Agentic AI Automates Parent Selection in Genetic Programming, Matching Hand-Coded Algorithms

A new paper from researchers at Cedars-Sinai (Jose Guadalupe Hernandez, Jui-Hsuan Chang, Anil Kumar Saini, Xi Li, Jason H. Moore) explores whether agentic AI—LLM-based agents augmented with retrieval-augmented generation (RAG)—can automate the design of genetic programming systems. Specifically, they target parent selection, a core evolutionary component that dictates which individuals reproduce. Using symbolic regression as a testbed, they ran ablation studies across four LLM types to isolate the effects of agentic reasoning and retrieval on algorithm generation, validity, implementation similarity, and downstream fitness.

The results show that both reasoning and retrieval substantially shape the types of algorithms generated, but downstream performance remains heavily dependent on the underlying LLM. The strongest configuration, a full agentic setup dubbed "5 mini–AR," consistently produced established ε-lexicase selection implementations while maintaining competitive accuracy. Benchmarked against fixed tournament selection and semi-dynamic MAD ε-lexicase across six symbolic regression problems, 5 mini–AR performed on par with ε-lexicase and generally beat tournament selection. This work demonstrates that agentic AI can translate domain knowledge into executable evolutionary components, advancing automated configuration and design of genetic programming systems.