The real time game code has to work on not only tyre physics, but other stuff as well. (With tyre physics getting more detailed, the other aspects of physics should improve as well.) It's more than 400.
Ooops. My eyes must have automatically skipped some words for me.
Speaking of the non-real-time model, not sure how paralleled the implementation was. If it was single threaded, with all those 4/6/8-core or HSA/hUMA stuff happening the gap might be narrowed, assuming that LFS devs (or whoever planning to use a similar design) do code for those hardware.
IIRC, low-end capability is still a high priority. I'm still of the opinion that a little venture into the world of neural networks might find a better way to approximate the bench quickly enough than just elbow grease and cunning... it could even be made into a collaborative effort or competition with a release of a set of test data and a time limit.
If you can translate a complex multi-dimentional non-linear physics problem into the domain of simple pattern matching and machine learning tools, then you might get some results from artificial neural networks...
Kinda like trying to service an F1 transmission using a glue gun, a roll of duck tape and a spork.
I said they were simple tools. You can use simple tools to solve complex problems - pattern recognition for example.
I think ANN's are not a good solution for real-time tyre modelling on consumer computing devices!
And now I hope that you are going to explain in technical detail how it can be done, and I will learn something useful. Otherwise, stop blowing hot air
"Plug the inputs into the input layer; read the outputs from the output layer; train the middle layers with data from bench" - something like that? Honestly, I don't see where you think the problem is.
Actually there are models that are specifically designed to address non-linear, multidimensional problems. The range of applications is very broad.
Nothing with that many syllables (eight?) can be that simple. I tried to read some studies where AANs were used, and most of the math was over my head (for now). I'm guessing that most people around here would have a similar experience.
Your pretentious antics push my buttons.
To get back to Racon's question:
I'm inclined to agree with col, but I couldn't come to any conclusions in a reasonable amount of time. My guess is that ANNs aren't so great for dynamic systems unless you have a huge amount of compute power at your disposal.
I consider a hammer to be a simple tool, and yet I would probably struggle with the math in a metallurgical analysis of the tempering process of steel.
Agreed - you put it better than me.
I think one of the main issues is that for a system like tyre physics, the NN cannot just look at the (significant number of) inputs at any single instance in time, it would need to process multiple sampling points on a time axis long enough to handle the 'impulse response' of the system being modelled, and at a high enough sampling frequency to represent the highest frequencies present in the system without aliasing. I would guess that if you take the number of time frames required, multiplied by the number of other inputs required for a detailed multi-dimentional model, you get a very big number. When you consider that each of these inputs must be connected to every single node of the ANNs hidden layer by a sigmoid function, then the processing requirements start to become astronomical.
Impulse response is an output, not an input. When you model a ball being dropped to the ground, for example, the speed of the ball changes over time - but we don't use the changing speed as an input. The inputs are the current speed and the acceleration due to gravity (etc), the outputs from the calculations are the new speed and the position. We've made the model produce the changing speed.
If you want to make a very simple model of a ball falling, and are happy to pretend your ball is an abstract, infinitely solid perfect sphere then that's fine. Unfortunately, that's not good enough for tyre physics in a racing sim.
I suppose there are various ways to go about it, but I don't see how you can ignore the fact that there is feedback in the system. If you want an ANN to model this without responding to the history of the system in some way, then you need an ANN with feedback and you need to divide the model up spatially with a higher granularity. You'll also need to run the sim at a much higher frequency.
Unfortunately, that doesn't solve the problem of the processing requirements being too heavy, it just moves the processing from one domain into another.
And as finite element analysis is a well explored technology, I doubt you'd get your spatially divided ANN version working as efficiently as Scawens existing FEA bench model if you went that route.
I'm not an expert in this area, so there must be alternatives that I'm not aware of.
Have you thought of a way to accurately model tyre physics using ANNs that can provide the high frequency feedback and dynamic nature of the system without finely grained input of some sort?
How would you reduce the complexity of the model without losing accuracy and realism?
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.
It seems like your primary interest here is winning arguments. I think you'd find a debate forum more enjoyable than this place, especially if the focus is on stuff that relates to your experience and expertise. Should Google around.
I got the point. And I accepted it. Then I pointed out that you would still need the heavy computation, you just moved it from time into space.
If the ball is not infinitely hard, then to model e.g. deformation and have the history encoded in the state, you need a finely grained FEA style model. Which to realize in an ANN would mean lots of nodes and lots and lots of connections.
And you need to run it very fast in order to model the high frequency behaviour of the system.