Fit a neural network
![]() |
|
![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() |
Fit a neural network
Steve Krug
Wisconsin Racing
Fine. But what does an ANN or an RNN bring to the table (read that STUDENT'S table) other than a model of the data ? Really, you want a model of the TIRE which was tested and have obtained a bag of data for it from a test facility.
May I point out that some of the TTC data has some questionable events and traits (Especially Round 5) that will be perfectly represented by an error minimizing NN. But is that what you want ? Or would you like a tire model that is easy to produce results for, fits PFG, ignores the blemishes and pimples from the testing process, can be factored into a vehicle model and used to predict it's stability, steering gain, max lat and that so wonderfully overstated 'yaw damping' metric ? All of this while looking hopelessly into a design judges blank stare resulting from a complete lack of understanding of the AI subject ? Extra points if your data model is common with the tire manufacturers. Otherwise there is an obvious translation cost. That won't sit well in closing arguments in an employment interview.
Or was this suggestion accompanied by a Flux Capacitor hack ?
I've found that using the cousin of a neural network, the genetic algorithm (followed by some gradient based cost minimisation), is a great method for determining the coefficient values for MF 5 onwards, provide the raw data is treated properly as per Mr Cobb. If you pay reasonable attention to the advice given in Pacejka's text, you can get get very reliable fits within the testing regime in relatively small computational time.
A big benefit of GA is that very large populations can be efficiently computed in parallel compared to gradient based methods (e.g. lsqcurvefit), as the algorithm doesn't need to dip out of the cost function solving to determine the slope. I'm no comp sci expert, but from my experience, gradient based methods are really good minimising the cost function within a particular n-dimensional well, while GA is quite good at identifying where all the wells are in that n-dimensional space.
OK, so what does Tesla use for setting tire and vehicle parameter specifications and submissions ? Is there actually a 'Tire & Wheel Release Group' empowered to analyze/synthesize, specify, integrate, validate and audit tire performance ? Or is this done by your tire suppliers ?
What models, Math products and testing methods are in place? This would be a good time to frame potential new Industry vehicle dynamics employees on how things get done as a template for FSAE project teams (IMHO).
Speaking on behalf of only myself and Wisconsin Racing:
Investing time into vehicle dynamics controls, sensor technology, and computer science seems like a good background for being involved in projects like FSAE-Electric, FSAE-Combustion, FS-Driverless, and dynamics analysis in industry.
Steve Krug
Wisconsin Racing
Dug out this thread from last year, I'll guess that some of you are working with Round 8 TTC data.
For those that don't follow XKCD, today's comic summarizes many of the problems in data analysis:
http://www.explainxkcd.com/wiki/inde..._Curve-Fitting
or the original at: https://xkcd.com/2048/