How will this work in a MineCraft simulation where there are no rules? Where will the feedback loop come from, e.g. a subjective evaluation by "humans"?
And then there are a bunch of statistics. http://minecraft.gamepedia.com/Statistics
I agree this would make a player that has an unusual play style (never making a base, for example), but it might make for some fun Let's Play videos.
And increasing YouTube subscriber count, and video view count, and video thumbs up, while minimising video thumbs down would be an interesting experiment. ("I welcome our AI Clickbaiting overlords")
The idea I had was to reward collecting blocks of different types, with exponentially decreasing reward. So finding a block it's never found before is 100 points, the second it finds is 50 points, and so on. It encourages exploration and touching on as many mechanics of the game as possible.
The important part is that the AI learns on its own though, not programmed to do specific tasks by a human. However it should perhaps be given recipes and some other meta information, because discovering those by chance is highly improbable.