AI Research methods

I can tell you how research in general is usually performed. You start by doing what's called a literature review. In the lit review stage, you just spend your time reading as many recent papers as possible. The papers should somehow by connected to the project that you're working on. Then, you start to re-implement some of the more interesting papers. Once you get into this stage, you usually become much more aware of the flaws in those papers and you can spend some time doing deep thinking about how to overcome those flaws. Then, you build your own system, run some experiments, and if it really is better than the original, you go publish your results.

read as much as you can get your hands on
code as much as you can
develop a skillset and a toolbox of tricks

talk to the designer about exactly what he wants the AI to do.
think of a way to do it

There's no real formula. 99% of Game AI is just a simple state machine with some icing poured over it to give it a bit more depth. The most important thing is to just let the design dictate how the AI will work. Once you have a behavioral set nailed down, it's generally just a relatively straight forward implementation issue. Clearly the more exposure you have to different architectures, the easier it will be to come up with an architecture that works. It's also important to realize that there are generally at least 3 completely different ways you can get the same or similar effect, so there's no "right" answer to this kind of problem. Different solutions will have different edge cases or performance issues, so it's always important to have an overall understanding of where the AI fits into the game performance-wise and design-wise,

-me