ai-f1-racing
Evolutionary racing AI • Deterministic sim
A top-down Formula 1 racing AI. A 38–16–3 neural network drives with analogue throttle, brake and steering, trained by an evolution strategy — antithetic perturbations, rank-normalised fitness, Adam on the search distribution — against a deterministic 2D simulation running at a fixed 60 Hz.
Determinism is a hard contract here: any function whose result is not bit-identical across JavaScript engines is banned from the simulation, and trigonometry is baked into tables at load. That is what makes a lap time mean the same thing twice and a training run reproducible. Spa-Francorchamps and Baku are traced from the real circuits to within about 0.1 % of their true lap lengths.
The simulation and the trainer work — a driver can be trained to lap. Cars cannot yet see or hit each other, so the racing itself is the part still being built.