I'm a CS MSc student, currently looking for thesis ideas. I'd like to have my thesis in the field of AI, preferably games AI. I'll be happy to hear interesting ideas (I have sound knowledge in the AI field, but not in the games AI field). All ideas and comment will be appreciated.

Only advice I can give (from someone who tried to write his thesis on the same subject, but gave up and picked a different one), is to try to make your subject as narrow as at all possible. Games AI is *really* vague, and has a hugely broad scope. It'll kill you if you don't find something much more specific to work on. [wink]

Choose a genre. Consider several top selling titles in that genre. If you have access to them, play them. If not, read about them. Work out what (in relationship to the AI) worked and what did not. As yourself whether or not there is the potential for better gameplay through better AI for that genre (not just those games). If there is, relate that back to what you know about AI and discuss this with your (potential) supervisor(s). Work out a research question (not just an idea for an implementation) that would be a worthy contribution to the body of knowledge in AI as related to games and/or simulations and/or robotics.

An example:
Genre: Strategy Games
Titles: Command & Conquer, Civilisation, Warcraft
What did not work (AI): AI was static in a level and did not change from one instance of playing the game to another. AI used similar strategies across 'levels/maps'.
Research questions:
(1) "Would dynamic strategy generation improved gameplay without being overly challenging?"
(2) "How can game agents construct dynamic strategies?"


The sorts of AI tools one might consider in this work would be plan recognition/ (recognising what the human player is attempting to do), opponent modelling (working out what the opponent knows (from what they're doing), planning systems, data mining and feature detection, etc, etc...

It would not be too hard for someone who has played a variety of games to ascertain what are the shortfalls in game AI across most genres. The difficuly for you is working out what is a worthwhile research question that will earn you a MSc if you can answer it!

Cheers,

Timkin

Quote: Original post by Timkin
Choose a genre. Consider several top selling titles in that genre. If you have access to them, play them. If not, read about them. Work out what (in relationship to the AI) worked and what did not. As yourself whether or not there is the potential for better gameplay through better AI for that genre (not just those games).

If one isn't comfortable with subjective metrics like "better gameplay" (or even "fun"), there is a different genre of research project: the really strong, superhuman-level AI.
I know of a few subgenres:
1) Completely solving a simple game (e.g. reduced-size mancala) with clever representations, clever heuristics, all sorts of clever optimizations and serious processing power. Advantages: leaving a durable and popular tombstone in the literature about that game, possibility of compensating lack of ideas with brute force. Disadvantages: narrow usefulness.
2a) Playing better than existing computer players because of more detailed evaluation of positions, smarter representations, and generally increased complexity. Advantages: pride, application to practical software and to competitions. Disadvantages: significant risk of failure or unexpected effort.
2b) Playing merely well, but using much less processing power than previously needed thanks to optimal algorithms, better heuristics etc. Like above, with a different emphasis.
2c) Playing better than existing computer players using similar techniques, but worse than the more sophisticted best ones. A good fallback plan for case 2a.
3) Playing merely well with a new or advanced technique applied to some kind of game for the first time, thus proving its usefulness and its equivalence with other techniques. Advantages: AI techniques have much warmer fans than AI problems, both in general and among thesis advisors. Disadvantages: possibly constrained to "pure" approaches.

I was thinking the toerh day about how boss battles in certain games can get repetative when you replay the game, or lose against a boos and try again. So something which I've termed "strategy blending" came to mind. Instea of bosses using the same AI strategies to fight you (or even normal enemies), how would a strategy blending system work.

You know like animation blending, only with AI instead.

Just a thought.

Oh, completely random idea. A year or so back, I saw two guys from my uni present their thesis about AI. Their idea was to see if a Neural Network could play Risk. :D

Thanks for the replies!
One of the major issues I have is implementation. After building a theory and testing it on simple environments, it's time to test it on more complicated games. With shooters, I know that Quake3 is open source now and that Unreal has a scripting language that one can use to modify the AI, but what about other genres? Are there any real time strategy games or RPG's that allow you to easily edit the AI?

Thanks!

Take a look here: http://gameai.com/. Under "Extensible AI" you'll find some of what you are looking for.

Well, with the new gaming devices that have multiple processors, multi-threading is going to be an iteresting issue for AI. One way to go about it is to somehow create a "anytime" algorithm that is constantly evolving enemy behavior. So, it's kind of like create a NPC bootcamp on the second core or seoncd processor. Then when their needed, the best are pulled out. The system then gets feedback from their performance for further improvement.

Personally, I feel that just simple better solutions won't do as the biggest problem nowadays with the fast changing hardware it how to implement things optimally to hardware. So, you pretty much need to start designing AI that fits the hardware. The whole topic in itself may end up being broad enough for a dissertation, but part of it will definitely be enough for a thesis.