Thats the problem I'm currently facing...

You cant provide it with feedback about its behavior, or it wont be a true A.I. You must provide it with feedback ONLY of its own behavior changing the situation...It must be able to 'decipher' the change in the situation and judge if its action was in the right or wrong direction...

So, the feedback can be constant, from a whole picture (video feed) to numbers it doesnt matter, and it needs a way to judge the change in the situation if its bringing it closer to the goal at hand performing its current action. (Plus, as always, it needs a note of randomness to experiment with new actions)

Lets try a little example, based on your post. And lets consider in this example a Counter-Strike bot (we're all familiar with those).
You can begin by building a simple AI that can move and shoot and give it a goal of killing the enemy bot.
Even if you dont give it the knowledge of what movement is, it can experiment in its input channels and based on distance to judge if its getting closer to its goal.
This is the simplest form of 'learning' AI, but it only learns how to move.
People gave the advanced bots programming to hide or jump from behind, but this has to be left outside of the programming, the A.I. needs to come to those things on its own.
Imagine thats all the bot learns and knows from its first round. Now it also knows where it met the enemy and how well that encounter went.
It can take into account the last encounter when it meets the bot again, perhaps in the same spot, and judge if it came quicker or slower and what is the difference in its health from the last encounter.
And somehow take into account this new information in the second round and perform better or worse, and make an even bigger assessment the third round based on the first two, learning the patters in behavior of the other bot, like most visited place, speed of shooting and speed of movement to that place, and basically take all this information at random when it needs to and judge how it should perform.

Lol, this sounds a lot more complicated than what I meant to describe

I was thinking about Neural Networks today. Instead of if-else statements for behavior, Neural Networks play out exactly the same role, but they can be modified easily, meaning it can be changed, unlike hard-coded conditioning... The 'action decision' part of the A.I. brain could be sculpted with a Neural Network acting like a group of if-else conditions to decide from possible actions, but it still needs a separate module for modifying the Neural Network based on the input,output (difference in the situation) of the system.

In this module lies the difficulty, making it judge how to modify the NN to perform better. Its like building an A.I. to guide another A.I. I guess And it'll probably take even more A.I. modules than those two to have something at least as smart as a bug or a bird.

Any thoughts on that? How could an A.I. be built to modify an N.N. that decides the actions needed to be performed.

PS.: I dont have a proof, but I believe the key to (at least a game A.I.) true A.I. in all this lies in the input towards the A.I. Maybe giving it just a little info of positions and other object statuses is not enough. Perhaps it needs a full video feed and an object detection system, in order to really 'feel' its environment, to have a correct interpretation of the distance to every point in its surroundings, and possible paths (where it can go and cannot go at every given time). But this would require an even greater amount of A.I. to detect and recognize and remember objects and forms...

And, by the way, this thing I'm talking about in THIS reply (not a really smart A.I. but in the right direction A.I.) will be more universal, not just acting as a CS bot, but for other games or even programs with different goals than kill that guy... This will be a smart A.I., compared to the current A.I., which are really built for a specific task and cannot operate outside of their given environment (meaning you cannot take a Google bot and plug it in Battlefield or vice-versa...) Could that be a foundation of true A.I.??