Hi all, I'm coming towards crunch time for my MSc in Computer Games Technology in uni and thesis topic time has come around! Basically I have to come up with topics to pitch to my lecturers and depending on demand for certain lecturers, I get one out of three suggested topics. My main interest in games is in AI and have previously done a classical AI course in my BSc so have a pretty good grounding in the basics, perhaps not specific implementation in games. I've been reading up a bit and my first topic is 'Improved NPC decisions using Hierarchial Planning' or something along the lines of deliberative intelligent agents perhaps using POP. I think I'll pitch this as my first choice but need other ideas if ye would be good enough to help. I'm interested in Emergent behaviour but aside from flocking can't really see any other major benefits, but am open to major correction, ideas hopefully :) Aside from Emergent behaviour, what other areas in AI are en vogue at the moment or any areas that companies you know of are actively working on? Any ideas will definitely be considered so I appreciate any feedback. Regards, Eddie

I personally believe that emergent behavior has tremendous room for growth and applications. I recently finished reading “Turtles, Termites, and Traffic Jams: Explorations in Massively Parallel Microworlds”, which has helped by giving me new ideas in my own pursuit of emergent behavior. The only fault that the book has is its dependance on StarLogo, which I admit had turned me off at first, but StarLogo is simple enough to not distract from the attempt to teaching how to model emergent group behavior.

Often times (as cited in the book) it is difficult to see or accept emergent behavior is based on our own desire to assign a cause or leader to an outcome vs something that has dynamics of its own.

In terms of "en vogue", I'm sure that Neural Networks (or at least practical applications of) are still high on the list.

What you tackle really depends on your timeframe (how many hours do you have to complete it in) and the expectations of your examiners regarding the level of novelty of the research. Remember that you're doing this for examination, not to produce a new product for the games market (although obviously a good idea might lead to further development work and an actual product, but don't make that your goal right now).

Here are two ideas that you might find interesting, or that might stimulate your own ideas.

1. Dynamic Quest Engine.
RPGs and MMOs that centralise gameplay around the completion of quests often have reduced replayability due to the lack of a 'surprise factor' that we all experience when we first play a game. Repeating a quest, even with a different character, is very much a grind through the gameplay elements, rather than an exploration of them. Maintaining freshness with regards to exploration of landscapes has been tackled with randomly generated worlds, continents, dungeons, etc, but little effort has been applied to the problem of dynamic quest generation. A dynamic quest engine would provide a greater sense of immersion in a developing landscape, allowing the players character to live an evolving life with continual novelty.

The quest generator itself should have several features:
a. generate quests that advance the games story
b. generate quests that enhance the games story
c. generate quests not related to the games story (diversionary quests)
d. generate quests that follow on from previous quests
e. generate quests that are inter-twined with any other quest it generates


2. Strategy Predictor
NPC AI is greatly enhanced by the ability to react to the players behaviour. In fighting games, pulling out the right move that defeats the players move presents an impetus for the player to try a different move, thus generating a fight as a sequence of attempted actions to defeat the opponent. The same is true in other genres such as strategy games. However, simply reacting to the players actions is not usually sufficient to defeat them (and so we often see 'cheating AI', or the an adjustable 'tougness' factor). Plan recognition is a problem of understanding the underlying decisions in another agent that generates a sequence of actions. Recognising the players plans allows for the formation of strategies consistent with the gameplay goals of the designers. This won't generally be 'always beat the player', but rather gives the designers the capacity to implement smart yet defeatable plans in their NPCs.


Cheers,

Timkin

If I were doing my Msc in AI instead of Computer Vision, I would try the use of the "Script"/"Scenario" AI technique in games. I think thats an awesome, very very underused class of AI technique. Brillantly simple. Could be two folds, applying scenarios for deductive reasoning/ action prediction and also learning reccurent scenarios from player actions. The only AI book I have which talks of this technique is "Artificial Intelligence" by Rich and Knight.

Please dont add anything else to the crappy connectivist fad that has been going on for 30 years. Planning is good, very hot in 'practical' game AI right now. You may want to look into architectures for parallel planning. Like it or not, parallel hardware are *finally* comming, and AI must take advantage of it.

Physics reasonning in game might be interesting too. Improving on reverse kinematics, let the agent reason about its own motion, and the physics of its surrounding.

Awww, there is so much fun to have in game AI :)

Thanks for the helpful replies. I've pitched one idea which is Goal Orientated Action Planning in games and will involve investigating GOAP's use in current games,its difference from , its advantages to developers, why it has only been used in FPS(as far as I can tell) and suggest possible implementations for other genres.

As far as timeframe, I only have 15 weeks for it so am not looking to do something ground-breaking!
Timkin, I would be very interested in the second idea you suggested and will definitely research this topic further, is it essentially building a player model which the AI uses to monitor + react to player patterns?
On the parallel planning topic.....would subsumption architecture fall under this topic? Its not really planning I guess as it is still rather reactive but the lower stages send messages upwards through the levels and can plan the long term goals of the AI so could it be considered planning? I'm only new to the idea of subsumption architecture so am still getting my head around it.

Quote: Original post by parrotbait
Timkin, I would be very interested in the second idea you suggested and will definitely research this topic further, is it essentially building a player model which the AI uses to monitor + react to player patterns?

The model that you're trying to build is of the cognitive process of the opponent, or at least it's mapping into a decision space. I can recommend a good starting point for your investigation would be the Bayesian Poker Player of Korb et al (Monash Uni). Read some of their papers and chase up some of their references.

Good luck!

I'll throw this term in to the mix. If your familiar with rendering and LODding meshes and texture LODs (levels of detail).

The how about 'LODding AI'? The games approach obviously for performance reasons, and scalability of worlds.

Well, I did mine on the use of Machine Learning to train competitive agents. The idea being that users train agents and those agents then compete against other agents in a game world to complete goals. I only had a little time to put it all together so I ended up just demonstrating it in a little text based fighting game but the theory was promising. I just lacked a rich enough action domain for it to be truly effective.

Another Idea that I have been meaning to research that could be promising is the "The use of semantic nets to mold malleable events" The idea being that a game writing creates soft events to set the theme or tone of events but the details are generated based on player actions preceding that event. So that if a writer was creating a romance plot arch they would create events that occur but the details such the subject of the characters romance would be determined by the player. For example the writer creates an event that consists of a huge fight at the players 1 year anniversary. They writer can write the hard details they want to occur but others left soft and shaped by the semantic net such as when the anniversary occurs, who it is with, where they go, and why they have the fight. The next event might be that either the player or their girl friend sleeps with someone else. If the player chooses not to then their girl friend will. In this way can blend script events with the actions of the player, without having to have heavily scripted plot archs, and hard storylines.

Well I eventually pitched my Masters as Enhanced NPC behaviour using GOAP.
Still at a reasonably early stage but thing I'm going to do a Unreal Tournament mod and have squad behaviour controlled using HTN and other behaviour using hierarchial FSMs and compare results.

Sounds good, you could always throw in a RL one to see if it can improve over time against its fixed opponents.
But Good Luck