A lot of people target Starcraft since it's a pretty well-developed game with well-understood rules. That gives you fewer variables to worry about (in the science-experiment sense) when trying to isolate what effects your AI implementations have on gameplay. I'd suggest starting there and broadening out to other games/engines as you get more familiar with things and have a better sense of what your own requirements look like.

One feature that you probably need to qualify is having some kind of program interface (external access to output data and input channels) or scripting interface that can push data to somehow create a program interface.

Games which you have access to the source code and build pipeline you could add such an interface if needed.

How many commercial games have any of the above ?? Open sourced games will make this access possible (but varying degrees of documentation might still make interpretting the gorilla code/data schemes a chore)

Thats the first pass criteria -- unless you somehow can decode packet traffic and can get internal knowledge of what the data means to get ahold of game state/mapping data etc.

--

Some games like the old Neverwinter Nights really could have used a program interface to allow players (who in that game ran the Multiplayer servers) .to make needed improvements the game company completely ignored (like an inventory 'bank' system' that players had to go to extreme lengths/difficulties to add to the published server software). Compentant player programmers were fully willing to improve that game (and were NOT in competition with the company) -- result -- the company lost some significant amount of business for being shortsighted.

Such an interface would be invaluable for debgging tools to the game company itself, but its possible that they also see it as a vector for easy cheating (which has added emphasis now with the online 'Store' connections and competative 'player achievemnent' features used to justify that online component.)