Geno-Synthetic Algorithm — experiment battery

Reference implementation and reproducible experiment battery for the Geno-Synthetic Algorithm (GSA), a type-factored coevolutionary optimization method that evolves heterogeneous genotypes (integer, real, Boolean, categorical, complex, embedding) in separate typed subpopulations using type-native operators, then assembles them into phenotypes for joint fitness evaluation.

This repository contains the code, raw run logs, and analysis reports used to produce the empirical results in the GSA paper.

What's here

• Core GSA implementation — typed subpopulations, type-native operators, active/passive phenotype assembly, three credit-assignment schemes (direct, elite, ensemble), and an asynchronous variant.
• Eight registered GSA variants for ablation: GSA_FULL_ENSEMBLE, GSA_DIRECT, GSA_ELITE_CONTEXT, GSA_NO_DIVERSITY, GSA_GENERIC_OPERATORS, GSA_NO_ASSEMBLY, GSA_ASYNC, GSA_ASYNC_DIRECT.
• Flattened baselines — DE, EA, and Mixed-Variable GA over a single decoded representation, for like-for-like comparison.
• Benchmarks — typed synthetic suite (additive, epistatic, deceptive, noisy, mix), TypedGated (active-assembly stress test), an IOH Boolean bridge, and a COCO BBOB-MixInt adapter (bbob-mixint).
• Experiment battery — programmatic run matrices in scripts/, parquet logs in results/raw/, and rendered reports + figures in results/reports/.
• Analysis — Wilcoxon paired tests, Vargha–Delaney Â12, Friedman with Holm correction, rank summaries, and paper-grade plots.

Quickstart

python -m venv .venv
# Windows
.venv/Scripts/activate
# Unix
# source .venv/bin/activate

pip install -e .[test]

# Sanity check
python -c "import gsa; print(gsa.__version__)"  # -> 0.1.0
pytest                                           # 164 tests, all green

Optional extras for the BBOB-MixInt suite and the IOH Boolean bridge:

pip install coco-experiment ioh

Reproducing the paper results

Each script writes raw runs to results/raw/<matrix>/runs.parquet and a matching report script renders markdown + figures into results/reports/<matrix>/. All seeds are deterministic.

Matrix	Run	Report
Paper headline (6 cells × 5 seeds)	python scripts/run_paper_experiments.py	python scripts/generate_paper_report.py
Async vs sync sub-matrix (80 runs)	python scripts/run_async_experiments.py	python scripts/generate_async_report.py
Larger-budget sweep (225 runs, {5k, 15k, 30k})	python scripts/run_budget_experiments.py	python scripts/generate_budget_report.py
BBOB-MixInt @ 5k (360 runs)	python scripts/run_bbob_mixint_experiments.py --budget 5000 --output-dir results/raw/bbob_mixint	python scripts/generate_bbob_mixint_report.py --input results/raw/bbob_mixint/runs.parquet --out-dir results/reports/bbob_mixint
BBOB-MixInt @ 100k (360 runs)	python scripts/run_bbob_mixint_experiments.py --budget 100000 --output-dir results/raw/bbob_mixint_100k	python scripts/generate_bbob_mixint_report.py --input results/raw/bbob_mixint_100k/runs.parquet --out-dir results/reports/bbob_mixint_100k --budget 100000
BBOB-MixInt 5k-vs-100k comparison	—	python scripts/compare_bbob_mixint_budgets.py + python scripts/generate_crossover_figure.py
TypedGated H3 ablation, strengthened (240 runs)	python scripts/run_gating_strengthened.py	python scripts/generate_gating_strong_report.py

Pre-rendered reports for each matrix are committed under results/reports/ so the figures can be browsed on GitHub without re-running anything.

Headline findings (calibrated)

• Architectural reach is GSA's clearest win: it is the only method that runs on Cx/E gene families; flattened baselines crash by encoder ValueError, deterministically.
• Active assembly matters when the optimum sits inside the gating region (TypedGated + Cx, n=20, D ∈ {20, 40, 80}): active wins on 51/60 paired-by-seed comparisons (85%), pooled Wilcoxon p = 2.9 × 10⁻⁷.
• Type-native operators matter: GSA_GENERIC_OPERATORS is consistently the worst variant.
• Credit assignment: GSA_ELITE_CONTEXT significantly outperforms GSA_FULL_ENSEMBLE (p = 5.4 × 10⁻⁷, Â12 = 0.0). Default to elite credit, not ensemble.
• Asymptotic crossover on an external suite (BBOB-MixInt): at 5k evals FLATTENED_DE wins (mean rank 2.17); at 100k evals GSA_DIRECT is statistically indistinguishable from FLATTENED_DE (Â12 = 0.499, p = 0.61) and FLATTENED_EA collapses from rank 2.67 to 4.25.
• Honest negative result: at small budgets on smooth multi-family synthetic problems, flattened DE wins — the per-iteration K-ensemble cost burns budget faster than the typed-operator advantage compounds.

See results/reports/<matrix>/report.md for full per-cell statistics.

Layout

src/gsa/
  core/         # typed subgenomes, operators, assemblers, credit schemes
  baselines/    # flattened DE / EA / Mixed-Variable GA
  benchmarks/   # typed synthetic suite, TypedGated, COCO BBOB-MixInt, IOH
  experiments/  # run specification, runner, evaluation budget
  analysis/     # statistics, aggregation, plots, tables
configs/        # experiment YAML configs
scripts/        # run_*.py and generate_*.py entry points
tests/          # 164 pytest tests
results/
  raw/<matrix>/runs.parquet
  reports/<matrix>/{report.md, figures/, tables/, *.csv}
docs/           # design specs, planning notes

Citing

If you use this code, please cite the accompanying paper. Citation metadata is available in CITATION.cff; the arXiv identifier is 2605.13365.

License

MIT — see LICENSE.