Genetic Neural Network

A neural network trained via genetic algorithm, implemented from scratch in JavaScript with no ML frameworks.

How It Works

The system evolves a population of 10,000 neural network individuals over thousands of generations. Each individual's fitness is evaluated by simulating decisions on historical data. The best-performing individuals survive and their weights are mutated to create the next generation.

Training Pipeline

1. Initialize - Create 10,000 individuals with random neural network weights
2. Evaluate - Each individual processes the dataset and makes decisions
3. Select - Rank individuals by fitness (cumulative performance)
4. Mutate - Best individuals survive; weights are randomly perturbed
5. Repeat - Evolution continues for configurable number of generations

Features

• Neural network forward pass implemented from scratch
• Genetic algorithm with selection, crossover, and mutation
• Population-based training (10,000 individuals per generation)
• Configurable network architecture and training parameters
• Binary weight serialization for saving/loading trained models
• No external ML dependencies

Tech Stack

JavaScript (ES Modules), Node.js

Usage

node index.mjs

Configuration

Key parameters in index.mjs:

• Population size
• Number of generations
• Network layer sizes
• Mutation rate and range

License

MIT