Deep Genetic Learning Cars
Cars driven by small neural networks, evolved by a genetic algorithm instead of trained with gradients.
2022
A simulation that evolves neural-network drivers with a genetic algorithm instead of training them. A population of cars drives a track, each steered by its own small neural network reading distance sensors. There is no backpropagation anywhere. Every generation, the cars that travel furthest before hitting a wall are selected, their weights crossed over and mutated, and the next population runs the same track. The simulation draws the active network live beside the track. Over the generations, the cars progress from crashing at the first corner to completing the circuit.
My role
Sole author.