This is actually pretty interesting... Wouldn't mind seeing this on a bot for Motocoin, but since Motocoin's levels are random would this approach work?
Maybe you could train it to select levels by showing it completed levels in lots of the last blocks. After it has created a neural network (I think thats what it is called) for selecting levels, you can then train it for a second time but this time completing levels. Maybe one day, when I know how to code. I can read code pretty well, I just need to learn it...