in game AI, if you want to win the nobel prize, come up with a faster and easier A*.

some devs favor planners.

some favor HFSMs.

some favor scripted rule based behavior.

i'm partial to modular hierarchies of expert systems myself, where each "module" is an "expert" at some type of decision making in a behavior tree (i guess you would say). a module may be a behavior tree, a FSM, a planner, NN, whatever works best for that bit of the overall system.

i've considered all types of AI over the years, but have never found one superior to the hierarchical "expert system" using appropriate AI types. although expert system really describes a behavior, its not a particular implementation type, such as decision tree, although decision trees are a common means of implementing them.

so as you can see, the intersection of popular academic AI topics and game AI is almost the empty set.

we get someone on here about every six moths asking this type of question.

most are surprised to learn that game AI and academic AI research are so far apart. and FYI, I completed the software engineering program at OSU, and i took classes in AI.

i think there are four realities in game AI that one doesn't usually have to consider in academia, which drives the difference in the evolution of AI approaches:

1. realtime - you get 5ms to do all AI for 1000 entities.

2. kiss - keep it stupid simple

3. gotta ship - the design must implemented, debugged, and working correctly in less time than it really take to thoroughly test all cases and combos, yet all must work.

4. underpowered hardware - you must design for the most abysmally under powered processor the game can possibly run on. i have a saying: "put a Cray on every users desk, and i'll build you a REAL game".

another one might be:

5. candy sells - at the 11th hour, your millisec budget is cut in half, because render is slow. now you must update those 1000 entities in 2.5ms, not 5ms - yet their behavior can't be adversely affected.

A suggested project

Analysing what is happening in a played game - either by output analysis (graphics, io events) or hooking into the guts of the program (few published games have a 'program interface' available though).

Observing and classifying activities going on (both the environment and what the player(s) do in reaction to it).

Interpreting, Classification and Factoring (to build an internal symbolic model)

This is a basic 'tool' of anything further in AI - converting and boiling down everything into machine useable data -- and then reexpressing it in a human readable form (needed to verify the whole thing works and THEN to be made use by the human component of AI - where the human is asked to give hints/teach about what is good or bad - which always has to be done one way or another)

Game companies want to see how players react to their game (during development) to spot patterns which indicate it is too simple or overcomplicated or too narrow in solution/activities done by real players (having a player use just one or two optimal tactics the whole game makes it a waste to develop a whole lot of possibilities and make the game overly repetitive/less interesting.

and one more question:

i chose a machine learning system that bots learn from best players to make their behave and performance better. what are most important multiplayer games with open source script ?

You realize that there are probably 3000 academics working on this exact thing (with fairly tepid success rates), don't you?