Learning High-Order Feature Interactions in Supervised Learning Classifier Systems Using Code Fragments

Abstract

Learning Classifier Systems (LCSs) are rule-based evolutionary frameworks that combine global search with local learning to produce interpretable predictive models. Supervised LCSs, such as ExSTraCS, have demonstrated strong performance on noisy and high-dimensional datasets; however, their reliance on fixed-length, interval-based rule conditions limits their ability to capture higherorder feature interactions. To address this limitation, we propose Code Fragment-based ExSTraCS (CF-ExSTraCS), a novel supervised LCS framework that integrates genetic programming-based code fragments into the ExSTraCS architecture. Experimental results demonstrate that CF-ExSTraCS substantially outperforms the baseline ExSTraCS system on high-dimensional Boolean tasks and consistently achieves superior or comparable performance across diverse visual feature representations.