Hello our current project is an arcade racing game. The simulation of the cars is physics based and practically impossible to predict. Right now the AI just tries to follow a target point that get's pulled along the race track in front of the car. I tweaked this system a lot to make the AI faster and more competitive, but now I'm in doubt that will ever work. Especially since we added a track with very little grip where you have to slide through the corners, my approach fails badly. That's why I want to try neural networks to train the AI on every track. My hope is that someone can recommend an easy to use library to me where I can setup and train a NN with just a few lines of code. I don't have practical experience with NNs. I think I need two input variables which indicate the car's position on the track, and three output variables for steering, throttle and brakes. Would this work? And what happens if the car spins, get's stuck or dynamically placed obstacles get in the way? My current approach can handle these situations (well it tries at least), so in such cases the AI would be able to fall back. I hope these are not too many questions. I'm primarily looking for a simple to use library so that I can instantly experiment a bit, instead of trying to get back-propagation to work for a week.

Neural networks *could* work, but i'm not sure that they're the best tool for this job. racing car AI has been around forever (at least since pole position on the atari) and I'm not aware of any commercial release that used neural networks. It'd be pretty hassling to set one up and train it to drive; and would be almost impossible to tune. NNs aren't magic AI entites that can learn anything and perform well at all tasks; think of them more as pattern recognizers or classifiers. There are elements of pattern recognition and classification in a race situation, but I think it's better solved with a little pre-computed or designer placed data.

You should check out the Article & Resources section of this site. There is a whole series of articles about writing a race-car AI system.

-me

You might find this article interesting:
http://www.ai-junkie.com/misc/hannan/hannan.html

Thank you for your replies.

Well I thought about the problem again. Until now I refused to implement an optimal racing line, but I guess it would help quite a bit.
Maybe that will even be sufficiently good on tarmac tracks. However, I don't see how that could work on gravel tracks, where really aggressive sliding is required to be fast.
The interview mentioned by viblo deals with this problem, but it's very vague.
So I think I'll just have to try and find out by myself. In the other thread about Neural Nets "FANN" (http://leenissen.dk/fann/) is mentioned, and from their tutorial page it looks like it's just what I searched for.

AI is a lot more work than I expected but at least it's interesting, too :)

Quote:Original post by vokuhila
Thank you for your replies.

Well I thought about the problem again. Until now I refused to implement an optimal racing line, but I guess it would help quite a bit.
Maybe that will even be sufficiently good on tarmac tracks. However, I don't see how that could work on gravel tracks, where really aggressive sliding is required to be fast.
The interview mentioned by viblo deals with this problem, but it's very vague.
So I think I'll just have to try and find out by myself. In the other thread about Neural Nets "FANN" (http://leenissen.dk/fann/) is mentioned, and from their tutorial page it looks like it's just what I searched for.

AI is a lot more work than I expected but at least it's interesting, too :)

If you are unable to describe the behavior you want, it is unlikely that you will be able to make any method works, machine learning even less.

Take some time to think and write the behavior you want to see. You want it to run ahead in lines, break before curves (what is each curve ideal entry speed?) slide in gravel (what are the mechanics of sliding? Facing vector vs speed vs velocity?), avoid other cars, pass other cars...

Only after you know exactly what you want, will it be time to choose a method to implement these behaviors.

Hello Steadtler

I thought about what you wrote, and my conclusion was that I'm not able to write down rules that would let the AI drive fast enough. It's just too complicated because it depends on many parameters. Every turn on every track is different. They have different ground materials, different decline, some have to be cut, some must not be cut. etc. etc.
So the strange thing is that I perfectly know how to drive around a track, but I cannot formulate rules which describe how to drive.

That's why I implemented a system today which simply collects data while I'm driving along the track:
- positon of the car
- velocity
- heading vector

The AI simply tries to make its car imitate these parameters. This works much better than I expected. The AI can slide through corners nicely and is just 2% slower than me which means that it's almost impossible to catch up if you are a bit behind :-)

So in the end, I used a neuronal net, but not an artificial one. I drove arround the tracks hundreds of times, training the thing between my ears without any overfitting or generalisation problems. And now I just need to harvest the results.
Hm. Pretty easy and probably done a thousand times before, but I'm not that much familiar with AI techniques.

Thanks all. I think with some more tuning my problem is solved.

Quote:Original post by vokuhila

That's why I implemented a system today which simply collects data while I'm driving along the track:
- positon of the car
- velocity
- heading vector

The AI simply tries to make its car imitate these parameters.

Thats pretty much the same as following a well-defined raceline, dont you think?

Quote:Original post by vokuhila
I thought about what you wrote, and my conclusion was that I'm not able to write down rules that would let the AI drive fast enough. It's just too complicated because it depends on many parameters. Every turn on every track is different. They have different ground materials, different decline, some have to be cut, some must not be cut. etc. etc.
So the strange thing is that I perfectly know how to drive around a track, but I cannot formulate rules which describe how to drive.

That's because you're 'over-thinking' the problem. A very simple rule for turning a vehicle through an arc of any radius: apply a force to the vehicle directed radially inward, with sufficient magnitude to achieve the desired turn rate.

Because you want the vehicle to rotate as it changes velocity, you need to ensure that you generate torque on the vehicle body (that's why we have steerable wheels at one or both ends of a vehicle, rather than in the middle). If you have a tight turn, then the radial force generated by the steerable wheels is insufficient and you must align the drive wheels with the direction of resultant force required (remembering that there are two components... the radial component for cornering and the tangential component for overcoming friction and maintaining tangential speed of the vehicle). You can do this by initiating a skid and using the steerable wheels to control the angle of the skid (and hence the angle of force application on the road)... and the drive wheels to control the magnitude of the applied force. The two together give you the applied force vector. For any turn its fairly easy to determine the turning force required given the entry point and forward speed at that point, mass and rotational inertia of the vehicle and the friction properties of the road surface.

As for a controller for this problem: a simple bang-bang control scheme on the steering wheel and constant drive wheel speed should do the trick.

Cheers,

Timkin

1. Those who understand the tradeoffs between NNs, bayesian networks, naive bayesian networks, ridge regression, linear regression, logistic regression, etc.

2. Those who think that "neural network" sounds all cool and brainy.

One of these groups should use NNs where they're appropriate. The other group needs to pick up a textbook on artificial intelligence, and another on machine learning.

Now, neural networks are actually potentially not a bad tool for racing AI. But they're probably not the best tool for a job. Until you understand NNs in a ML context you really shouldn't be prescribing them for yourself.