As machine learning systems are used more and more in important areas such as healthcare, finance, and public policy, the demand for transparency and trustworthiness has grown significantly. However, many state-of-the-art models, particularly deep learning models, are often treated as "black boxes," offering limited insight into how predictions are made. This lack of explainability raises concerns around accountability, fairness, and trust. To address this challenge, this thesis explores the use of Genetic Algorithms (GAs) as a powerful tool for building interpretable models and generating explanations for AI systems.
Genetic algorithms, inspired by natural evolution, are well-suited for exploring complex search spaces where multiple objectives—such as accuracy, simplicity, and transparency —must be balanced. In this work, GAs will be investigated as a method to evolve human-understandable models, such as symbolic expressions, rule sets, or surrogate explanations for black-box classifiers.
The goal is to explore how evolutionary search can be used not just for optimization but for generating explainable results that help users understand AI decisions. This work will provide both theoretical and practical insights into the intersection of evolutionary computation and explainable AI. This thesis will test genetic algorithm-based explanations on standard datasets and compare them to popular methods like SHAP and LIME, to see how understandable, accurate, and practical they really are. Our broader aim is to show how evolutionary methods can mediate between raw model performance and meanings that are graspable in lived practice.
If you want to learn more or get involved, please contact Pooya Zakeri .