🧬 Self-Evolving Neural Network — 4524d43d
Evolved locally with neuroevolution (genetic architecture search), then uploaded from this machine.
Model
- Architecture: gated MLP — top-3 features → layer 1, remaining features join every layer after
- Layers:
192(swish), 96(selu), 32(swish), 192(linear), 256(relu), 64(sigmoid) - Params: 102,216
- Learning rate: 0.005 · Optimizer: rmsprop · Batch size: 16
- Top-3 feature gate: [4, 9, 16]
- Fitness: 331118.82
Evolution history
- Generations: 18
- Initial best fitness: 37948.7707
- Final best fitness: 240545.7623
- Improvement: +533.9%
Files
| File | Description |
|---|---|
evo_checkpoints/best_model.keras |
Trained best model (Keras) |
evo_checkpoints/best_genome.json |
Best architecture genome |
evo_checkpoints/evolution_history.json |
Fitness across generations |
self_evolving_model.py |
Core evolution engine |
evo_gui.py |
Flask web dashboard |
requirements.txt |
Dependencies |
How to run
pip install -r requirements.txt
# Continue evolving from this state
python3 self_evolving_model.py --continue --generations 50
# Or launch the web GUI
python3 evo_gui.py --port 5000
Scaling up (v2/v3 on GPU)
python3 self_evolving_model.py --continue --max-units 512 --max-layers 8 --train-epochs 20
See evo_v2_colab.ipynb for a ready-to-run Google Colab notebook.