Neutrality, Simplicity and Search: Lessons from Genetic Programming

Created by W.Langdon from gp-bibliography.bib Revision:1.9078

@InProceedings{Banzhaf:2026:GECCO,

• author = "Wolfgang Banzhaf",
• title = "Neutrality, Simplicity and Search: Lessons from Genetic Programming",
• booktitle = "Proceedings of the 2026 Genetic and Evolutionary Computation Conference",
• year = "2026",
• editor = "Nelishia Pillay and Gabriel Kronbergerr and Gabriel Kronberger and Leonardo Trujillo and Ting Hu and Sebastian {Rojas Gonzalez} and Saul {Calderon Ramirez} and Stephan Winkler and Yazmin Maldonado and James McDermott and Jose Manuel {Munoz Contreras} and Giorgia Nadizar and Marcella {Scoczynski Ribeiro Martins} and Hirad Assimi and Mario Andres Munoz and Oliver Cuate and Daniel Hernandez and Yanan Sun and Edgar Galvan and Aniko Ekart and Nadarajen Veerapen",
• pages = "1",
• address = "San Jose, Costa Rica",
• series = "GECCO '26",
• month = "13-17 " # jul,
• organisation = "SIGEVO",
• publisher = "Association for Computing Machinery",
• publisher_address = "New York, NY, USA",
• note = "Invited ACM SIGEVO Keynote",
• keywords = "genetic algorithms, genetic programming, neutrality, simplicity bias, genotype-phenotype maps, evolutionary computation, Speaker Bio",
• isbn13 = "9798400724879",
• URL = "https://gecco-2026.sigevo.org/Keynotes#Neutrality_Simplicity_and_Search:_Lessons_from_Genetic_Programming",
• DOI = "10.1145/3795095.3822498",
• size = "1 page",
• abstract = "The role of neutrality, a phenomenon where many genotypes map to one phenotype, has long been considered in evolutionary computation a (sometimes annoying) side effect rather than a principle. I argue that it is, in fact, a central force in search over redundant genotype-phenotype maps. Its main consequence is a bias toward simple solutions. Using linear genetic programming on Boolean functions, where complexity can be measured directly, we show that simpler phenotypes occupy exponentially larger neutral sets, and that search is therefore drawn toward them. It turns out that this bias is a property of the map itself, not of evolution as such. Notably, it appears in other evolutionary systems with redundant genotype-phenotype maps, such as RNA, and even in non-evolutionary settings like deep neural networks and other redundant, complexity-bounded representations. It is absent, however, in the one-to-one representations often used in optimization. We then suggest that hierarchical systems built from modules may be understood as realizing neutrality at a higher level, because composing phenotypes from reusable modules enlarges their neutral sets combinatorially. This connects the classical 1960s argument of Herbert Simon about the architecture of complexity with more recent observations about depth in modern machine learning. From this follow lessons for the design of fitness functions and representations, as well as a broader question, namely whether genetic programming can serve as a model system for understanding simplicity bias in machine learning at large.",
• notes = "GECCO-2026 A Recombination of the 35th International Conference on Genetic Algorithms (ICGA) and the 31st Annual Genetic Programming Conference (GP)",

}

Genetic Programming entries for Wolfgang Banzhaf

Citations