The evolution toward Sixth-Generation (6G) networks marks a transition from human-operated infrastructure to AI-native systems, where Artificial Intelligence and Machine Learning (AI/ML) constitute the operational core. Zero-Touch Service Management (ZSM) enables this vision through autonomous closed loops that translate high-level intents into network actions with minimal human intervention. Standardization by the 3rd Generation Partnership Project (3GPP) through the Network Data Analytics Function (NWDAF) and by the European Telecommunications Standards Institute (ETSI) positions AI/ML models as first-class network functions. This thesis argues that realizing ZSM depends on AI/ML techniques that are autonomous, resilient, explainable, and compliant.
Deploying AI/ML on the network control path exposes challenges hidden by offline evaluation. Existing NWDAF implementations lack integrated Machine Learning Operations (MLOps) and intent-driven orchestration, while live networks suffer from data, label, and concept drift, catastrophic forgetting during adaptation, and degradation of Deep Reinforcement Learning (DRL) policies. Furthermore, autonomous AI must satisfy the transparency and accountability requirements imposed by the European AI Act, the General Data Protection Regulation (GDPR), and emerging 3GPP standards.
To address these challenges, this thesis develops seven complementary techniques organized into three pillars. The first pillar focuses on autonomous AI/ML operations. It introduces the first comprehensive open-source NWDAF aligned with 3GPP Release 17, implementing the separation of the Analytics Logical Function (AnLF) and Model Training Logical Function (MTLF) while integrating an MLOps-enabled orchestrator for autonomous model lifecycle management. It also proposes Agentic-NWDAF, an intent-aware orchestration framework that combines Agentic AI with the Model Context Protocol (MCP) to translate natural language intents into executable NWDAF service calls, achieving high accuracy over one hundred 3GPP-derived intents.
The second pillar ensures long-term model reliability. It proposes a Dual Self-Attention AutoEncoder that jointly detects data, label, and concept drift, improving detection accuracy by more than 13.6% while reducing latency by 94.7%. Continuous adaptation is achieved through Long Short-Term Memory with Elastic Weight Consolidation (LSTM-EWC), improving knowledge retention by up to 51% compared with catastrophic forgetting baselines. The pillar further introduces a Zero-Touch Drift Management framework whose Composite Drift Index (CDI) combines Key Performance Indicators (KPIs), policy statistics, and epistemic uncertainty from Evidential Deep Learning (EDL), improving drift detection by up to 26.0% across four 3GPP-relevant control problems.
The third pillar governs autonomous decision making. It proposes a Composable eXplainable Reinforcement Learning (XRL) framework that leverages Large Language Models (LLMs) to generate profile-aware explanations of DRL decisions for technical and non-technical stakeholders. It also introduces EUROCOMPLY, an Agentic AI framework that combines vector and graph retrieval to automatically audit datasets and AI/ML pipelines against the European AI Act, GDPR, and 3GPP standards, validated across twenty telecom use cases and four realistic datasets.
Together, these contributions form a unified zero-touch AI/ML architecture in which autonomous model deployment, continuous drift management, explainability, and regulatory compliance operate as an integrated closed loop. By aligning with 3GPP and ETSI specifications while combining uncertainty-aware learning, attention mechanisms, and LLM-based reasoning, this thesis advances AI/ML that is deployable as the operational core of self-managing 6G networks and provides a concrete step toward autonomous, resilient, explainable, and compliant Zero-Touch Service Management.