Driving Evolution

A lightweight, browser-based simulation that demonstrates how simple neural networks can learn to drive using an evolutionary process. Cars sense their environment with ray sensors, make steering decisions with a small feed‑forward neural network, and improve over generations through elitism and mutation.

Demo

• Try it out here
• Watch the overview video: YouTube demonstration

Features

• Neural-network controlled cars with 5 forward-facing ray sensors
• Evolutionary loop with elitism and mutation across generations
• Moving obstacles and road boundary collisions
• Real-time HUD (generation, distance, alive count)
• Configuration modal for population and mutation parameters
• Statistics modal with a graph of average distance per generation

How it works (high level)

• Cars cast 5 rays to measure obstacle proximity; distances are normalized.
• A compact neural network (inputs: 5, hidden: 3→2, output: 1) outputs steering in [-1, 1].
• Fitness is proportional to distance traveled before collision.
• After all cars die, the next generation is created via:
  • Elitism: best performers are cloned directly
  • Mutation: clones of elites are randomly perturbed
  • Occasional random newcomers for diversity

Usage

1. On first load, the configuration modal appears. Adjust parameters:
   • Number of Obstacles (default 7–8)
   • Number of Cars (population size)
   • Parent Count (elites carried over)
   • Mutation Rate (0–1)
   • Mutation Magnitude (0–0.5)
2. Click “Start Simulation”.
3. Use controls:
   • Start/Pause: toggles simulation
   • Reset: reconfigure and restart
   • ℹ️ Info: opens a primer on concepts and tips
   • 📊 Stats: opens a graph of average distance per generation

Project Structure

DrivingEvolution/
  index.html          # Canvas, modals, and controls
  style.css           # UI and canvas styling
  main.js             # App bootstrap, loop, UI wiring
  helpers/
    SimulationManager.js   # Population lifecycle, evolution, rendering HUD
    StatisticsManager.js   # Modal and canvas-based stats graph
    SimulationState.js     # Run/config state machine
    ObstacleManager.js     # Spawning, updating, drawing obstacles and walls
    roadLines.js           # Center line animation
  classes/
    Car.js                 # Car entity, sensor rays, fitness, drawing
    NeuralNetwork.js       # Feed-forward NN, predict/clone/mutate/crossover
    Obstacle.js            # Obstacle entity and collision
    Ray.js                 # Ray casting against rectangles

Technical Notes

• Neural network: tanh activations, output in [-1, 1] maps to horizontal movement.
• Sensors: 5 rays spanning a ~60° field; exponential scaling improves near-field sensitivity.
• Fitness: cumulative distance while alive; collisions end the car’s run.
• Evolution: elites cloned directly; additional population filled by mutated clones and some random cars.