Neural nets training to race cars
This project by Tomasz Rewak hits a bunch of my interests: it uses neural nets and genetic algorithms to train generations of simple race cars as they gradually learn to navigate the track. It’s also user-editable in the browser, if you want to play with the parameters yourself.
It takes a few generations for the cars to start figuring out how to get around, though once they do it won’t take long for them to start finishing the race. You can watch as they get better: at first they may overfit and only turn in the direction of the first curve they’ve learned, but once one figures out that there are two directions to turn, the next generation will learn from that one.