Dynamic Resource Allocation for Wireless Networks and Radar Systems via Deep Reinforcement Learning

The rapid advancement of wireless communication technologies and the proliferation of smart devices have led to increasingly complex and dynamic network environments. These developments have posed significant challenges to traditional radio resource management (RRM) techniques, which often struggle with scalability, adaptability, and real-time decision-making. In response to these limitations, this dissertation explores the application of advanced machine learning, particularly deep reinforcement learning (DRL), to develop intelligent, adaptive, and data-efficient solutions for resource management in wireless networks and radar systems. We begin by addressing joint channel access and power control in wireless interference networks using centralized, distributed, and federated multi-agent DRL frameworks. These methods demonstrate strong adaptability and performance with limited information exchange. To tackle the data inefficiency and generalization limitations of conventional DRL, we introduce meta-reinforcement learning strategies using Model-Agnostic Meta-Learning (MAML) for dynamic channel access, power control, and UAV trajectory design. These frameworks enable rapid adaptation to new environments with minimal retraining, making them well-suited for real-world deployments. Expanding into radar systems, we investigate adaptive radar resource management for multi-target tracking. We develop a constrained deep reinforcement learning (CDRL) approach that optimally allocates time under budget constraints. To address data scarcity in radar environments, we propose hybrid learning frameworks combining offline and online CDRL. This work is extended to track-while-scan radar systems and integrated sensing and communication (ISAC) platforms, highlighting the trade-offs between sensing and communication tasks. Multi-objective reinforcement learning techniques, implemented with soft actor-critic (SAC), are also employed to find Pareto-optimal policies in radar scanning and tracking tasks. Finally, recognizing the importance of interpretability in AI systems, we introduce a novel explainable artificial intelligence (XAI) framework—Deep learning assisted local interpretable modal-agnostic explanations (DL-LIME)—which enhances the traditional LIME method by incorporating deep learning into the sampling process. This approach improves both fidelity and task performance, providing greater transparency in neural network decision-making. Overall, this dissertation contributes a suite of innovative learning-based frameworks for efficient, robust, and interpretable resource management across wireless communication and radar systems. The proposed methodologies pave the way for the development of next-generation intelligent networks, capable of operating effectively in highly dynamic and data-constrained environments.