The online racing simulator
How complicated can a tire model become
(143 posts, started )
#126 - col
Quote from Mountaindewzilla :Better: "I don't need to know anything about physics to know when hammering something is a good solution to a problem".

ANNs are not hammers that you can use to beat problems into submission (or wood).
You have to understand some of the underlying theory in order to effectively use an ANN to solve any non-trivial problem. Prove me wrong.

Most important is an understanding of the non-trivial problem. e.g. If your roof is leaking, just going and hitting it with a hammer won't help. You need to understand how a roof works, and how slate works etc. When you have all that knowledge processed, you can use your hammer as a tool for fixing your roof - you still don't need to understand the physics of the hammer.
people, neural nets are good at recognising patterns. who ever suggested they were good at computing results to non-linear functions?

a good neural net learns whether to return 0 or 1 when asked "is there a letter A in this bitmap data?". how is that related to "what is the resultant force vector when the following objects come together under the following conditions?"
I'm done.
Quote from CarlLefrancois :Who ever suggested they were good at computing results to non-linear functions?

This question was addressed in a previous post. Check this out for more information.


col, the hammer metaphor is not valid. I replaced your invalid metaphor with a similarly invalid but more appropriate metaphor. I should have been more clear about what I was trying to do there.

You didn't address the crux of my argument.

Does arguing with people online make you feel like a Bad Mother****er?
It makes me feel like a pretentious douche, so I'm going to stop here. Cheers.
#129 - col
Quote from Mountaindewzilla :

You didn't address the crux of my argument.

Does arguing with people online make you feel like a Bad Mother****er?
It makes me feel like a pretentious douche, so I'm going to stop here. Cheers.

Hmm, the crux of your argument seems to be calling me names. I was trying to ignore that and continue the tyre physics discussion. It's a shame you're intent on pursuing the ad hominem angle.
#130 - col
Quote from Mountaindewzilla :
col, the hammer metaphor is not valid.

Yes it is!

I clicked your 'let me google that for you' link and the second result includes various comments the back up my stance here. Both that ANNs are not good for modelling complex systems due to efficiency issues, and that you don't need to understand them to use them:

I've emboldened the pertinent sections.
Quote :
A. K. Dewdney, a former Scientific American columnist, wrote in 1997, "Although neural nets do solve a few toy problems, their powers of computation are so limited that I am surprised anyone takes them seriously as a general problem-solving tool." (Dewdney, p. 82)
Arguments for Dewdney's position are that to implement large and effective software neural networks, much processing and storage resources need to be committed. While the brain has hardware tailored to the task of processing signals through a graph of neurons, simulating even a most simplified form on Von Neumann technology may compel a neural network designer to fill many millions of database rows for its connections – which can consume vast amounts of computer memory and hard disk space. Furthermore, the designer of neural network systems will often need to simulate the transmission of signals through many of these connections and their associated neurons – which must often be matched with incredible amounts of CPU processing power and time. While neural networks often yield effective programs, they too often do so at the cost of efficiency (they tend to consume considerable amounts of time and money).
Arguments against Dewdney's position are that neural nets have been successfully used to solve many complex and diverse tasks, ranging from autonomously flying aircraft[46] to detecting credit card fraud[citation needed].
Technology writer Roger Bridgman commented on Dewdney's statements about neural nets:
Neural networks, for instance, are in the dock not only because they have been hyped to high heaven, (what hasn't?) but also because you could create a successful net without understanding how it worked: the bunch of numbers that captures its behaviour would in all probability be "an opaque, unreadable table...valueless as a scientific resource".
In spite of his emphatic declaration that science is not technology, Dewdney seems here to pillory most of those devising them are just trying to be good engineers. neural nets as bad science when An unreadable table that a useful machine could read would still be well worth having.[47]

Quote from Mountaindewzilla :This question was addressed in a previous post. Check this out for more information.

i don't think you understand what neural nets are for.

just because google gives you search results with "neural net" and "nonlinear" in the same article doesn't mean what you think it means.
Quote from col :I got the point. And I accepted it.

I beg to differ - I didn't mean that the state should contain a copy of the history, but that the history will change the state in such a manner that you don't need lug it around, much less feed it into the net.

You're telling me that the ball's position over the last x frames must be used to calcute the next position, I'm telling you to just store the current speed in the state and do away with all that calculation.

Quote from col :I've emboldened the pertinent sections.

I notice you didn't embolden "neural nets have been successfully used to solve many complex and diverse tasks".

Quote from CarlLefrancois :i don't think you understand what neural nets are for ... people, neural nets are good at recognising patterns. who ever suggested they were good at computing results to non-linear functions?

NNs are highly suited to reducing complex interactions between numerous inputs into relatively few outputs.
Quote from Racon :NNs are highly suited to reducing complex interactions between numerous inputs into relatively few outputs.

yes. my argument is that the outputs generated by neural nets are not the floating-point vector results of complex physical interactions.
Well, you can do floats in ANNs if you want to...
#135 - col
Quote from Keling :Well, you can do floats in ANNs if you want to...

Or vectors!

You could rig up an ANN to give you vectors as well, you could theoretically rig up an ANN to solve most problems.
The issue here is that they are only _good_ at solving a limited set of problems.

There's no point in using Neural Nets in areas where they would be hopelessly inefficient or otherwise impractical. You would choose a more appropriate tool for the job.
Quote from CarlLefrancois :it's the equivalent of plotting the eventual path of the eye of a hurricane given (many computations...), etc. the trick is in reducing these incalculable bodies to points with minimal additional descriptive parameters, and making these bodies appear and disappear over time as their influence becomes more and less important to the result

http://www.youtube.com/watch?v=n_hiDuaZ_0Y

note from 1:30 to 2:00 how points appear and disappear.

the simulation undertaken in this video is of another order of computation cost than the one LFS can afford for tire rubber calculations, but the concept of maintaining a set of the most influential points could be useful.
Is tire physics a finished problem? I majored in math at the university but that's a long ways from really understanding physics. Well, I tried looking around since I can sort of read the math and I remember seeing one really old paper from around the 70s (nothing newer seemed to show up on Google) about how, this was unsolved.

What I mean is, areas of physics that are well understood show up on Wikipedia as well understood, tires don't seem to be done. Of course, it's entirely possible that the Pirelli engineers consider this a finished problem but I'm sort of doubtful considering the way the tires were graining in Austria.
No it's not. Even with a lot of HPC power and plenty of time.

The fundamentals of the dynamic behavior of the materials involved remains kinda mysterious atm. Math/HPC alone can't do much.
Quote from rowdog :Is tire physics a finished problem?

from the Milliken book i'm reading (race car vehicle dynamics) it seems like the state of the art technology is to take a tire and scrub it against a moving belt at various angles and speeds and come up with tables of static data points. plotting these points can give you a bunch of curves for different variables, e.g. lateral force vs slip angle.

i believe LFS uses a set of fitted polynomials made to match these curves.

as Keling implies, this totally misses any subtle nonlinear effects. in reality the traction obtained by the tire is based on a lot of dynamic effects and what we're missing now is how the traction changes as the tire moves from one angle to another, or goes from over the limit back onto the limit.

the static data point interpolation gives a useful estimate but to get closer we need to simulate the dynamics. one example of something that needs simulating is the ridge of pressure that exists in a small part of the contact patch and that makes up for a lot of the lateral force that is generated. this ridge moves around as you switch from accelerating to braking and otherwise go from one extreme to the other.
Quote from CarlLefrancois :from the Milliken book i'm reading (race car vehicle dynamics) it seems like the state of the art technology is to take a tire and scrub it against a moving belt at various angles and speeds and come up with tables of static data points.

This is how you get your reference data, yes.

Quote from CarlLefrancois :i believe LFS uses a set of fitted polynomials made to match these curves.

That's generally how empirical models work, for example Pacejka's Magic Formula (though that's a mixture of trigonometric curves with linear or polynomial modifiers). LFS as released uses Scawen's custom implementation of an empirical model.

State of the art these are not though, and this is where more physically based modelling comes in (e.g. brush, Dugoff, finite element, etc).
It might (probably not) be good idea to use neural nets to calculate some smaller more specific parts of tire physics. But creating a huge net, which inputs stuff like velocity, position, temperature, etc... and outputs stuff like... contact frictions?, sounds quite silly to me. Probably it is possible, sure, why not, but slower than the current test bench algo.

Also lookup tables can be small part of the solution.. but not the whole thing, this is not NFS where you fetch the friction coefficient from LUT by angle of attack. It doesn't mean that anything can't be acquired from LUT. Maybe some minor stuff, which adds some spice to the soup, but would be too heavy to calculate realtime.
RCVD is quite old, but the method is still standard. Whatever your type model is, this kind of data is still used as reference.

But with more physics based (or should I just say more detailed?) models you can't utilize the measurement directly. You have to inverse a set of model parameters.
Quote from Bob Smith :State of the art these are not though, and this is where more physically based modelling comes in (e.g. brush, Dugoff, finite element, etc).

thanks for the clarification, it helps raise my level of knowledge above "next to nothing"

Quote from Keling :But with more physics based (or should I just say more detailed?) models you can't utilize the measurement directly. You have to inverse a set of model parameters.

what does it mean to inverse a set of model parameters?


what i mean by "simulating the dynamics" instead of doing interpolation of static data points is keeping track of a few points in the carcass along the circumference of the tire and giving them some basic interactions among each other.

the idea would be to attempt to capture some of the subtleties that get missed by summing the forces they represent into one vector.


the neural net idea isn't bad either. i assume it would be possible to train one to take all input torques and current state and predict if the tire is currently exhibiting non-linear behaviour. a few of these nets might allow the simulation to switch between members in a family of functions.

How complicated can a tire model become
(143 posts, started )
FGED GREDG RDFGDR GSFDG