Support NeoGAF

[Lathentar]

At the University of Texas, I had the joy of working on a proof of concept video game which demonstrates a new use of Neural Network technology in video games. Having agents adapt in real-time right before your eyes has never really been done. Today, the first demo of the game has been released.

Thanks guys, don't be too harsh on it, a bunch of college students worked on this! I'd love to see what kind of behavior you can come up with!

Download the Demo at www.nerogame.org

Description: NERO is the result of a joint project between the Digital Media Collaboratory (DMC) and the neuroevolution group at the Department of Computer Sciences at the University of Texas at Austin (UTCS). NERO utilizes state-of-the-art academic artificial intelligence research in order to demonstrate its effectiveness in a game environment. The game uses AI to allow simulated robotic agents to cope with changing environments and situations, and form adaptive tactical solutions. The end result is a game that adapts itself to the strategies desired by the player, while still allowing the AI-controlled entities to operate as autonomous agents. NERO introduces a new genre of video game that is only possible with machine learning technology: The player takes the role of trainer and teaches a team of novice soldiers the skills for battle. When sufficient skills have been acquired, players can take their teams to battle mode where their skills are tested against teams trained by other players.

 

[pj]

At the University of Texas, I had the joy of working on a proof of concept video game which demonstrates a new use of Neural Network technology in video games. Having agents adapt in real-time right before your eyes has never really been done. Today, the first demo of the game has been released.

Thanks guys, don't be too harsh on it, a bunch of college students worked on this! I'd love to see what kind of behavior you can come up with!

Download the Demo at www.nerogame.org

Description: NERO is the result of a joint project between the Digital Media Collaboratory (DMC) and the neuroevolution group at the Department of Computer Sciences at the University of Texas at Austin (UTCS). NERO utilizes state-of-the-art academic artificial intelligence research in order to demonstrate its effectiveness in a game environment. The game uses AI to allow simulated robotic agents to cope with changing environments and situations, and form adaptive tactical solutions. The end result is a game that adapts itself to the strategies desired by the player, while still allowing the AI-controlled entities to operate as autonomous agents. NERO introduces a new genre of video game that is only possible with machine learning technology: The player takes the role of trainer and teaches a team of novice soldiers the skills for battle. When sufficient skills have been acquired, players can take their teams to battle mode where their skills are tested against teams trained by other players.

No normal mapping, no sale


jk, I'll download it and screw around even though I'm not entirely sure wtf it is..

Edit: Word of advice: If you want more than one reply, keep the word "learning" out of the title, and try to throw some controversy in there. "HALO 2's AI SUCKS SHIT COMPARED TO THIS"

 

[Ghost]

I was thinking about doing something similar to this, what kind of neural network are you using?



Defintely a good proof of concept

 

[Lathentar]

I was thinking about doing something similar to this, what kind of neural network are you using?



Defintely a good proof of concept

Using real-time NeuroEvolution of Augmenting Topologies for the learning.

You can find papers on the concept at by googling rtNEAT.

 

[Zaptruder]

We're lazy bastards, so can you give us a laymans explanation for what this technology exactly does and give us some examples of how it's better than other touted adaptive AI systems?

 

[Matlock]

Too bad there ain't an uninstaller.

Cool game, tho.

 

[FiRez]

Looks too heavy taxing for my pc

such a shame

 

[JackFrost2012]

To paraphrase Mark Twain, "People don't want a thinking computer, they want to THINK the computer's thinking." Actually intelligent and adaptive AI will kill humans dead everytime (c.f. Terminator) Still, good job trying this.

 

[Argyle]

Kinda reminds me of...

 

[Lathentar]

We're lazy bastards, so can you give us a laymans explanation for what this technology exactly does and give us some examples of how it's better than other touted adaptive AI systems?

This technology is the most efficent neural network learning algorithm out there and allows for real-time adaptation. Just do a simple demo like making the enemies run toward a flag. There is no path finding there is nothing...

The only things these guys know how to do it fire, turn left and right and move forward and backward. That is IT. Over the course of training they can learn to do so much more and you can witness them learning. You can train them to fight how you want and pit them against other players.

It's crazy what kind of things the agents can figure out. Bizarre dodge tactics and teamwork is not uncommon.

 

[Lathentar]

To paraphrase Mark Twain, "People don't want a thinking computer, they want to THINK the computer's thinking." Actually intelligent and adaptive AI will kill humans dead everytime (c.f. Terminator) Still, good job trying this.

I know, its the biggest problem currently.

However, if you allow AI to adapt imagine an MMORPG where the AI adapts to your tactics, but never actually gets stronger. You can't use the same cheap trick over and over again.

 

[Chrono]

Sounds very cool. You're a graduate student?

 

[sprsk]

ever play vf4?

 

[Parallax Scroll]

Just installed it, but I have to head off to work. I'll try it out later today.

 

[Lathentar]

ever play vf4?

I've heard VF4 adapts to your play. I'm guessing they use a Bayesian Logic for learning.

 

[Parallax Scroll]

I've been playing around with this for a bit. It's definitely cool to see the troops get better at something over the course of several minutes. They seem to forget a lot of things pretty quickly, though.

For example, I'll set up a target so they have to run clockwise around an obstacle to reach it. They fumble around for a while until they figure this out. Then I switch it up and make them run counter-clockwise. They have to learn this all over again. Fair enough, since those two situaitons are not identical. Once they get good at the counter-clockwise motion, I switch to clockwise again, and they're no better at it then when I started.

 

[BugCatcher]

Having agents adapt in real-time right before your eyes has never really been done.

Actually, agents that adapt in real-time have been around for quite a while. I remember being fascinated by Evolve! Lite 10 years ago (download), watching pixels battle it out and change behavior over time. The learning AI in Galapagos was supposed to be pretty revolutionary at the time, although I hear the game itself was pretty frustrating. Neither used neural networks, though. Black & White's creatures did, however, but the way it learned isn't really comparable to what you're trying to do here.

Neural networks are fun, but I still think it's going to be a long time before they become feasible in games. The biggest problem, as I see it, is that no matter how complex your network becomes, agents can still only be trained to excel at one task at a time. They can adapt if they need to perform a new task to succeed, but then they're only trained for that, and lose their ability to handle the original situation. It's akin to physical evolution; animals can evolve to adapt to changing environments, but that doesn't make them any more intelligent. It doesn't have anything to do with intelligence, really, which is why I think it's kind of inaccurate to talk about evolution, mutation algorithms and fitness functions as part of AI. It's more like artificial biology. Don't get me wrong; NNs are cool, and this is interesting research, but there's still a lot of work to be done before these kinds of agents can work in realtime games.

 

[StRaNgE]

Sounds like the early stages of skynet and countdown till self awarness.

T2's one day becomes a reality. run John Conner, RUN!