Skip to content

RaggedR/categorical-evolution

 
 

Repository files navigation

Composition Determines Diversity

Paper (camera-ready) · Supplementary materials · Talk script

Accepted to GECCO 2026 (AABOH Workshop)

Robin Langer, Claudius Turing, Lyra Vega


The Result

Migration topology explains 23.9–49× more variance in diversity than model or domain choice across evolutionary multi-agent systems. Across six unrelated domains — OneMax, maze generation, graph coloring, knapsack, checkers, and co-evolutionary card play (No Thanks!) — the ordering

none > ring > star > random > fully connected

holds with perfect rank correlation (Kendall's W = 1.0, p = 0.00008). The first Betti number β₁ predicts this diversity ordering perfectly. The full algebraic invariant is the first sheaf cohomology group H¹(G; F).

A spectral bridge connects algebraic connectivity λ₂ to the diversity ordering with a further falsifiable prediction: at n ≥ 7 islands, ring preserves more diversity than star (reversing their n=5 relationship). Confirmed at p < 0.0001.


GECCO 2026 (Start Here)

Camera-ready paper: gecco2026/paper-camera-ready-gecco-v1.pdf

Supplementary materials: gecco2026/supplementary-materials-README.md — index, reproduction instructions, and download link.

Talk script (Lyra's narration, 15 min): gecco2026/gecco-talk-script-lyra.md

Slides: slides/gecco-talk.pdf · slides/gecco-talk.tex

Earlier submitted version (pre-camera-ready): gecco2026/submitted-gecco2026-aaboh/paper-submitted.pdf


Key Figures

Figure File
Topology ordering across all six domains experiments/plots/multi_domain_topology_ordering.pdf
Variance decomposition (topology vs domain) experiments/plots/multi_domain_variance_decomposition.pdf
Coupling onset timing by topology experiments/plots/multi_domain_coupling_onset.pdf
Per-seed diversity fingerprints experiments/plots/fingerprints_panels.pdf

All publication figures live in experiments/plots/ (PNG + PDF).


Experiments

Python experimental suite validating the central claim across six domains.

Guide: experiments/README.md

Quick reproduction of the main result:

cd experiments
pip install pandas matplotlib numpy scipy
python multi_domain_analysis.py

Key scripts:

Script Purpose
*_domain.py Domain sweep implementations (OneMax, Maze, Graph Coloring, Knapsack, No Thanks!, Checkers)
multi_domain_analysis.py Cross-domain topology ordering, Kendall's W, variance decomposition
early_convergence_analysis.py Diversity trajectories, Mann-Whitney tests
coupling_onset_analysis.py Coupling onset timing by topology
plot_fingerprints.py, plot_multi_domain.py Publication figures

Raw CSV data (experiment_e_*.csv) — five topologies × 30 seeds × 100 generations per domain — lives alongside the scripts in experiments/.


Haskell Framework

Categorical framework implementing GA operators as Kleisli morphisms over an MTL effect stack.

Source: haskell/src/Evolution/

cd haskell
cabal build
cabal test
cabal run categorical-evolution -- --demo maze-migration-sweep

Key modules: Category.hs (GeneticOp type), Island.hs (topology-parameterized migration), Effects.hs (EvoM monad stack), Operators.hs (selection, crossover, mutation).


Other Documents

EUMAS 2026 Draft

An expanded journal-length version (12–15 pages LNCS) building on the GECCO result with new experiments (β₁ vs λ₂ two-timescale decomposition, ring vs star at constant β₁, LLM multi-agent sign-flip).

CAIS 2026 Abstract

Short abstract version of the core result.

Grant Materials

XTX AI4Math Fund application for research on categorical foundations for provably correct AI agent orchestration.

Reference and Background


Archive

Earlier drafts and superseded material:


License

MIT

About

Category-theoretic genetic algorithms using MTL effects. Composes GA operators as Kleisli category morphisms. Inspired by cgibbard/category-printf and sjshuck/mtl.

Resources

License

Stars

0 stars

Watchers

0 watching

Forks

Packages

 
 
 

Contributors

Languages

  • Python 41.4%
  • TeX 37.8%
  • Haskell 14.6%
  • BibTeX Style 6.1%
  • Makefile 0.1%