Disclaimer: I spent over 2 years working on both Prince of Persia and Assassin's Creed, I am not legally allowed to tell you how these games are made, but I will try to give you general knowledge about game AI. Most games do not use learning as you described.
First, the language or the platform (or device) you target has nothing to do with the artificial intelligence
tools you will be using.
Second, your two first examples, tic tac toe and chess are turn-based
games, this is totally different than a real-time
game. Also, both of these games work on a limited board with a somewhat limited search space
. These games would be way more easily solved using other algorithms than learning (like minimax for example).
In the case of a real-time
adventure game in 3D space like those you described, just even finding the context
or input
you will be feeding to your learning system will be hard... what elements are pertinent for the decision making: recent audio stimuli, visual stimuli, relativeness of these stimuli in space and time, surrounding enemies or allies, etc... Then, given the complexity of these world and what NPC
s are allowed to do, how do you define the output
of your learning algorithm? How does this output remain cohesive over subsequent queries?
Keep in mind that whatever changes you make to your game design will require rerunning the learning process and possibly the definition of the input
and output
.
You are better off going the standard route, use well known methods such as finite state machines
, fuzzy state machines
, behavior trees
, planning
, etc... You will have more control over the decision making and will actually be able to design your NPC
behaviors, which you wouldn't be able to do with learning.