EvolutionaryOptimization

Evolutionary optimization of a CNN for MNIST digit classification — no backpropagation required.

Overview

This project evolves a population of CNNs using genetic algorithms instead of gradient descent. The evolutionary strategy includes:

• Elite selection: Top performers survive to the next generation
• Layer crossover: Random swapping of entire convolutional layers between parents
• Parameter crossover: Weighted averaging of parent parameters
• Mutation: Reinitialization of the last convolutional layers

The model (SimpleCNN) has 10 convolutional layers with BatchNorm and achieves ~98.9% accuracy on MNIST after 160 generations.

Project Structure

EvolutionaryOptimization/
├── src/
│   ├── __init__.py       # Package exports
│   ├── model.py          # SimpleCNN architecture
│   ├── evolution.py      # Crossover, mutation, population evolution
│   ├── evaluate.py       # Model evaluation and visualization
│   ├── utils.py          # Data loading, model save/load
│   └── logger.py         # EvolutionLogger for tracking metrics
├── train.py              # Main training script
├── Experiment_1.ipynb    # Original experiment notebook
├── requirements.txt      # Python dependencies
└── README.md

Setup

pip install -r requirements.txt

Usage

Run the full evolutionary training:

python train.py

Or use the notebook Experiment_1.ipynb for interactive experimentation.

Results

After 160 generations with a population of 10:

• Best accuracy: 99.1% (generation 150)
• The first ~100 generations show slow improvement (~19% accuracy) as structure-based crossover explores the search space
• A breakthrough occurs around generation 102, after which accuracy rapidly climbs to 98%+