On 12/11/2024 at 9:34 AM, brinx said:

I am working with AI/ML in some of my hobby projects

So in your experience then, which model would you use to train this? I think it would actually be an interesting experiment to run with XP even now. There are python bindings out there that can talk to the XP drefs. Python tends to be the language of choice for the AI/ML crowd, no? It could be used to both train and then run the flight model after training is complete.

23 hours ago, blingthinger said:

which model would you use to train this?

the model stays the same, the "magic" is in the "loss", "backpropogation" and perception functions: How the marrix weights move from wrong answers to right answers, and does so in a way that gives generalisable results.

Getting those working for a specific problem set is more art than science.

But once they are trained, chips like those being worked on tenstorrent are (very likely) going to make conventional computation for many problem sets obsolete imho, because dotproduct functions are sooo much faster than for/if.

Edited December 16, 20241 yr by mSparks

1 hour ago, mSparks said:

"magic" is in the "loss", "backpropogation"

This is all "model"

So which of the menagerie of those above options that are out there would one use for something like this? This could be proof-of-concept'ed in XP right now. We can set relatively detailed geometry info from PM and can run a plugin to extract 1D forces from XP at any operating point to create a training set.

1 hour ago, blingthinger said:

This is all "model"

So which of the menagerie of those above options that are out there would one use for something like this? This could be proof-of-concept'ed in XP right now. We can set relatively detailed geometry info from PM and can run a plugin to extract 1D forces from XP at any operating point to create a training set.

No, the loss and back propagation functions are only active during training to calculate the weights.

once trained a NN is universally calculated as output = dotproduct(weights,input), where "layers" are several of those chained together (i.e. output = dotproduct(weights1,dotproduct(weights2,input)) etc), aka a markov chain (which is why GPUs are good at NNs - they hardware accelerate dotproduct over 10s of thousands of cores).

Edited December 16, 20241 yr by mSparks

Im wondering is this is not just a, my sim is better than your sim because, topic, Its going no where though interesting watching heads butting. Glad im not that smart.

Not at all. For my part at least, this is wanting to see how well a generalized AI/ML-based flight model might perform. Nobody's using that right now so there's no spy vs. spy here. I'd be stunned if python could be used to control asobo2020/4. (Was wrong here. Looks like there is something out there but it does look like it's been abandoned?). XP doesn't complain. Python is the sword of choice in AI/ML these days. XP would be a perfect platform to test something like this.

Edited December 17, 20241 yr by blingthinger

12 hours ago, mSparks said:

Getting those working for a specific problem set is more art than science.

10 hours ago, mSparks said:

several of those chained together

Yea, it's precisely because of statements like these that I doubt that this will work as well as you and brinx claim. There is more than a baker's dozen of these models available out there and they all sometimes, kind of, maybe lead to a similar-ish result. You can't have that with a generalized flight model. Different combinations of convolutions and types of convolutions and weights and layers and etc. etc.

What resource can be used to sift through all this and identify what would be needed here?

Also, there's a very strange lag in your posts...weird. You show as posting that many hrs ago yet I only see it now? Odd avsim quirk I guess.

Edited December 17, 20241 yr by blingthinger

3 hours ago, blingthinger said:

Yea, it's precisely because of statements like these that I doubt that this will work as well as you and brinx claim.

way back when I referred to "fails to converge".

Im not really making a claim one way or another, until someone gets it working it doesn't work, and companies could invest several lifetimes of effort and never get it working.

But when they do work they are fairly miraculous, with the main benefit being they turn what would otherwise be highly serial math into something that is "embarrassingly parrallel". lots of problem sets are "solved" now- voice and image recognition have been phenomenal for a while, chatgpt shows what they can do with language

https://chatgpt.com/

If you haven't tried it.

fluid dynamics is a far more contained problem with more easily testable outputs, so I would tend towards thinking it is more easily solvable than not.

3 hours ago, blingthinger said:

Also, there's a very strange lag in your posts

avsim mods kindly proof read my posts for me lest I trigger the more sensitive folk.

3 hours ago, blingthinger said:

What resource can be used to sift through all this and identify what would be needed here?

Its still fairly secretive tbh, traces most of its modern history back to Alan Turing and baysian stats, Brits would generally be much further ahead but Churchill ordered all his research destroyed and had him chemically castrated, so most of what survived is from what he taught in the US and has been rediscovered since. A lot is probably going to remain secret because its also used for developing nuclear weapons.

Edited December 17, 20241 yr by mSparks

On 12/12/2024 at 11:12 AM, mSparks said:

I guess it would depend how you frame the problem. ...

My guess is that would probably work well  because its the same kind of problem set as ...

You're making statements like this. Sounds confident but I guess also unsure of how it would be done? Noted.

I've used them in engineering contexts to design hardware. Not fun little chatgpt playtests. I'm watching other proponents around me also keep saying "no, no wait, I promise this will work!"

I've never trained one and I have yet to see anything that would lead me to believe this will work for a general purpose flight sim the way you and brinx seem to think that it will. Especially not as a CFD replacement.

19 hours ago, blingthinger said:

Python is the sword of choice in AI/ML these days. XP would be a perfect platform to test something like this.

Yes, because it has this: https://xppython3.readthedocs.io/en/latest/

Shouldn't be too hard to get into, as the documentation is A+.

12 hours ago, blingthinger said:

Sounds confident but I guess also unsure of how it would be done? Noted.

I'm reasonably confident it would work, I'm not confident it would be worth the investment to get it working, at least not for anyone that isn't already selling/developing CFD software. for CFD software it for sure has the potential to realise decent accuracy "real time" CFD, for flight sim specifically tho, a lot of time and effort required for something that "at best" is as good as laminar Research already offer.

To better answer your earlier question about resources to get it working, I did mention pytorch already

https://pytorch.org/tutorials/

 

 

Edited December 18, 20241 yr by mSparks

11 hours ago, Bjoern said:

Yes, because it has this: https://xppython3.readthedocs.io/en/latest/

Shouldn't be too hard to get into, as the documentation is A+.

Problem is Python is not compiled 😕 but interpreted... just like LUA, always less effective than C or C++ with an optimized compiler.

 

Edited December 18, 20241 yr by jcomm

9 hours ago, mSparks said:

pytorch

That's the tool base. Not the recipe. It's guidance on the recipe that I am more curious about. I guess you're calling that the proprietary/secret bits which makes sense. What's a good resource to learn about those various bits and bobs?

There's no reason this couldn't be put to the test now without CFD. Remember what you'd get from CFD? Surface pressures? You'd end up summing that all up to a 1D number for each of the 6DOF motions. XP provides that at the dataref level. The proof of concept is to maybe use the simple airfoil geometry profile from the .afl file as a proxy for the CFD surface mesh and use maybe a hershey bar wing so that you just need chord and span (aspect ratio) to describe each of the lifting surfaces. It's a very high level geometric break down. And then train with control surface movements and resulting 1D forces, etc. also all available in the XP dataref list. Building the training set is just a plugin away. What to do with that set after it's been recorded, is the question.

Train it on a few wing geometries, e.g. aspect ratios 1,2,4,6, and then use it to fly a wing with aspect ratio 10 and a little forward sweep. See how it handles extrapolation on a highly non-linear space. You've said this will be easy as cake for a NN.

 

3 hours ago, jcomm said:

Problem is Python is not compiled 😕 but interpreted... just like LUA, always less effective than C or C++ with an optimized compiler.

 

I had this very same worry too. I think though that with the size of test my brain was conjuring, the optimizations baked into python's numpy library would probably be fast enough. It can be pretty quick for matrix operations and these might not be huge, huge matricies. Though I do wonder if pytorch could get access to the GPU from within a XP plugin...

Edited December 18, 20241 yr by blingthinger

4 hours ago, blingthinger said:

It's guidance on the recipe that I am more curious about. I guess you're calling that the proprietary/secret bits which makes sense. What's a good resource to learn about those various bits and bobs?

probably the nicest lecture on the background of the evolution of the mathematics involved I've seen is

Covers all the secret squirrel stuff very nicely to.

you can interpret the NN weights as the Bayes probabilities she is talking about in there (1=100% likely, -1= 100% unlikely), although there is a little more to it that just that.

bugs she talks about at the end is now here

http://bes-qsig.github.io/fge/docs/IntroWinBUGSwithR/

although pytorch is a lot more feature rich and includes hardware acceleration support (especially for those on OS's that lack even basic hardware acceleration like windows).

Edited December 18, 20241 yr by mSparks

13 hours ago, jcomm said:

Problem is Python is not compiled 😕 but interpreted... just like LUA, always less effective than C or C++ with an optimized compiler.

A bit of execution time penalty is totally worth not having to deal with the utter pile of BS that is anything C.