MorphoNAS: Embryogenic Neural Architecture Search Through Morphogen-Guided Development

This repository contains the reference implementation and experiment scripts for the paper:

MorphoNAS: Embryogenic Neural Architecture Search Through Morphogen-Guided Development, submitted for peer review. Preprint: arXiv:2507.13785v1.

Authors: Mykola Glybovets, Sergii Medvid

Overview

MorphoNAS is a morphogenetic neural architecture search (NAS) system. Compact genomes encode morphogen dynamics and threshold-based developmental rules that deterministically grow a neural network from a single progenitor cell via local chemical interactions. We evaluate MorphoNAS in two domains:

• Structural targeting: evolving genomes that develop into predefined random graph configurations (8-31 nodes)
• Functional performance: solving CartPole with compact evolved RNN controllers (6-7 neurons under size pressure)

The codebase includes experiment pipelines, analysis scripts, and utilities to reproduce the results reported in the preprint.

Repository structure

• src/
  • experimentA.py: structural targeting experiments (graph properties)
  • experimentB.py: RNN controller experiments (CartPole)
  • regenerate_displays.py: regenerate visualizations for experiment outputs
  • Additional modules: genome, evolution, morphogen and neural propagation, displays, etc.
• experiments/
  • _ExpA_graph_properties/: configs and results for Experiment A
  • _ExpB_RNN_controller/: configs and results for Experiment B
• experimentA.sh, experimentB.sh: end-to-end scripts for setup, running, and post-processing
• requirements.txt: Python dependencies

Quick start

1. Create a virtual environment and install dependencies

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

1. Run experiments via helper scripts

# Experiment A: structural targeting
bash experimentA.sh

# Experiment B: RNN controller (CartPole)
bash experimentB.sh

Each script will set up the environment, run the experiments, and regenerate displays.

Running components manually

You can also run components directly:

# Experiment A (generate configs, run, then regenerate displays)
python src/experimentA.py --generate-configs --run-experiments --seed=54210 --experiment-seed=65420
python src/regenerate_displays.py experiments/_ExpA_graph_properties/results/*/

# Experiment B (run multiple trials, regenerate displays, and run analyses)
python src/experimentB.py --run-experiments --experiment-seed=65420 --num-runs=100
python src/regenerate_displays.py experiments/_ExpB_RNN_controller/results/*/
python src/experimentB_analysis.py experiments/_ExpB_RNN_controller K01_cartpole
python src/experimentB_analysis.py experiments/_ExpB_RNN_controller K02_cartpole_min

Outputs (plots, metrics, and intermediate JSON/CSV artifacts) are stored under the corresponding experiments/*/results/ directories.

Reproducing figures and analysis

• Experiment A visualizations are regenerated via src/regenerate_displays.py over experiments/_ExpA_graph_properties/results/*/.
• Experiment B includes additional analyses:
  • src/experimentB_analysis.py experiments/_ExpB_RNN_controller K01_cartpole
  • src/experimentB_analysis.py experiments/_ExpB_RNN_controller K02_cartpole_min

Citation

If you use this repository or build upon MorphoNAS, please cite the preprint:

@misc{glybovets2025morphonas,
  title        = {MorphoNAS: Embryogenic Neural Architecture Search Through Morphogen-Guided Development},
  author       = {Mykola Glybovets and Sergii Medvid},
  year         = {2025},
  eprint       = {2507.13785},
  archivePrefix= {arXiv},
  primaryClass = {cs.NE},
  url          = {https://arxiv.org/abs/2507.13785}
}

Notes

• Python version: it was run on Python 3.13. See requirements.txt for dependencies.
• The experiments were run on a single machine in CPU-only mode. CPU: Apple M2 Pro, RAM: 96GB.
• Random seeds can be controlled via the CLI flags shown above to support reproducibility.

Contact and issues

For questions, bug reports, or feature requests, please open an issue in this repository. For paper-related inquiries, refer to the contact information on the preprint page: arXiv:2507.13785v1.