I stumbled upon this really cool use of evolving neural network application.
This demo uses a method called Neuroevolution to train the robot Mario player. The theory is based entirely on biology and evolution. Training is split into sessions, after each of which, the best performing neural networks are 'bread' together and a small amount of random mutation is introduced. Just like how all living things on earth do.
This particular implementation uses NeuroEvolution
of Augmenting Topologies (NEAT). You can read the full paper here. Thanks to the author for this awesome demo!