I'm not 100% sure this use case applies, but since we are talking about mining data, I guess in some sense it applies...
I remember when Quake was the state of the art FPS and they allowed 3rd party modules (scripts, don't remember the details) to be added to the game engine. There were many bot scripts that allowed you to control the bots using these add-ons. Some were getting to be so good that sometimes you could not tell whether it was a bot or a real person playing.
Sadly the bots you find in today's games are extremely stupid. For instance, in Gears of War (1/2/3) the bot behavior is so predictable that you can quickly learn how to defeat them. It seems game developers gave up on making truly smart bots.
So I was thinking whether it would be possible to have learning bots. Basically, you start out with a "blank" bot that just stands there. When he gets shot, he learns the movements from the opponent that killed him. So this is added to his database of moves. As he gets into more combat, he becomes smarter and smarter.
I know that translating the moves of the opponent to work in a generic way in a dynamic environment is probably quite complex, but I believe much more difficult problems have been solved in the past already.
