What is it?

This simulation uses neuroevolution to evolve both car shapes and neural networks that control them. Each car has a simple neural network that improves through generations without backpropagation. Based on BoxCar2D but rewritten from scratch with neural networks, using Box2D physics engine.

Neural Network

• Inputs: Velocity, angle, distance sensors
• Hidden layer: Information processing
• Outputs: Motor speed, wheel angle
• Learning: Evolution through selection and mutation

Genome

• Car shape and physical properties
• Neural network weights and biases

Controls

• Save/Restore: Manage populations
• New Population: Restart evolution
• Fast Forward: Speed up simulation
• Mutation rate/size: Control evolution parameters
• Elite clones: Best cars copied to next generation

Graph Colors

• Red: Top score
• Green: Top 10 average
• Blue: Generation average

Credits

• Original BoxCar2D concept by BoxCar2D
• Box2D physics engine by Erin Catto
• Based on: Genetic cars 2
• seedrandom.js by David Bau