LASIGE integrated researcher Sara Silva, together with researchers from NOVA IMS, Ines Magessi and Leonardo Vanneschi, developed two new Genetic Programming selection methods that prevent overfitting by minimising both the training error and the functional complexity of the evolved models in a pseudo-multiobjective fashion.

Both methods rely on an improved version of the Slope-Based Complexity (iSBC) that approximates the mathematical curvature of any model in linear time. With the iSBC, evolution produces models that generalise better, use fewer features and simplify to smaller sizes. Since iSBC measures the complexity per feature along its entire domain, it can reveal not only how much but also where the output of a model is affected by changes in its input features, akin to using Partial Dependence Plots and SHAP values. The approach supports the development of transparent, interpretable, and reliable symbolic models, particularly valuable in domains where understanding model behaviour is crucial.

This work was published in the top 10% journal IEEE Transactions on Evolutionary Computation. The paper, titled “Controlling Functional Complexity for Overfitting Reduction and Improved Interpretability in GP”  is availabe here.