This is cool, though (due to overfitting) it's really just "memorizing" the first level rather than learning how to play in general. It'd be interesting to see the same technique with separate test and training data sets.
I wonder if there is a way it could learn to get over that tall pipe in this one. It might be sufficient for it to learn to deliberately attack any enemies it finds, combined with its tendency to move right all the time. But I suppose that might be too complex a set of actions for this.