Homotopy-Based Multi-Objective Optimization for Enhanced Machine Learning in Black-Box Environments

This dissertation presents a unified framework for optimizing machine learning models in black-box environments where accuracy, robustness, fairness, and adaptability must be jointly addressed. Building on homotopy continuation methods and surrogate modeling, the work develops stable and efficient hyperparameter optimization strategies, then extends them to multi-objective settings that balance competing performance criteria. Applications span multimodal automatic target recognition, ecological image segmentation, and fair face parsing for generative AI. Adaptability is achieved through self-supervised vision transformers for multi-sensor data and adaptive autonomy mechanisms for human-on-the-loop robotics. Complementary interpretability tools, such as wavelet-based analyses of vision transformers, enhance transparency into model decision-making. Together, these contributions establish foundations, fairness-aware optimization, and adaptability strategies within a coherent methodology, enabling the design of robust, trustworthy vision systems with demonstrated utility in defense, environmental monitoring, and ethical AI domains.