Benchmarking GSGP: Still competitive 10 years later?

• Published in: Genetic programming and evolvable machines 2025, Vol.26 (1)
• Main Authors: Muñoz Contreras, Jose Manuel, Trujillo, Leonardo, Hernandez, Daniel E., Cardenas Florido, Luis A.
• Format: Article
• Language: English
• Subjects: Artificial Intelligence; Benchmarks; Biomedical Engineering and Bioengineering; Compilers; Computer Science; Electrical Engineering; Genetic algorithms; Interpreters; Programming Languages; Programming Techniques; Semantics; Software Engineering/Programming and Operating Systems
• ISSN: 1389-2576; 1573-7632
• DOI: 10.1007/s10710-024-09504-3

Geometric Semantic Genetic Programming (GSGP) reimagined how to search for symbolic models using an evolutionary process. It addressed one of the main weaknesses of standard Genetic Programming (GP) by performing the search directly within the semantic space of a problem, which can define a convex and unimodal fitness landscape. Since the method does not require the syntactic evaluation of offspring, GSGP allowed more efficient and effective learning systems compared to standard GP. However, recent benchmarking results have suggested that GSGP is no longer a competitive approach, particularly in the symbolic regression domain, despite its previous success in several real-world tasks. Therefore, the research question of this work is an empirical one, stated as: Is GSGP still a competitive symbolic regression method 10 years after it was proposed? A comprehensive benchmark of black-box problems and comparisons with state-of-the-art methods were used to answer this question. In particular, a recently developed parallel version of GSGP is used, extending the implementation to also include the previously proposed optimal mutation step computation, as well as using the analytical quotient operator instead of a protected division. Results show that with simple, but important, extensions to the original GSGP algorithm, the answer to the research question is yes .