See the GitHub repository here.

I’m a massive fan of the Box2D physics engine and have wielded it in a few languages as part of game and simulation projects. I spotted a Go port and wanted to somehow combine this with the Pixel 2D game library and decided this top-down car demo written in JS with Box2D would be a fun base to start with. So, I went ahead and migrated it to Go with Pixel & Box2D, adding enhancements such as vehicle health and destruction.

I’d studied evolutionary algorithms at uni but had never actually implemented them. I decided to encode the car’s movements as a genetic sequence, and iteratively evolve the car’s performance in order to solve a basic problem. The car evolves to drive to and gracefully brake on a green target point. It uses roulette wheel selection combined with random cross-overs, swaps and mutations to produce the following generation. The fitness function is a combination of the distance from the target and the final velocity at the end of the sequence execution to encourage the car to precisely park on the target rather than speed over it.

The following video illustrates the improvement in fitness between evolutionary iterations. I’ve also included a surprising solution which evolved after hundreds of iterations, showing how machine learning and evolution in general can provide interesting and less obvious solutions…