5G/6G

This work proposes a deep neural network-based framework for jointly optimizing RIS placement, power allocation, and reflective element assignment in double-active RIS-assisted systems. Existing approaches, such as Block Coordinate Descent (BCD), while mathematically sound, often yield suboptimal and computationally intensive solutions. To address these limitations, eight neural models are developed, each differing in constraint-handling and output parameterization. A synthetic dataset of 200,000 samples is used for training and evaluation. Results show that constraint-aware models, specifically the softmax-constrained hinge (SCH) and sigmoid-normalized hinge (SNH), achieve both 100% feasibility and consistent improvements in achievable downlink rate. These gains are statistically validated using Wilcoxon signed-rank tests. Furthermore, the models generalize well across urban and rural settings and remain effective with shallow architectures, supporting their practical use in dynamic wireless environments. This work presents a scalable, data-driven solution for RIS optimization, laying the groundwork for intelligent, real-time configuration in future 6G wireless systems.

Effective resource allocation in O-RANs is crucial for ensuring quality of service, minimizing operational costs, and maintaining network reliability. While mathematical programming can provide optimal solutions, it may not be computationally feasible for large-scale and dynamic scenarios. Recent approaches using deep reinforcement learning (DRL) have shown promise by facilitating adaptive and autonomous decision-making; however, these methods often face challenges such as slow convergence and poor generalization to unfamiliar scenarios. This work proposes a novel LLM-driven curriculum-based DRL framework for resource allocation in O-RANs to address these issues. The proposed solution integrates curriculum learning into the DRL training process to enhance convergence and stability. Additionally, it introduces LLMs to automatically generate adaptive curricula based on the agent’s performance and the network environment, thereby replacing manual curriculum design and improving learning efficiency and generalization.