dialtone wrote: ↑05 May 2022, 20:44
PlatinumZealot wrote: ↑05 May 2022, 19:16
matteosc wrote: ↑05 May 2022, 18:07
Agree. ML needs a set of inputs and outputs to be "trained". If you do not have the data you do not go anywhere. Also this is not the kind of problem you want to feed to a AI.
So Mercedes don't have any data for a machine learning model after two months and 4 races. Hmm OK.
Let's just move on then to other developments.
I think we're bordering on the OT here but... Driving 60 laps with the exact same setup is useless for ML. When you build a model, you need to setup the car in many different ways and then collect enough data on each setup so that, whatever computer model you have, it can lookup the behavior of the system at a given setup.
Once you have enough setups collected, and metrics on each of them, you should end up with a decent range of parameters to be able to run regressions on the data and generalize the model of how your suspension works.
At this point, even with all this data and model, you still need to know what you want to accomplish as you can't just type: "how to solve porpoising" in the computer and it will figure it out. First Merc needs to find the trigger of said behavior or they need to come up with the exact behavior that they want the suspension to work, and actually know (otherwise it's a bet) that it will solve porpoising.
Once you have the final model, and all of the setup plus its data, only then you have barely enough information to digitally make a new suspension, which then needs to be validated for manufacturing, weight, size and so on. And last you can test it in the car and if that doesn't work you are free to figure out what went wrong in any of the previous steps.
And keep in mind that anyway the computer needs to know what are its degrees of freedom when designing the suspension anyway, so if whoever is setting the experiment up misses a key freedom in the design, but chooses to constrain it in a particular way, you will end up with no possibility of a design that solves your issue.
All in all I doubt any team is actually doing this, what's more likely is instead that the ML system is going to tell you what is the best setup given everything that you have, as is. ML aided design is pretty limited for this stuff IMHO.
