PhD Oral Exam - Junior Momo Ziazet, Computer Science

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

This thesis proposes a unified framework for enabling cognitive and sustainable communication networks by integrating four key components: data-centric modeling, energy-aware optimization, scalable distributed learning, and predictive resource control. As communication networks evolve towards 5G and 6G, managing complexity, energy consumption, and distributed intelligence becomes paramount. This work addresses these challenges through a series of complementary contributions.

At the data layer, strategies are developed for generating high-quality synthetic packet-level data and transforming real-world flow-level traces into structured, machine learning-ready formats. These approaches tackle data scarcity, heterogeneity, and realism, laying a foundation for reliable Artificial Intelligence (AI) model development in network environments.

At the infrastructure layer, the placement of disaggregated 5G functions, Radio Units (RUs), Distributed Units (DUs), Centralized Units (CUs), and User Plane Functions (UPFs) is formulated as a large-scale optimization problem. Decomposition-based and heuristic methods achieve up to 14% improvement in energy efficiency while ensuring Quality of Service (QoS) and responsiveness. The results also highlight limitations of traditional peak-time-based planning strategies.

In the learning layer, AFSL, an asynchronous federated-split learning framework, and its energy-efficient extension, AFSL+, are introduced. These protocols accommodate client heterogeneity, reduce convergence time, and selectively engage participants to minimize energy usage without compromising accuracy or stability.

To further reduce communication overhead, Ada-AFSL is proposed, a dynamic compression strategy for split learning that adapts to real-time bandwidth fluctuations. This method achieves up to 82% data reduction while maintaining model performance and improving generalization in certain scenarios.

Finally, ST-SplitGNN and ST-SplitGNN+ are presented as a spatio-temporal learning and control framework enabling accurate traffic prediction and uncertainty-aware resource allocation. These models support proactive scaling strategies that balance service reliability with sustainability goals.

In essence, this thesis constructs an ecosystem where data generation acts as the soil, optimization as the roots, learning protocols as the growth structure, and predictive control as the flowering system. Together, these elements yield a living, intelligent network architecture that adapts, learns, and sustains itself. The proposed framework advances the vision of AI-enabled, energy-conscious communication systems, offering concrete solutions for scalable, reliable, and green network management in the 5G era and beyond.