From Six to Two: SLIM-DUO, a Simplified Extension of SLIM-GSGP

Geometric Semantic Genetic Programming (GSGP) is an extension of Genetic Programming (GP) that captured the interest of researchers because of its ability to induce a unimodal error surface for any supervised learning problem. Although still a recent development, the Semantic Learning with Inflate and Deflate Mutations (SLIM-GSGP) extension of GSGP has already attracted significant attention due to its novel ability to generate offspring that are smaller than their parents, effectively addressing the problem of steady model growth in GSGP. This paper presents SLIM-DUO, an extension of SLIM-GSGP that integrates the six existing SLIMGSGP variants into solely two unified formulations: DUO-MUL and DUO-SUM. This integration streamlines benchmarking and hyperparameter exploration while aiming to retain comparable predictive performance at approximately one third of the computational cost. Across five test problems, SLIM-DUO achieves predictive performance comparable to that of SLIM-GSGP, with model size differences that remain within previously observed dataset-dependent variability among SLIM variants. Overall, SLIM-DUO substantially reduces the computational effort required during both the configuration and benchmarking phases, highlighting a favorable balance between model size and computational complexity and preserving solution quality.