Machine learning optimization for dynamic spectrum awareness

Abstract

This dissertation proposed a machine learning-based approach focused on improving dynamic spectrum awareness in wireless communications. The approach is comprised of four main components: network optimization with genetic algorithm convolutional neural networks (GACNN), which focuses on optimizing neural network architectures for specific tasks to reduce the complexity and cost of the network optimization process; unsupervised prototype learning, which uses hierarchical clustering and a prototype-based learning objective to estimate signal-to-noise ratio (SNR) regions and perform modulation classification to improve the classification accuracy and the identification of new signal classes; deep neural network explainable AI (DNN XAI), which increases the transparency and interpretability of machine learning models in deep neural networks to ensure compliance with spectrum regulatory standards; and lastly, incremental learning class representation drift, which evaluates the performance of incremental learning methods in baseband modulation classification to establish a continuous learning process that adapts to dynamic environments. By addressing gaps in current spectrum management techniques, these components can improve spectrum utilization, increase machine learning-based communication interpretability, and provide an informed model for future spectrum management strategies