Result Details
Towards ML-Based Position-Aware Mutation Operator in Cartesian Genetic Programming
Drahošová Michaela, Ing., Ph.D., DCSY (FIT)
Mrázek Vojtěch, doc. Ing., Ph.D., DCSY (FIT)
Cartesian Genetic Programming (CGP) relies primarily on point mutation, which is uninformed and treats all genes uniformly. Combined with positional bias -- nodes near inputs being substantially more likely to be active -- most mutations target irrelevant parts of the chromosome. Existing approaches address this by restructuring the genotype; in this paper, the mutation operator itself is made position-aware. The Mean number of Active Output Connections (MAOC), a theoretical metric quantifying positional bias from CGP parameters alone, is derived and validated against an archive of over 11 million mutation records from CGP runs on the PMLB dataset. Three operators of increasing complexity are proposed: probability-based, Graph Neural Network (GNN)-based, and Large Language Model (LLM)-based. Evaluation on 38 symbolic regression tasks shows all three achieve fitness comparable to baseline while accelerating convergence in the early generations of evolutionary search. The probability-based operator reduces runtime by approximately 32 % compared to baseline point mutation, while the GNN-based operator adds runtime overhead of approximately only 21 %
Cartesian genetic programming; Mutation operator; Positional bias; Graph neural network; Large language model
@inproceedings{BUT212057,
author="Martin {Hurta} and Michaela {Drahošová} and Vojtěch {Mrázek}",
title="Towards ML-Based Position-Aware Mutation Operator in Cartesian Genetic Programming",
booktitle="Parallel Problem Solving from Nature – PPSN XIX
19th International Conference, PPSN 2026, Trento, Italy, August 29 – September 2, 2026, Proceedings, Part IV",
year="2026",
journal="Lecture Notes in Computer Science",
volume="16988",
pages="17",
publisher="Springer Nature Switzerland",
address="Cham",
doi="10.1007/978-3-032-36229-2\{_}17",
isbn="978-3-032-36229-2",
issn="0302-9743",
url="https://link.springer.com/chapter/10.1007/978-3-032-36229-2_17"
}