Aristotle University Of Thessaloniki (AUTH)
Scalable frameworks for sustainable and distributed intelligence systems at the edge
Abstract
dc:descriptionSustainability has attracted increasing attention in recent years as the number of interconnected and Internet of Things devices continues to grow and computational workloads become progressively more complex. At the same time, artificial intelligence is being adopted more widely by organizations and industries, while AI models are becoming larger and more computationally demanding. These developments increase both the energy consumption and the resource requirements associated with data processing, model training, and inference. Under these conditions, conventional centralized data-centre architectures are gradually evolving toward hybrid multi-cloud and multi-edge environments in which multiple distributed resource providers participate in computation and service delivery. Although this transition improves flexibility and scalability, it also introduces several new challenges. At the system-design level, platforms ranging from low-level hardware components to high-level neural network implementations are becoming increasingly complex, and their design choices can significantly affect energy consumption, resource utilization, and system performance. At the orchestration level, hybrid multi-cloud--multi-edge environments introduce highly dynamic and heterogeneous infrastructures, making efficient resource allocation and scheduling increasingly difficult. Furthermore, the growing relevance of edge computing has encouraged the emergence of collaborative and distributed edge-AI paradigms, with federated learning representing one of their most prominent implementations. This dissertation addresses these challenges across the main stages of the edge and hybrid cloud--edge continuum. First, it proposes machine-learning- and reinforcement-learning-based methodologies for automated system design, with the objective of satisfying user-defined specifications and sustainability constraints, such as energy efficiency. Second, it introduces a distributed orchestration methodology that decomposes a complex centralized decision-making problem into several simpler local orchestration problems. The participating agents collaborate to determine efficient scheduling decisions according to the principle of acting locally to optimize globally. In addition, the dissertation proposes a fully serverless and asynchronous federated learning scheme based on direct peer-to-peer communication. The scheme is combined with an adaptive aggregation strategy designed to reduce the impact of stragglers, redundant updates, and statistical heterogeneity in client data distributions. This approach enables the development of decentralized and scalable serverless AI clusters without relying on a permanent central coordinator. Finally, the proposed methodologies and frameworks are evaluated using real-world systems and real-time decision-making applications. The examined use cases include an acoustic monitoring system for detecting queenless beehives and an ultra-wideband indoor positioning system enhanced through federated learning. The dissertation therefore follows a hierarchical progression, beginning with single intelligent-node approaches for automated system design, advancing to distributed multi-node intelligence with central coordination for system orchestration, and culminating in fully serverless, peer-to-peer, and asynchronous federated learning. Collectively, these contributions address key challenges in the development of sustainable, scalable, and intelligent systems across the edge continuum.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Sad, Christos
- Σάντ, Χρήστος
Subjects
dc:subject × 14- Κλιμακούμενα πλαίσια
- Ευφυή συστήματα
- Κατανεμημένη μάθηση και ενορχήστρωση
- Ενεργειακή αποδοτικότητα στην άκρη του δικτύου
- Scalable frameworks
- Intelligent systems
- Distributed learning and orchestration
- Energy efficiency at the edge
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Υπολογιστές, Υλικό (hardware) και Αρχιτεκτονική
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Computer science, Hardware and Architecture
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/62948