Maybe I'm wrong somewhere...
I tried thinking about an A.I. outside of a game world, and imagined it in an email context where it has to sell stuff through emails, by constructing a message of its own chosen words list from a word db.
It can get feedback by receiving email responses, which can be interpreted into 'energy' and the bigger the email, the more 'energy' it gets. But, this is still a fixed solution, because I'm telling it that the more emails it gets (and bigger) the more 'correct' its actions are. I'm not sure if this is the correct way to inform it of its success, and it will be the correct in the context of selling stuff online, eventually it will start constructing meaningful emails that actually sell stuff, but if you get it outside of this context it wont be able to perform with a different set of input/output info, say flying the RC helicopter from my last post, or at least I'm not sure it will still be able to work.

What do you think?

Now, to clarify, the email seller A.I. wont know its actually picking words out of a db and selling stuff to someone, it is just manipulating data, its environment, into getting more 'energetic' responses... The same goes for the RC example, it wont know what a rotor is, it'll just have input/output of dangerous or safe distance from the ground and its available controls and real energy (battery charging?, but this requires also location mapping)...