Periodic Reporting for period 1 - CENTRIC (Towards an AI-native, user-centric air interface for 6G networks)
Période du rapport: 2023-01-01 au 2023-12-31
The CENTRIC European project aims to leverage Artificial Intelligence techniques through a top-down, modular approach to wireless connectivity that puts the users’ communication needs and environmental constraints at the center of requirements for the wireless network communications protocol stack design.
The CENTRIC project is designing a user-centric AI Air Interface (AI-AI) applying AI techniques to create and customize tailor-made waveforms, dynamically configure transceivers, and enable appropriate signaling and medium access protocols to support these requirements. The results of CENTRIC will be validated and demonstrated in laboratory prototypes and its breakthroughs will enable future 6G use cases, such as self-driving vehicles, the internet of nano bio-things, or multi-sensory holographic communications.
PROJECT VISION
CENTRIC positions the AI-AI as the essential fabric of future wireless connectivity systems, for the benefit of mobile network operators and private networks markets, which will profit from highly customizable systems wherein different customers have entirely distinct personalized needs. For instance, a university campus is unlikely to have the same connectivity requirements as an indoor smart factory, as well as new and radically different communication needs, will emerge in the near future. Addressing them with a traditional multi-purpose wireless network would significantly limit the scope of future applications, and potentially prevent many cutting-edge inventions from fulfilling their potential.
The CENTRIC aims at being “future-proof,” endowing the networks of tomorrow with AI tools to support a range of new applications.
CENTRIC will explore and deliver AI-based breakthroughs in some of the most pressing 6G Radio Resource Management (RRM) areas, such as sustainable data management at the wireless edge, Electro-Magnetic Field (EMF) exposure control in novel network architectures and energy savings, seamless interoperability among services and applications, etc.
PROJECT OBJECTIVES
Objective 1 – To develop AI methods for the discovery of novel and efficient waveforms
- Machine Learning (ML) techniques that enable the discovery and optimization of tailor-made waveforms targeting the user communication needs while accounting the radio-front and computing hardware limitations as well as the propagation environment
Objective 2 – To develop AI methods for the discovery of novel and efficient transceivers
- Efficient AI-based transceivers for two particularly challenging use-cases: large-scale extreme massive multiple-input multiple-output (MIMO) deployments beam-based communications in millimeter-wave (mmWave) spectrum
Objective 3 – To develop AI methods for the discovery of customized lightweight communication protocols
- Learning-to-communicate (L2C) techniques based on multi-agent reinforcement learning (MARL), semantic communications, and edge control/caching.
Objective 4 – To introduce novel end-to-end hardware co-design solutions for energy-efficient AI-native transceivers
- CENTRIC will explore, design and validate AI accelerators that are based on both standard digital technology and emerging mixed-analog digital platforms based on non-volatile memory elements.
Objective 5 – To develop training and monitoring environments as enablers for AI-AI deployments
- tools for network designers and operators to deploy, train, and monitor AI models general CENTRIC framework based on digital twins (DT), which encompasses cloud-based agents running networked AI models of physical agent counterparts
Objective 6 – To validate user-centric AI-AI solutions in a lab setting
- CENTRIC will define novel testing and benchmarking frameworks for user-centric AI-based communications.
Objective 7 – To demonstrate and disseminate AI-AI concepts
- Laboratory proof-of-concept (PoC) implementations to validate the expected gains and shed light on the feasibility and cost of the AI-AI implementation.
The CENTRIC project is designing a user-centric AI Air Interface (AI-AI) applying AI techniques to create and customize tailor-made waveforms, dynamically configure transceivers, and enable appropriate signaling and medium access protocols to support these requirements. The results of CENTRIC will be validated and demonstrated in laboratory prototypes and its breakthroughs will enable future 6G use cases, such as self-driving vehicles, the internet of nano bio-things, or multi-sensory holographic communications.
PROJECT VISION
CENTRIC positions the AI-AI as the essential fabric of future wireless connectivity systems, for the benefit of mobile network operators and private networks markets, which will profit from highly customizable systems wherein different customers have entirely distinct personalized needs. For instance, a university campus is unlikely to have the same connectivity requirements as an indoor smart factory, as well as new and radically different communication needs, will emerge in the near future. Addressing them with a traditional multi-purpose wireless network would significantly limit the scope of future applications, and potentially prevent many cutting-edge inventions from fulfilling their potential.
The CENTRIC aims at being “future-proof,” endowing the networks of tomorrow with AI tools to support a range of new applications.
CENTRIC will explore and deliver AI-based breakthroughs in some of the most pressing 6G Radio Resource Management (RRM) areas, such as sustainable data management at the wireless edge, Electro-Magnetic Field (EMF) exposure control in novel network architectures and energy savings, seamless interoperability among services and applications, etc.
PROJECT OBJECTIVES
Objective 1 – To develop AI methods for the discovery of novel and efficient waveforms
- Machine Learning (ML) techniques that enable the discovery and optimization of tailor-made waveforms targeting the user communication needs while accounting the radio-front and computing hardware limitations as well as the propagation environment
Objective 2 – To develop AI methods for the discovery of novel and efficient transceivers
- Efficient AI-based transceivers for two particularly challenging use-cases: large-scale extreme massive multiple-input multiple-output (MIMO) deployments beam-based communications in millimeter-wave (mmWave) spectrum
Objective 3 – To develop AI methods for the discovery of customized lightweight communication protocols
- Learning-to-communicate (L2C) techniques based on multi-agent reinforcement learning (MARL), semantic communications, and edge control/caching.
Objective 4 – To introduce novel end-to-end hardware co-design solutions for energy-efficient AI-native transceivers
- CENTRIC will explore, design and validate AI accelerators that are based on both standard digital technology and emerging mixed-analog digital platforms based on non-volatile memory elements.
Objective 5 – To develop training and monitoring environments as enablers for AI-AI deployments
- tools for network designers and operators to deploy, train, and monitor AI models general CENTRIC framework based on digital twins (DT), which encompasses cloud-based agents running networked AI models of physical agent counterparts
Objective 6 – To validate user-centric AI-AI solutions in a lab setting
- CENTRIC will define novel testing and benchmarking frameworks for user-centric AI-based communications.
Objective 7 – To demonstrate and disseminate AI-AI concepts
- Laboratory proof-of-concept (PoC) implementations to validate the expected gains and shed light on the feasibility and cost of the AI-AI implementation.
In the scope of research towards novel E2E learning methods for waveforms, transceivers, and modulation, the CENTRIC identified the following algorithms for further consideration in the project, including six scenarios for the CENTRIC waveform benchmarking framework:
- Multi-user MIMO neural receiver: an enhanced neural network-based receiver with support for flexible MU-MIMO components and 5G NR physical uplink shared channel (PUSCH) compatibility
- AI-based modulation learning for sub-THz: a scalable end-to-end autoencoder-based PHY layer technique for symbol modulation/demodulation
- AI-based CSI compression: an improved variant of a separate sequential training scheme called NW-first with a raw dataset-sharing scheme
- AI-aided UE array dimensionality reduction: an approach for reducing UE array dimensionality in large MIMO systems.
- Enhanced NW first separate training for two-sided AI/ML CSI feedback compression.
- ML based device-constrained assistance information beam prediction and device agnostic framework for beam measuring and codebook selectionReinforcement learning based resource allocation for joint communication and sensing in mmWave environments
Furthermore, novel AI-based techniques for automating the generation of lean layer-2 (L2) & RRM protocols have been investigated. In this respect, a list of relevant challenges in protocol emergence has been constructed and open-source software repositories released, including the following:
- Random channel access with multi-agent reinforcement learning
- Multiple access to radio channels for robotic tasks
- Realistic generator of Downlink Control Information (DCI) messages
- Simulator of in-factory 6G subnetworks for ML-based sub-band allocation
In addition, the CENTRIC project has main significant progress in the area of AI-AI hardware and enablers resulting in the following main contributions:
- Implementation of a transformer architecture using neuromorphic computing
- Novel calibration algorithm for ray tracing in digital twins via phase-aware optimization.
- Implementation and performance evaluation of a neural receiver on Synthara hardware platform.
- New ARQ protocol for edge-based AI
- Multi-user MIMO neural receiver: an enhanced neural network-based receiver with support for flexible MU-MIMO components and 5G NR physical uplink shared channel (PUSCH) compatibility
- AI-based modulation learning for sub-THz: a scalable end-to-end autoencoder-based PHY layer technique for symbol modulation/demodulation
- AI-based CSI compression: an improved variant of a separate sequential training scheme called NW-first with a raw dataset-sharing scheme
- AI-aided UE array dimensionality reduction: an approach for reducing UE array dimensionality in large MIMO systems.
- Enhanced NW first separate training for two-sided AI/ML CSI feedback compression.
- ML based device-constrained assistance information beam prediction and device agnostic framework for beam measuring and codebook selectionReinforcement learning based resource allocation for joint communication and sensing in mmWave environments
Furthermore, novel AI-based techniques for automating the generation of lean layer-2 (L2) & RRM protocols have been investigated. In this respect, a list of relevant challenges in protocol emergence has been constructed and open-source software repositories released, including the following:
- Random channel access with multi-agent reinforcement learning
- Multiple access to radio channels for robotic tasks
- Realistic generator of Downlink Control Information (DCI) messages
- Simulator of in-factory 6G subnetworks for ML-based sub-band allocation
In addition, the CENTRIC project has main significant progress in the area of AI-AI hardware and enablers resulting in the following main contributions:
- Implementation of a transformer architecture using neuromorphic computing
- Novel calibration algorithm for ray tracing in digital twins via phase-aware optimization.
- Implementation and performance evaluation of a neural receiver on Synthara hardware platform.
- New ARQ protocol for edge-based AI
n/a for the periodic report - to be provided in the final report.