NeuroEvolution of Ant Dynamics (NEAD)

Authors

• Michael Crosscombe (@toohuman)
• Ilya Horiguchi (@NeoGendaijin)

Project Overview

Real Ants	Simulated Behaviour

The purpose of this project is to evolve neural networks that can accurately reproduce realistic ant dynamics and, eventually, colony-level collective behaviours.

Project Structure

• WANNTool/: Core WANN implementation and tools
• prettyNeatWann/: Main implementation directory containing:
  • Domain-specific environments
  • Training and testing scripts
  • Analysis tools
  • State space visualisation

Features

• Ant trajectory analysis and state space representation
• Behavioural clustering and pattern recognition
• Neural network evolution for ant behaviour reproduction
• Colony-level behavioural analysis
• Visualisation tools for behavioural state space

Dependencies

• Python 3.11+
• NumPy
• Pandas
• SciPy
• Scikit-learn
• Matplotlib
• Seaborn
• Gymnasium
• PyTorch

Usage

Main scripts can be found in the prettyNeatWann/ directory:

• wann_train.py: Train WANN models
• wann_test.py: Test trained models
• ant_state_space.py: Analyse ant behavioural states

License

[To be added]

Acknowledgements

This project builds upon the Weight Agnostic Neural Networks (WANN) implementation from the brain-tokyo-workshop repository by Google Research. The original WANN implementation is described in:

@article{wann2019,
  author = {Adam Gaier and David Ha},
  title  = {Weight Agnostic Neural Networks},
  eprint = {arXiv:1906.04358},
  url    = {https://weightagnostic.github.io},
  note   = "\url{https://weightagnostic.github.io}",
  year   = {2019}
}