Watch RL at Work
Ninety neural networks start knowing nothing at all. The ones that get furthest survive and breed. Nobody tells them what a corner is. Switch to Drive to take the same car round yourself.
What you are looking at
Nobody shows it a good lap
There is no dataset of correct steering. Each car reads five distance sensors and decides for itself; the only feedback is how far round it got before hitting a wall.
The furthest survive and breed
At the end of a generation the best networks are kept, paired and mutated. Everything else is discarded. Corners are learned because cars that take them badly do not have children.
Nothing is running off screen
The networks, the physics and the evolution all run on this page, in JavaScript. Reload it and the population starts over knowing nothing at all.
A demonstration of the method, not a product. The same approach an agent improving against an objective nobody has hand-coded a rule for is what we apply to pricing, routing and scheduling problems for clients.
Talk to us about your own problem →