behalf may be held responsible for the use which may be made of the information contained therein. This project has received funding from the European Union’s Horizon 2020 research and innovation programmed under grant agreement No 101015956. Hexa-X Deliverable D4.1 Executive Summary This report is the first deliverable of project Hexa-X work package four (WP4) “AI-driven communication and computation co-design”, led by task T4.1 - “Gap analysis for AI-driven communication and computation co-design”. It focuses on the main gaps related to WP4 work and concentrates on associated problem areas (statements) to be addressed, as well as on the description of solution directions. Work concludes with an outlook on planned next steps. The purpose of this document is to depict motivations for the utilization of Artificial Intelligence (AI) and, in particular, Machine Learning (ML) mechanisms in 6G systems and identify the major challenges that arise. In addition to performing an extensive gap analysis, it also investigates potential approaches and delivers a set of recommendations for future work. Input is yielded from deliverables: D1.1 “6G Vision, use cases and key societal values”, D1.2 “Expanded 6G vision, use cases and societal values – including aspects of sustainability, security and spectrum”, D2.1 “Towards Tbps Communications in 6G: Use Cases and Gap Analysis”, D6.1 “Gaps, features and enablers for B5G/6G service management and orchestration” and D7.1 “Gap analysis and technical work plan for special-purpose functionality”. The resulting guidelines are intended to direct the work in the following tasks of WP4, namely task 4.2 (T4.2 - “AI-driven air interface design”) and task 4.3 (T4.3 - “Methods and algorithms for sustainable and secure distributed AI”). The overall storyline of introducing AI in 6G networks, including the motivating challenges and aspired benefits is presented, followed by the definition of fundamental AI concepts and a summary of common practices. The role of data is clarified, including considerations related to data quality, quantity, availability, ownership and monetisation, along with aspects on in-network data privacy, security and integrity. Focusing on a subset of identified 6G use cases and their connected performance and value indicators, as detailed in Hexa-X deliverable D1.2, potential applications are investigated, starting with applications to the air interface and continuing with in-network learning methods. Regarding AI-based air interface design, four main pillars are considered. The first is about novel, data-driven transceiver design approaches, accounting for hardware impairments in the transmitter and receiver radio frequency (RF) chains. Secondly, AI-driven transmitters are considered with an investigation of AI optimization of beamforming and Reinforcement Learning (RL) methods for fast initial access. Thirdly, AI-driven receiver design is discussed, including exploration of functionalities, such as channel estimation and channel decoding, as well as receiver side processing as a single block (end-to-end optimization). Finally, concerning AI- driven radio interface functionality, this document examines different approaches for radio resource management (RRM), cell-free and distributed massive MIMO (multiple input, multiple output) systems, as well as model predictive control of antenna systems. The topic of in-network learning is organized in three main parts. The section of joint communication and computation co-design investigates different approaches for distributed learning taking into account aspects of trustworthiness, sustainability, efficiency and resilience. Subsequently, enablers for in-network AI security, privacy and trust are analysed, including privacy concerns, explainability features and prevention and mitigation of in-network AI functionality attacks. Lastly, AI-powered network operation incorporating predictive orchestration for behaviour-driven adaptation along with intrusion detection procedures is presented. Dissemination level: public Page 4 / 101 Hexa-X Deliverable D4.1 Table of Contents Executive Summary .................................................................................................................... 4 Table of Contents ........................................................................................................................ 5 List of Figures .............................................................................................................................. 7 List of Tables ............................................................................................................................... 8 List of Acronyms and Abbreviations ......................................................................................... 9 1 Introduction......................................................................................................................... 15 1.1 Objective of the document ........................................................................................... 15 1.2 Structure of the document ............................................................................................ 15 2 Connecting Intelligence in 6G - overall trends and challenges ....................................... 16 2.1 Artificial Intelligence and Machine Learning in 6G networks .................................... 16 2.2 AI/ML primer .............................................................................................................. 17 2.3 AI-driven air interface design ...................................................................................... 18 2.3.1 ML-based modelling at the air interface ................................................................. 18 2.3.2 ML-based optimisation of the air interface............................................................. 19 2.3.3 ML at the air interface - the key challenges ............................................................ 19 2.4 In-network learning methods and algorithms .............................................................. 20 2.5 Architectural implications to in-network AI/ML ......................................................... 21 2.5.1 AI agent discovery and selection ............................................................................ 22 2.5.2 AI service pairing inferencing tasks to learning algorithms and topologies ........... 24 2.6 Use cases with AI/ML relevance ................................................................................. 25 2.6.1 Use cases, KPIs and KVIs relevant to AI-driven air interface design .................... 25 2.6.2 Use cases, KPIs and KVIs relevant to in-network learning methods and algorithms ............................................................................................................... 28 2.7 Challenges for data management, ownership, and privacy .........................................