Explainable AI with Evolutionary Computation

Overview

When an AI model decides something about a person, the person should be able to see how the decision was made. Most modern machine learning models make that impossible: their reasoning is spread across millions of weights that nobody can read. Genetic programming (GP) takes a different route. It evolves models as explicit symbolic expressions, so the model is something a human can potentially interpret and understand. This work uses GP and related evolutionary and swarm methods to build explainable AI models.

Themes

• Counterfactual and local explanations. Hayden Andersen's PhD is evolving human-friendly explanations. We use particle swarm optimisation and differential evolution to find counterfactuals: the smallest realistic changes that would have flipped a prediction. Niching lets a single run return a diverse set of alternatives rather than one answer, so a person can choose the explanation that is most relevant to them. More recently, we have evolved small GP models as local explanations of a black-box model's behaviour around a single decision, giving an interpretable substitute for methods such as LIME.
• Large language models as explainers. An evolved GP tree is interpretable in principle, but a page of nested functions is not explainable to laypeople. Paula Maddigan tested whether a large language model can turn an evolved tree into an accurate plain-language explanation for a non-expert.
• Explainable dimensionality reduction. Nonlinear dimensionality reduction methods are crucial for simplifying and processing complex datasets. But leading methods such as t-SNE and UMAP reduce data in an opaque way, where the new, lower-dimensional space cannot be traced back to the original features. GP can learn the mapping as an explicit formula instead. Ben Cravens's work on GP for explainable manifold learning is the latest in a line that began with my PhD, and the embeddings remain competitive with standard non-linear methods while staying readable.
• Making other models interpretable. Finn Schofield replaced the encoder of an autoencoder with GP, keeping the reconstruction quality that deep learning provides while making the compressed representation interpretable. Our other related work has proposed a method to cluster features so that learning classifier systems search more effectively and remain understandable.

Team and collaborators

Selected outputs

• Cravens, B., Lensen, A., Maddigan, P., & Xue, B. (2026). Genetic Programming for Explainable Manifold Learning. IEEE Transactions on Emerging Topics in Computational Intelligence. PDF
• Andersen, H., Lensen, A., Browne, W. N., & Mei, Y. (2024). Interpretable Local Explanations Through Genetic Programming. GECCO. PDF
• Maddigan, P., Lensen, A., & Xue, B. (2024). Explaining Genetic Programming Trees using Large Language Models. PDF
• Andersen, H., Lensen, A., Browne, W. N., & Mei, Y. (2023). Producing Diverse Rashomon Sets of Counterfactual Explanations with Niching Particle Swarm Optimization Algorithms. GECCO. PDF
• Mei, Y., Chen, Q., Lensen, A., Xue, B., & Zhang, M. (2023). Explainable Artificial Intelligence by Genetic Programming: A Survey. IEEE Transactions on Evolutionary Computation. PDF
• Schofield, F., Slyfield, L., & Lensen, A. (2023). A Genetic Programming Encoder for Increasing Autoencoder Interpretability. EuroGP. PDF
• Andersen, H., Lensen, A., Browne, W. N., & Mei, Y. (2022). Evolving Counterfactual Explanations with Particle Swarm Optimization and Differential Evolution. IEEE CEC. PDF
• Lensen, A., Xue, B., & Zhang, M. (2022). Genetic Programming for Manifold Learning: Preserving Local Topology. IEEE Transactions on Evolutionary Computation. PDF
• Lensen, A., Xue, B., & Zhang, M. (2021). Genetic Programming for Evolving a Front of Interpretable Models for Data Visualization. IEEE Transactions on Cybernetics. PDF

See the publications page for the full list.