Periodic Reporting for period 1 - AIoTwin (Twinning action for spreading excellence in Artificial Intelligence of Things)
Reporting period: 2023-01-01 to 2025-12-31
AIoTwin is a twinning coordination action for spreading excellence in Artificial Intelligence of Things, bringing together researchers from three European universities (UNIZG-FER, TUW, and TUB) and one research institute (RISE). Its strategic objective is to significantly strengthen the scientific excellence and innovation capacity of UNIZG-FER in the area of the Internet of Things and Artificial Intelligence (AI) through knowledge transfer and strategic networking activities between UNIZG-FER researchers and world-class scientists from leading EU institutions.
This is achieved through an elaborate work plan and a comprehensive set of activities that have increased researcher mobility and improved the scientific and innovative capacities, as well as the attractiveness and visibility, of UNIZG-FER in the European Research Area. The knowledge gained from networking activities drives a joint research project on data-driven orchestration middleware for energy-efficient IoT supporting machine learning workflows. The middleware was released as open source software and tested on a selected use case using the consortium's shared infrastructure. In addition, the project strengthens the research management and administrative skills of UNIZG-FER staff.
AIoTwin offers a unique opportunity for UNIZG-FER researchers to collaborate closely with leading international researchers to develop new ideas, foster creativity, and enhance the impact of all participating partners at an international level. It helps reduce disparities between Croatia and other EU Member States participating in the Horizon Europe programme.
This is achieved through an elaborate work plan and a comprehensive set of activities that have increased researcher mobility and improved the scientific and innovative capacities, as well as the attractiveness and visibility, of UNIZG-FER in the European Research Area. The knowledge gained from networking activities drives a joint research project on data-driven orchestration middleware for energy-efficient IoT supporting machine learning workflows. The middleware was released as open source software and tested on a selected use case using the consortium's shared infrastructure. In addition, the project strengthens the research management and administrative skills of UNIZG-FER staff.
AIoTwin offers a unique opportunity for UNIZG-FER researchers to collaborate closely with leading international researchers to develop new ideas, foster creativity, and enhance the impact of all participating partners at an international level. It helps reduce disparities between Croatia and other EU Member States participating in the Horizon Europe programme.
The AIoTwin project investigates topics in the field of Artificial Intelligence of Things (AIoT), which connects the Internet of Things (IoT) and Artificial Intelligence (AI), introducing AI into physical environments. The research component of the AIoTwin project develops a data-driven orchestration middleware for energy-efficient IoT to support machine learning workflows. We developed open-source software called the AIoTwin orchestration middleware, which addresses the technical challenges that arise when AIoT solutions are deployed in operational edge-to-cloud environments with heterogeneous devices.
The architecture for the data-driven orchestration middleware for energy-efficient IoT supporting machine learning workflows includes a “learning pipeline” for hierarchical federated learning and an “inference pipeline” capable of meeting specific quality of service constraints associated with inference requests.
The AIoTwin orchestration middleware includes the following components which are published in the AIoTwin GitHub repository https://github.com/aiotwin(opens in new window):
1. Framework for adaptive orchestration of federated learning (FL) supporting hierarchical architectures: enables ML engineers to easily set up HFL pipelines across available edge-to-cloud infrastructure, potentially spanning different administrative domains. HFL training tasks can be initiated with a predefined cost budget, such as communication or energy costs, and the orchestration tool steers the training process to ensure that global models achieve improved accuracy while pipeline costs remain within the specified budget.
2. Implementation of hierarchical federated learning (HFL) services on top of the Flower framework: extension of the popular federated learning framework to support hierarchical setups.
3. QEdgeProxy, a QoS-aware load balancer for dynamic service routing across the computing continuum: serves as an intermediary between IoT devices and inference services placed in the edge-to-cloud continuum.
Components #1 and #2 form the learning pipeline of the AIoTwin orchestration middleware. The solution is based on Kubernetes (K3s), uses the Flower framework, and integrates our original Reconfiguration Validation Algorithm (RVA) to provide a dynamic and intelligent mechanism for deploying, monitoring, and reconfiguring HFL pipelines at runtime. Extensive experiments investigate HFL pipeline reconfiguration under specified latency or energy budgets, as well as performance on the joint infrastructure.
Component #3 forms the inference pipeline of the AIoTwin orchestration middleware. QEdgeProxy provides guarantees that a high percentage of requests originating from the same IoT device are processed within a predefined latency limit or energy budget, even in a highly-dynamic environment with many IoT devices. We reported the results of an extensive evaluation experiment of QEdgeProxy, demonstrating its superiority over state-of-the-art solutions.
In addition to the previous three components of the orchestration middleware for edge environments, the following component was also designed and implemented:
4. Simulator for decentralised training of Spatio-Temporal Graph Neural Networks (ST-GNNs): used to evaluate various distributed ST-GNN training strategies in traffic forecasting scenarios for decentralised prediction of vehicle speed in the context of a Smart City.
The architecture for the data-driven orchestration middleware for energy-efficient IoT supporting machine learning workflows includes a “learning pipeline” for hierarchical federated learning and an “inference pipeline” capable of meeting specific quality of service constraints associated with inference requests.
The AIoTwin orchestration middleware includes the following components which are published in the AIoTwin GitHub repository https://github.com/aiotwin(opens in new window):
1. Framework for adaptive orchestration of federated learning (FL) supporting hierarchical architectures: enables ML engineers to easily set up HFL pipelines across available edge-to-cloud infrastructure, potentially spanning different administrative domains. HFL training tasks can be initiated with a predefined cost budget, such as communication or energy costs, and the orchestration tool steers the training process to ensure that global models achieve improved accuracy while pipeline costs remain within the specified budget.
2. Implementation of hierarchical federated learning (HFL) services on top of the Flower framework: extension of the popular federated learning framework to support hierarchical setups.
3. QEdgeProxy, a QoS-aware load balancer for dynamic service routing across the computing continuum: serves as an intermediary between IoT devices and inference services placed in the edge-to-cloud continuum.
Components #1 and #2 form the learning pipeline of the AIoTwin orchestration middleware. The solution is based on Kubernetes (K3s), uses the Flower framework, and integrates our original Reconfiguration Validation Algorithm (RVA) to provide a dynamic and intelligent mechanism for deploying, monitoring, and reconfiguring HFL pipelines at runtime. Extensive experiments investigate HFL pipeline reconfiguration under specified latency or energy budgets, as well as performance on the joint infrastructure.
Component #3 forms the inference pipeline of the AIoTwin orchestration middleware. QEdgeProxy provides guarantees that a high percentage of requests originating from the same IoT device are processed within a predefined latency limit or energy budget, even in a highly-dynamic environment with many IoT devices. We reported the results of an extensive evaluation experiment of QEdgeProxy, demonstrating its superiority over state-of-the-art solutions.
In addition to the previous three components of the orchestration middleware for edge environments, the following component was also designed and implemented:
4. Simulator for decentralised training of Spatio-Temporal Graph Neural Networks (ST-GNNs): used to evaluate various distributed ST-GNN training strategies in traffic forecasting scenarios for decentralised prediction of vehicle speed in the context of a Smart City.
The orchestration middleware includes a framework for adaptive orchestration of Hierarchical Federated Learning (HFL) pipelines with high innovation potential. This framework enables ML engineers to easily set up complex HFL pipelines across the available edge-to-cloud infrastructure, potentially spanning different administrative domains. HFL training tasks can be initiated with a predefined cost budget, such as communication or energy costs, and the orchestration tool steers the training process to ensure that global models achieve improved accuracy while pipeline costs remain within the specified budget. This is appealing for collaborative businesses that are unwilling to share their data but are willing to jointly train and develop improved ML models based on their datasets. In an HFL setup, the data remains local, while models – specifically, model parameters – are shared between the businesses.