Agentic Parent Selection

Research code for agentic parent selection in Genetic Programming: an LLM/agent framework generates parent-selection algorithms (selector.py) that are plugged into a shared GP engine and evaluated for symbolic regression, then compared against hand-written baselines (tournament and epsilon-lexicase selection).

The repository is organized into seven top-level directories, each with its own README.md giving the full details. This file is a map of the whole project.

Directory overview

Directory	Purpose
AGENTS/	The agentic-gp framework — the LLM/RAG agent that generates the parent-selection algorithms. Includes setup (Conda, Ollama/Azure), configuration, RAG documents, and the CLI (main.py) for running the recommender.
GP/	The tree-based Genetic Programming engine (Source/gp.py, DEAP + Ray) that evolves symbolic-regression models with a runtime-pluggable selector, plus the two hardcoded baseline selectors (tournament.py, epsilon_lexicase.py).
Experiments/	Self-contained HPC (SLURM) job bundles and precomputed splits for the two experiments — Experiment 1: Ablation and Experiment 2: Benchmark Evaluation — that drive GP/Source/gp.py. Distributed as archives that must be unzipped.
Data-Tools/	Dataset generation (Pull_Data/) and the analysis pipeline: Python scripts that aggregate raw results into .csv/.txt, and R Markdown reports that render the study's statistics and figures.
RESULTS/	Raw per-replicate experiment outputs (performance MSE, generated algorithms) and the aggregated, analysis-ready data. Available only after unzipping (see below).
UCI-DATA/	The final, polished datasets for the 6 regression problems (one data.csv each). Available only after unzipping (see below).
SUPPLEMENTARY_MATERIAL/	The paper's rendered supplementary PDFs: the exact prompts used in each condition, the Experiment 1 ablation reports (classification, execution, and per-model performance), and the Experiment 2 benchmark performance report. Each is a knit of an analysis report under Data-Tools/.

Compressed directories

Two directories are shipped as zip archives at the repository root and are not present until extracted:

unzip uci_data.zip   # -> UCI-DATA/
unzip results.zip    # -> RESULTS/

The Experiments/ bundles are likewise distributed as archives; see Experiments/README.md for the unzip steps specific to each experiment.

The rendered SUPPLEMENTARY_MATERIAL/ PDFs are also packaged for convenience as supplementary_material.zip at the repository root, which extracts to the same SUPPLEMENTARY_MATERIAL/ folder:

unzip supplementary_material.zip   # -> SUPPLEMENTARY_MATERIAL/

How the pieces fit together

1. Datasets are built in Data-Tools/Pull_Data/ and distributed as UCI-DATA/ (one data.csv per problem).
2. The agent in AGENTS/ generates parent-selection algorithms (selector.py) under different LLM/RAG configurations.
3. The experiments in Experiments/ split the datasets and submit SLURM jobs that run the GP engine in GP/, plugging in either an agent-generated selector or a hardcoded baseline.
4. Runs write raw outputs to RESULTS/, which Data-Tools/ aggregates and analyzes into the study's tables and figures.

See each directory's README.md for specifics.