🚀 Project Demo
Watch the evolution from random driving to optimized behavior over generations.
🧠About the AI System
This project simulates autonomous agents — virtual self-driving cars — trained using evolutionary machine learning techniques. Each car is controlled by a lightweight feedforward neural network, which takes in real-time sensor inputs to make navigation decisions.
The training is entirely unsupervised: no predefined routes or rewards are manually given. Instead, a genetic algorithm evolves the neural networks across generations by:
- Fitness Evaluation — scoring cars based on how far they travel without crashing.
- Selection — cloning the best-performing neural networks.
- Mutation — introducing random small changes to improve exploration of solutions.
Over successive generations, this process leads to increasingly competent AI agents, capable of navigating complex, dynamically generated tracks without direct programming of driving strategies.
🔬 Neural Network Visualization
Real-time visualization of the neural network adjusting control decisions.