Explainable ai and optimization theory based wireless radio resource management

The International Telecommunication Union's IMT-2030 framework identifies ``Integrated artificial intelligence (AI) and Communication'' as one of the core directions for 6G networks. This shift toward AI-driven solutions necessitates greater transparency in decision-making processes, where explainability is pivotal in establishing trust in AI systems, allowing network operators and engineers to understand, validate, and troubleshoot the decisions made by deep learning (DL) models. Additionally, robustness against out-of-distribution inputs and adversarial attacks is crucial to ensure reliable performance in diverse and evolving environments. These two factors—explainability and robustness—are essential to fulfilling the broader goals of AI-native 6G networks, where AI is not just an enhancement but a foundational component integrated into communication systems. This thesis examines the application of Explainable AI (XAI) for robust and transparent radio resource management (RRM) in AI-empowered 6G networks. First, the significance of this research is presented by unveiling the merits of explainability and robustness for 6G RRM, outlining a range of core explainability and robustness techniques for efficient RRM. Then, the practical implementation of these XAI techniques is analyzed, including a quantitative evaluation of their impact on key performance indicators (KPIs) in specific 6G scenarios. Two case studies are presented, showcasing their application in model simplification and enhancing RRM robustness. The first study proposes an explainable and robust DL-based framework for beam management in a millimeter-wave (mmWave) multiple-input multiple-output (MIMO) communications system. To minimize signaling overhead for beam alignment, a novel model-agnostic, feature relevance-oriented XAI framework is designed to rank and prioritize input features, reducing input size and beam sweeping time. To incorporate transparency and resilience into the DL-based beam alignment engine (BAE), a defense mechanism against adversarial inputs is developed, examining internal representations learned by deep neural networks (DNN) to provide interpretability and robustness against malicious/out-of-distribution inputs. The second case study focuses on deep reinforcement learning (DRL)-based RRM in vehicular networks for joint transmit power and spectrum allocation. First, a model-agnostic XAI-based methodology is devised to explain the inference process of multiple trained DRL agents. Then, a novel XAI-based feature importance ranking algorithm using Shapley additive explanations (SHAP) is developed to simplify the DRL agents' input state size, followed by an automated feature selection strategy to reduce model complexity. By eliminating low-importance features, the methodology produces a simplified model with fewer network parameters and lower training time while maintaining reasonable performance. This thesis further introduces a novel DRL-based framework for joint power control and block length allocation to minimize the worst-case decoding error probability for ultra-reliable and low-latency communication (URLLC)-based vehicular networks. Initially, an algorithm grounded in optimization theory is developed based on deriving the joint convexity of the decoding error probability in the block length and transmit power variables within the region of interest. Subsequently, inspired by event-triggered control systems, an efficient event-triggered DRL-based algorithm is proposed to solve the joint optimization problem. Incorporating event-triggered learning into the DRL framework enables assessing whether to initiate the DRL process, thereby reducing the number of DRL process executions while maintaining reasonable reliability performance. The results demonstrate a noteworthy reduction in training and runtime complexity compared to the conventional optimization-based solutions.