Automobilista 2 V1.4.3.3 & Racin´ USA Pt3 Officially Released!

I have been getting those with other Steam titles recently, so I am guessing a (temporary) Steam infrastructure issue.

 

Shhh.. Don't say it too loud. Someone may mark you as a fan boy who refuses the only true religion in Simracingland...

 

Even a minimal knowledge about how neural networks(NN) is used does reveal that the phase where a lot of computer power is neded is only in the learning/training phase of the network.
So when Sony in my link above does describe how ressource demanding this training period of Sophy was then it has nothing to do with the version of the NN that can/will be used in a commercial game.

This is the difference between a dynamic and a static NN.

The ressource heavy training of the dynamic NN ends up in some kind of static NN "algoritm" where the rules are hidden for the human eyes - because they are created under the training process by the dynamic NN itself.
The advanced learning skills of the dynamic NN is not needed in the commercial end product.
This is the reason it only does show lack of knowledge about the subject to believe that todays PCs is only able to run one or two NN AI bots at a time because this will demands so huge ressources.
The demands of running a static NN "algoritm" isnt much different than running the unintelligent AI algoritms most commercial sims of today is cheating the users to use.

 

NN inference generally still needs Tensor accelerators (GPU or otherwise). Training is more expensive, but inference is not a cheap CPU task and certainly not for a game requiring a v.high decision rate.

As discussed earlier, using such compute expensive models to help refine/define A.i. paths is one thing, but I don't believe anyone is going to put a NN model into a game at this point. Possibly (but still highly unlikely) a traditional ML model. If you used the model to define paths 'offline' you could possibly build a lookup table on which decisions to make in certain circumstances with the available paths identified by the model.

Or conceivably you could run inference on only the closest 2-3 cars and it would be used solely to pick a line/path out of options that have been pre-chosen by the model offline. However even then it would need to be a very small NN for a CPU to run inference at a decent tick rate.

If you have evidence to the contrary please supply, I'd be surprised if iRacing has an actual model and inference engine in-game.

 

Your description is a more detailed version of what I mentioned earlier is used by iRacing. They have the advantage of thousands and thousands of tracked/traced laps to use as the basis for the "multiple choices." This matches my understanding of the server-based, machine-learning process.

No, it is not some super intelligence embedded into each AI vehicle, nor would it matter whether the AI vehicles were using super-advanced or more moderate physics to control their moment-by-moment motions @donaldd.

 

After a mix of online sessions, watching replays, test day sessions and a 5 lap AI race, this is how my core usage looks like. Some cores maximized at some times, others not so much, and the average only looks higher in just some threads. Still, I think that an app that presents an historical graph of each value would help to pinpoint things better. Which one could I use?

 

HWiNFO has a logging function which creates a .csv file, and there's a viewer available for it which draws graphs from the .csv files.

 

TY, just tried the feature. Will try to remember starting a log next time before opening the sim

 

I logged a CSV with HWinfo, using the Formula V10 Gen 2 at Indianapolis Road Course. First, a Test Day. Then, a 5 lap 21 car AI race. Then, the same AI race as I found issues with damage being overly sensitive and had to go back to main menu to turn it down.

The timeline is in minutes. You can see on the first thread (top left) when the game is running, as then is where it's being used near max. The rest of the graphs relate to each thread, two per core, 12 in total.

Check what's happening with the rest of the threads: most of them have little activity when one is on track racing. They all spark into life when loading, but go back to sleep when that's done.

AFAIK, this does not look like a fully multithreaded software: there are 4 threads working of 12 available, and only one is properly loaded.

 

Now do this with Raceroom,you will be surprised. (or deflated)

 

Any chance we can get the decimal points restored in the seat adjustment messages.
eg. Previously would show 1.2 etc.... no it's just whole numbers.
There's a lot of movement while it stays on 0.

 

I'd even prefer to have them snap to a number instead of the current fluid adjustments. Doesn't tilt work this way? I may be wrong. Anyways, maybe make the steps smaller if you ever change it.

 

Will be there some more ovals in the future?
Maybe a Tri-Oval, a kind which is missing besides the two D-shape ovals (Fontana + Daytona), Quad-oval (Indianapolis) and the more/less standard oval (Gateway).
I would prefer the Tri-oval of Nazareth - it's something special

 

So you want the Jesus oval?

 

Is there any game that uses all of the cores available?

Multithreading is a bitch to deal with, at least that was my impression back when I had to write multithread code at Uni.

 

This could have something to do with it

maybe the madness engine is optimised to consoles so doesn’t use all cores and threads

 

I can't imagine it being easy to take a released game running on an engine that someone else built and turning it into one that does multithreading on a level where it's worth it (assuming they have the rights to do that). Not that I'm a coder.

Not to mention that if they managed to make the multithreading work like a dream and the AI suddenly had all the computational capacity you could ever wish for, that would mean another full pass to calibrate it for all the different car/track/weather combinations with its newfound capabilities.

 

Any idea how the Live For Speed AI was learning tracks on their own 20 years ago? Honest question.

 

Daytona is a tri-oval.

 

That’s what I thought