Periodic Reporting for period 1 - ADROIT6G (Distributed Artificial Intelligence-driven open and programmable architecture for 6G networks)
Reporting period: 2023-01-01 to 2024-06-30
The project’s overall objective include:
TO1: Design and implement a novel 6G system architecture that integrates a distributed AI framework for combined communication, computation and control and empowers the convergence of networks and IT systems to enable new future digital services.
TO2: Create an AI-driven Management & Orchestration (M&O) and control framework for 6G Networks.
TO3: Architect a distributed and secure CrowdSourcing AI.
TO4: Develop energy-aware models for multimodal Representation Learning.
TO5: Evolve the cellular infrastructure to allow the true integration of deep-edge devices in communication and computation functions.
TO6: Enable Non-Terrestrial Networks (NTN) connectivity for highly reliable Industrial IoT Services.
TO7: Extend and demonstrate the use of decentralized AI for Device-to-Device (D2D) communications.
TO8: Support data plane acceleration.
TO9: Integrate and demonstrate the potential and user value of ADROIT6G through relevant experimentation, testing, and validation of its innovations in PoCs in lab settings.
TO1: Design and implement a novel 6G system architecture that integrates a distributed AI framework for combined communication, computation and control and empowers the convergence of networks and IT systems to enable new future digital services.
TO2: Create an AI-driven Management & Orchestration (M&O) and control framework for 6G Networks.
TO3: Architect a distributed and secure CrowdSourcing AI.
TO4: Develop energy-aware models for multimodal Representation Learning.
TO5: Evolve the cellular infrastructure to allow the true integration of deep-edge devices in communication and computation functions.
TO6: Enable Non-Terrestrial Networks (NTN) connectivity for highly reliable Industrial IoT Services.
TO7: Extend and demonstrate the use of decentralized AI for Device-to-Device (D2D) communications.
TO8: Support data plane acceleration.
TO9: Integrate and demonstrate the potential and user value of ADROIT6G through relevant experimentation, testing, and validation of its innovations in PoCs in lab settings.
During the reporting period, in WP2 ADROIT6G has provided a comprehensive overview of the Use Cases (UC) facilitated by the ADROIT6G system. The Key Performance Indicators (KPIs) and Key Value Indicators (KVIs) definitions are also introduced to assess the project's success and effectiveness by providing the baseline and target values that need to be validated within the Proof-of-Concepts (PoCs) for KPIs and which KVI is relevant for every service class. Moreover, ADROIT6G has provided the initial version of the ADROIT6G architecture, by defining the three main frameworks that operate on top of a programmable inter-computing and internetwork infrastructure, i.e. the AI-driven Management and Orchestration Framework, the Fully distributed and secure AI/ML Framework for CrowdSourcing AI, and the (Belief-Desire-Intention eXtended) BDI- & AI-driven Unified & Open Control Operations Framework. In addition, the closed-loop function approach of ADROIT6G is also defined, targeting to address the complexities inherent in managing 6G networks. Furthermore, the physical infrastructure is also defined by providing details on the Far-Edge, the Edge, and the Cloud domains.
WP3 is devoted to the development of algorithmic solutions for the effective training of machine learning models and the execution of inference tasks, exploiting an efficient AI/ML architectural framework and information flows. The developed solutions aim to realize of the CrowdSourcing AI concept, i.e. sharing the already-acquired knowledge in the form of both datasets describing a phenomenon or process, and trained ML models so that decision making entities can save computational, network, and energy resources while ensuring high-quality decisions. The energy-efficient distributed schemes defined within this WP thus leverage the resources available across various, heterogeneous nodes while meeting the service requirements on learning and decision quality level.
WP4 is dedicated to the innovations that make up the Management and Orchestration and Control Frameworks as well as methods for data plane acceleration. For the Management and Orchestration, the WP is focused on innovative methods for providing intelligent vertical service and slice management using AI-driven decision solutions over distributed, heterogeneous systems including far edge, edge, and cloud resources, via terrestrial and non-terrestrial networks. The Control framework components are instrumental in enhancing the overall network's efficiency and agility through the UE-VBS concept for far-edge resources aided by BDIx agents for real-time and near real-time control functions.
WP4 has worked toward the development of the Management and Orchestration and Control Frameworks in the design and implementation of their functional components. The WP has also defined methods and mechanisms for multi-timescale and multi-domain closed-loop services, including internal components of BDIx agents as well as monitoring and decision closed-loop functions for network optimization. The UE-VBS concept has been studied and elaborated to fit into the ADROIT6G architecture, and algorithms for use in optimizing D2D communications through BDIx agents in the UE-VBS concept have been developed and evaluated. The WP has also made progress in the data-plane acceleration through P4 and the Satellite advancements.
WP5 focus on the integration of the ADROIT6G architectural components developed under WP3 and WP4, the setup the testbed environments and the technology validations and performance evaluation in a lab environment through three representative extreme 6G use cases (i.e. holographic telepresence, Industrial IoT, collaborative robots/drones) covering extreme eMBB, URLLC and mMTC service classes, in corresponding Proofs of Concepts (PoCs) over two Testing Cycles. For the validation of ADROIT6G concept and technologies, theoretical considerations, simulation, emulation and upgraded 5G testbeds will be used, as appropriate.
WP3 is devoted to the development of algorithmic solutions for the effective training of machine learning models and the execution of inference tasks, exploiting an efficient AI/ML architectural framework and information flows. The developed solutions aim to realize of the CrowdSourcing AI concept, i.e. sharing the already-acquired knowledge in the form of both datasets describing a phenomenon or process, and trained ML models so that decision making entities can save computational, network, and energy resources while ensuring high-quality decisions. The energy-efficient distributed schemes defined within this WP thus leverage the resources available across various, heterogeneous nodes while meeting the service requirements on learning and decision quality level.
WP4 is dedicated to the innovations that make up the Management and Orchestration and Control Frameworks as well as methods for data plane acceleration. For the Management and Orchestration, the WP is focused on innovative methods for providing intelligent vertical service and slice management using AI-driven decision solutions over distributed, heterogeneous systems including far edge, edge, and cloud resources, via terrestrial and non-terrestrial networks. The Control framework components are instrumental in enhancing the overall network's efficiency and agility through the UE-VBS concept for far-edge resources aided by BDIx agents for real-time and near real-time control functions.
WP4 has worked toward the development of the Management and Orchestration and Control Frameworks in the design and implementation of their functional components. The WP has also defined methods and mechanisms for multi-timescale and multi-domain closed-loop services, including internal components of BDIx agents as well as monitoring and decision closed-loop functions for network optimization. The UE-VBS concept has been studied and elaborated to fit into the ADROIT6G architecture, and algorithms for use in optimizing D2D communications through BDIx agents in the UE-VBS concept have been developed and evaluated. The WP has also made progress in the data-plane acceleration through P4 and the Satellite advancements.
WP5 focus on the integration of the ADROIT6G architectural components developed under WP3 and WP4, the setup the testbed environments and the technology validations and performance evaluation in a lab environment through three representative extreme 6G use cases (i.e. holographic telepresence, Industrial IoT, collaborative robots/drones) covering extreme eMBB, URLLC and mMTC service classes, in corresponding Proofs of Concepts (PoCs) over two Testing Cycles. For the validation of ADROIT6G concept and technologies, theoretical considerations, simulation, emulation and upgraded 5G testbeds will be used, as appropriate.
-Inter-domain Migration: Local RL Agent Dueling Double DQN (DDDQN) is used to design the Predictors for VNF Migration.
-Inter-domain VNF placement strategy: distributed RL agents in different domains generate an action based on their DDDQN assessment and bid with their confidence to host the VNF in their domain.
-Migration of ML models is implemented as VNFs, whenever such models are implemented as stateful micro services.
-New multi-modal generative models, capable of cross-generating an arbitrary number of modalities (Masked-Multi-Time Stochastic Differential Equations).
-Engineering of efficient schemes that allow seamless scalability in the number of modalities for the training procedure. Encoders and Decoders at different network edge points for a lightweight transmission.
-Dependable and fair distributed training strategies to (i) minimize learning time (proxy for cost), or/and (ii) reduce resource consumption while meeting requirements on loss.
-Meta-learning, distributed approach for ML model selection, Crowdsourced ML model creation, and adaptation for 1) matching nodes’ computing capabilities and 2)meet time & accuracy requirements.
-2-step Transfer learning: clusters of similar (in traffic distribution or geo proximity) gNBs coordinated by an edge server for hyperparameter optimisation during training (FL).
-A scalable and energy-efficient FL algorithm: a lightweight Hessian estimation and adaptive step sizes to improve convergence speed and reduce communication overhead in distributed networks.
-Satellite communication and computing framework emulation.
-Unified Programmable Data Plane and Traffic Isolation by using P4 on DPUs.
-UE-VBS Computing Resource Control.
-Service Orchestration/Management.
-SLA-driven FL solution for 6G zero touch network slice management
-BDIx agents
-Federated Learning and Deep Feedback Neural Networks in the DAI Framework for Autonomous Radio Resource Management in D2D Communications.
-Energy-Efficient Edge Resource Management in Multi-User Goal Oriented Communication.
-Semantic Communications Based on Adaptive Generative Models and Information Bottleneck.
-Information-theoretic analysis of multi-modal sources and adaptive off-loading.
-Monitoring functions for distributed data collection.
-Decision Engine framework
-Algorithms for predictive monitoring and SLA offering.
-Closed-loop Governance
-Inter-domain VNF placement strategy: distributed RL agents in different domains generate an action based on their DDDQN assessment and bid with their confidence to host the VNF in their domain.
-Migration of ML models is implemented as VNFs, whenever such models are implemented as stateful micro services.
-New multi-modal generative models, capable of cross-generating an arbitrary number of modalities (Masked-Multi-Time Stochastic Differential Equations).
-Engineering of efficient schemes that allow seamless scalability in the number of modalities for the training procedure. Encoders and Decoders at different network edge points for a lightweight transmission.
-Dependable and fair distributed training strategies to (i) minimize learning time (proxy for cost), or/and (ii) reduce resource consumption while meeting requirements on loss.
-Meta-learning, distributed approach for ML model selection, Crowdsourced ML model creation, and adaptation for 1) matching nodes’ computing capabilities and 2)meet time & accuracy requirements.
-2-step Transfer learning: clusters of similar (in traffic distribution or geo proximity) gNBs coordinated by an edge server for hyperparameter optimisation during training (FL).
-A scalable and energy-efficient FL algorithm: a lightweight Hessian estimation and adaptive step sizes to improve convergence speed and reduce communication overhead in distributed networks.
-Satellite communication and computing framework emulation.
-Unified Programmable Data Plane and Traffic Isolation by using P4 on DPUs.
-UE-VBS Computing Resource Control.
-Service Orchestration/Management.
-SLA-driven FL solution for 6G zero touch network slice management
-BDIx agents
-Federated Learning and Deep Feedback Neural Networks in the DAI Framework for Autonomous Radio Resource Management in D2D Communications.
-Energy-Efficient Edge Resource Management in Multi-User Goal Oriented Communication.
-Semantic Communications Based on Adaptive Generative Models and Information Bottleneck.
-Information-theoretic analysis of multi-modal sources and adaptive off-loading.
-Monitoring functions for distributed data collection.
-Decision Engine framework
-Algorithms for predictive monitoring and SLA offering.
-Closed-loop Governance