Evolutionary Driving Simulator

This project explores training neural networks without gradient-based methods by using evolutionary algorithms, providing a deeper understanding of learning beyond standard backpropagation. By removing reliance on differentiability, the approach enables experimentation in environments where gradients are unavailable or difficult to define.

The system is a 2D top-down driving simulator where agents navigate custom tracks using sensor inputs. Neural network weights are evolved using a genetic algorithm, with each generation evaluated based on driving performance and improved through selection and mutation.

The architecture cleanly separates simulation, evolution, and visualization. A custom map editor allows rapid environment design, while a real-time viewer enables observation and analysis of agent behavior during training.

To improve generalization, a retention-based training strategy was implemented across multiple tracks, encouraging agents to retain useful behaviors instead of overfitting to a single environment.

Key Features

• Genetic algorithm training for neural network controllers
• Multi-track retention strategy for improved generalization
• Real-time visualization of agent behavior and training
• Interactive map editor for rapid experimentation

This project demonstrates a strong understanding of machine learning fundamentals, including neural network representation, optimization strategies, and system design for experimentation.