Selective rationalization via genetic-based learning: extending the GenSPP architecture with the CMA-ES algorithm

Deep learning models often lack transparency, preventing their adoption in high-stakes decision-making domains. The Gen-SPP framework addresses this by generating sparse, human-readable rationales for classification tasks using a Genetic Algorithm (GA) for feature selection. However, GAs suffer from poor modeling of parameter correlations and require extensive hyperparameter tuning. This thesis proposes extending the Gen-SPP by replacing the GAs with the Covariance Matrix Adaptation Evolution Strategy (CMA-ES), a more robust, distribution-based optimization algorithm. CMA-ES adaptively models parameter dependencies during search, enabling a more effective exploration of the search space. Our experiments on benchmark datasets demonstrate that CMA-ES can improve both classification accuracy and rationale fidelity compared to the GA baseline. Although CMA-ES demands higher computational memory, it can offer a stronger optimization strategy for explainable AI systems.