Genetic algorithm + neural networks race procedural circuits. No ML libraries — built from scratch.
Choose a circuit, press Start, then adjust settings while it runs.
Population: number of cars per generation. Mutation: how much offspring weights change. Sim Speed: training speed multiplier.
Rays: number of distance sensors used by each car (more rays = more context, lower performance).
NN Layers: hidden layer depth. Neurons/Layer: hidden width for each layer.
Starts are fully random with no driving priors, so steering/throttle/brake behavior must be learned from evolution.
Reset starts over with the current settings.
Inputs: 9 rays + speed. Outputs: steer, throttle, brake.