TOWARDS ENHANCING RESILIENCE AND ACCURACY OF AI MODELS AND WIRELESS NETWORKS UNDER REAL-WORLD CHALLENGES

Abstract

The rapid evolution of wireless networks toward ultra-dense, user-centric, and AI- and machine learning-based architectures introduces unprecedented challenges in ensuring reliable, efficient, and resilient network operation. Key obstacles towards these networks include (1) frequent occurrences of the outages and inadequacy of traditional methods to handle extensive volume especially in dense deployments and high shadowing scenarios, (2) performance degradation and subsequent adverse affect on optimization methods due to positioning errors in users and remote radio heads, and (3) the lack of resilience of the traditional machine learning based solutions against the scarcity and distribution shift of high-quality data required for effective machine learning modeling. These issues undermine the effectiveness of traditional management and modeling approaches, which are often incapable of handling the complexity, heterogeneity, and real-world uncertainties of next-generation networks. This dissertation systematically addresses these challenges through data-driven frameworks that leverage advanced artificial intelligence and domain-informed machine learning. First, we propose a robust, automated two-tier outage management solution for dense cellular networks. An enhanced XGBoost-based detection model achieves superior accuracy in scenarios with high shadowing and sparse data. At the same time, an actor-critic reinforcement learning compensation strategy ensures fair and efficient service restoration for both outage-affected and already-served users. Next, we tackle the adverse impact of positioning errors in user-centric ultra-dense networks, which can severely degrade area spectral efficiency and energy efficiency by misguiding service zone formation and user association. To mitigate this, we introduce a data-driven optimization and error compensation framework that combines residual learning, automated machine learning, and multi-objective optimization. Our results show that the data-driven optimization and error compensation approach recovers up to $23%$ of performance lost due to localization errors, outperforming baseline methods. Complementary time-series forecasting techniques, including seasonal auto-regressive integrated moving average (SARIMA) and multilayer perceptron regression, further enable dynamic compensation for the degradation of key performance indicators, with SARIMA achieving the highest prediction accuracy. Addressing the pervasive challenge of data scarcity in wireless propagation modeling, we develop a domain-informed generative adversarial network framework. By integrating analytical propagation equations into the training of generative adversarial networks, the proposed method generates high-fidelity synthetic data even under extreme data limitations. Experimental evaluations demonstrate up to 50% improvement in data quality metrics and a 48% reduction in root mean square error for downstream machine learning tasks compared to conventional augmentation techniques, establishing a new benchmark for data-driven modeling in data-scarce scenarios. Finally, to enhance model robustness against distribution shifts and support resilient digital twin applications, we introduce a multistage framework combining conditional tabular generative adversarial network-based data augmentation with attention-through-segmentation training. This approach accelerates the convergence of generative adversarial networks by up to 35%, while attention-through-segmentation reduces the root mean square error by up to 67% and improves generalization across diverse and dynamic wireless environments. Collectively, the methodologies developed in this dissertation advance the state of the art in artificial intelligence-powered wireless network management and modeling. By addressing outage resilience, positioning error compensation, and robust propagation modeling under data scarcity and distribution shifts, this work lays a solid foundation for the next generation of intelligent, self-sustaining wireless networks that can meet the demands and uncertainties of future networks.