AI, Where it's at!

[ronnoc10]

For the record, I don't know too much about what Game AI says in the code, but I'm going to rant about what it does on my screen.
It seems to me that modern AI is directly deterministic, as in, 'if these conditions are met, do this'. I believe that AI should be programmed as a Complex System, or, more specifically, a system of desire. All human behavior can be boiled down to desire. As of now, I am typing this because I desire too.
If a AI is given a system of desire, conditions, such as player position, no longer directly dictates the AI's behavior. The player's position instead influences the desire system. The item at with the most desirability is what the AI does. Of course, the AI will have to be made up of many systems, an attack, movement, ex. all tied together with more systems, calumniating in one big system.


I'm not sure if what I'm saying is very clear.

[HagNasty]

Hesham and I where actually talking about this at the beginning ot the thread. The link below is a online book I have been reading to improve my knowlege of AI. It has opend many doors to me. Basically it is about learning and decision making for long term goals(or desires).

The way some games have AI built into them now is based off of values and projections. The AI does not know what the player will do but it knows what it might do and takes a rout that will lead to it's ultimate goal.

Some squad based AI might do the following:
larger creatures will take up a defencive position to soak up player damage while weaker ranged character try and pick off the player from behind them. This follows the smaller creatures final goal more directly (Attack and kill the player) though the larger characters don't they see that they may have a better chance of killing that player if they soak up the damage for the ranged characters.

I would suggets you read this book it's very good and have been enjoying it myself. Only wish I had more time with it.

http://www.cs.ualberta.ca/~sutton/book/the-book.html

[ronnoc10]

Thanks for the link! I will certainly look into it!

[ronnoc10]

Just saw this article:
http://www.gamasutra.com/view/featur...al_.php?page=1

Personly, I thought it was horrible. For one thing, he seems to think appling the Uncanny Valley to AI is an origanal idea, despit the fact that it is a theroy in robotics, and was latter applied to CGI. Over all, I thought it was pretty bad, what'd you guys think?

[HagNasty]

I'm almost afraid to look now :P

[Dominia]

Hey guys, I'm currently still in school and will be looking for a game company to join this winter when I graduate. Just wanted to stay as a student, this site is an amazing resource and I've highly recommended GameCareerGuide (and it's sister site) to many of my peers. Keep up the great work.

It's interesting I found this thread because for last year or so (going back to January '06) I've been working on an in-house video game here at the University of Texas at Austin that was designed specifically to make use of learning algorithms and Neural Networks. The game is called NERO and we will be releasing a 2.0 version in a matter of weeks (or whenever we update the website, darn marketing ><).

Basically in NERO you get a team of 'dumb' robots, train them, then send them off to battle. Nothing is scripted and there are no finite states in the AI, it is all real-time self learning. We're using a modified NEAT algorithm called rNEAT ('r' standing for real-time) as the AI and the Torque engine for graphics.

It is not pretty and there is no storyline unfortunately (yet anyway) but overall once you get the hang of training the NERO agents, one can get some amazing behaviors. Currently I'm working on an open source version of NERO (completely redesigned the code since we had a myriad of students working on the original project with no version control ) that should be ready at least for the research community in the next couple of months.

There is only 3 of us over the summer making the port so the code and doc will be up to snuff thankfully lol. It is our hopes at least that we can release some rNEAT AI libraries which could interface with other games by the end of the year, but that is further down the road.

I didn't want my first post here to sound like a shameless plug (it did, I know sorry) but I do want to expose ya'll to some current implementations of advanced AI in games where the AI is the focus, and not a back seat driver.


--Jason Hooten

[ronnoc10]

If you don't mind me saying, your game sounds a lot like Robocode...

[Dominia]

Originally Posted by ronnoc10

If you don't mind me saying, your game sounds a lot like Robocode...

Fair enough from a pure design perspective. In both games you do send tanks, squads, w/e in to war to fight against others, but the battle part of Nero is not the point.

Nero is designed so you can actively teach the robots what to do when they come across a wall, other enemies, or certain situations. All of the learning is done in real time so you really are teaching them. A great example is one we call bunker.

Basically in training mode, you set up a small 4 sided room with 2 openings on either side. In the center you place a flag (their objective) and an automatic firing turret facing one of the openings. Then you spawn your team outside of the building and watch them try to get to the flag through the opening only to get blasted away by the turret.

Eventually they realize that there is another opening (all on their own). They run to that side and grab the flag all in one piece; safe from the turret. Then you turn the turret around to face the other opening and eventually they learn to come in through the original door.

All of a sudden now when you play in battle mode, if they try to come into a room and get fired upon, they'll leave and try to find safer way in. Though it is more of a tech demo (displaying the rNEAT algorithm) then a game, comparing it to a game where you program your own AI manually instead of teaching is a little premature. =)

Large steps in self-learning AI are all over the place in the academic world, but with it being unreliable compared to FSM's (though many times NN get much better and interesting behaviors), the gaming world has not and will not take it seriously for a long time.

--Jason Hooten

[Dominia]

I completely agree. When I joined the project version 1.0 was about to be released and didn't have a chance to contribute on game design, namely making it fun.

The problem always with research projects in academia is everyone has their own motivations. The professor wants a research platform, the grad students want a self learning interface they can plug into other applications, the undergrads want a to do new research they can put on a resume, and a few undergrads (my self included) pushed to make a video game out of it.

In 2.0 (once it gets released) you'll see more 'game -like' features; territory mode, squad based combat (mixing of different teams with different behaviors), a expanded tutorial and LAN and server multi-player support. Still though it comes up short in my personal opinion of being a fleshed out game; like I said before it was not in everyone's best interest to make a game. It is times like these I wish I had gone to a gaming school lol.

Do I think there is a place rNEAT in the gaming and industrial world? Yes. Does it work as a stand alone product? Not really.

The foundation is there for rNEAT and other NNs that if used correctly, for a GUI system in a OS, quick time reactionary responses to player gun fights or even a general over all abstract enemy class that all other enemies inherite from which learns the gamers tendencies over time, would be interesting applcations the years to come.

Honestly I have been playing too long and I'm very tired to learning boss encounters and reverse engineering their FSN to beat a level. Learning algorithms and more importantly, randomness, 4tw ^^.

--Jason Hooten

[HagNasty]

I'm not sure but I think we may have used this a a example in one of my AI classes. Or at least a very early version of it.

You created robots who wandered, Put a target location and let them aim for it.

You would smite the ones who did not do the right thing and as a result only the smart survived in the 'gene pool'

You could build walls and stuff too to make it more harder.