Automobilista 2 September 2024 Development Update - The Road to V1.6 Pt2

Unlisted, to be exact - if you can find the link, you can still see it

 

But why do I have to do this?
Is the AI really learning something during a pre-race session? Do they need to adjust themself to the car & track & other parameters?
Isn‘t it just hard coded behavior for every car & track combination?
I can‘t really imagine it‘s the first thing.

 

At least by deleting the stream it shows that the quality required wasn't here

 

I have taken the time to watch several races in different categories and the AI works more realistically in long races, in those of less than 30 minutes they tend to get stuck in groups of several cars with clumsy movements.
I notice that the car that goes along the ideal line usually wins 90% of the duels, which makes it difficult to pass.
To really see how the AI works you have to do the classification because it helps to organize the cars in their natural racing positions and see it from the outside with a TV camera, and the truth is that it works much better than what is seen in the videos that have published, always talking about long races, I take the trouble to choose circuits that are in the championships set by Reiza and I respect the laps they set and there the AI is much better.
I am quite optimistic about 1.6, I don't see much bad things, I hope it comes out soon, greetings.

 

I think part of the reason people are picking up on the grouping is that in a previous version it very much seemed to be issue. You would get 3 AI cars running together, jockeying for position but not doing much, then a gap, and then another group doing the same seemingly all the way down the field. The only time the behaviour seemed to change was when you were in amongst them.

 

You should do it because without qualifying a slow driver could start in the front row and leads to more skilled drivers being held up by "him". If you do a qualy session this won't happen.

 

Yep, it seems to be the most effective way of getting the best distribution of AI drivers by skill/performance across the field

 

You don't have to do it, if you don't mind the AI tripping over themselves a bit. Whatever the ins and outs behind their programming, this is as recommended by Renato himself many many times - run a qualifying session that allows all cars to run at least five laps so they can sort themselves out into pace order, which in turn leads to tidier races.

 

It's mainly so AI sort themselves out into a realistic pace order factoring in all the available things that affect it, like their skills and the car ability and suitability to that track.

However, in AC yes the AI make setup tweaks during qualifying to amend their fuel limits (they calculate how much is needed for their average lap) and final drive ratios - I'm not sure if they adjust more than that. Frustratingly, that's not saved session to session. Not sure if AMS2 has any of that happening.

 

That's something that would interest me too! I mean, how much "real" AI is in the AI, i.e. is it really using current methods of ML? Geoff Crammond's GP series had excellent AI, and he had certainly never heard of neural networks.


Modern ML worflows consist of (essentially and very simplified) two parts.
- Training the model
- Using the trained model

Are the devs really training an AI model? Is this technology already the basis of sim AI? Or are we still "only" talking about (complex) "algorithms"? The latter would never learn anything, they do exactly what they are supposed to do, predictably. Of course, you can artificially build in some random parameters to make it look "human".

But "real" AI, based on ML, could learn from the player, but it would have to do this separately on each PC/console. That would no longer be comparable. Technologically, it might even be possible, because most of our GPUs now have the ability to handle tensors, which is already used for anti-aliasing (although the model is trained by the developer).

Does anyone here have more detailed information?

 

Ok, I understand the point, that after Qualification the AI field is sorted by their skill.
When I look at a custom Ai file, I see that the skill gets entered in plain numbers.
So why do a Qualification to do the sorting? Just let them start in order of their skill level.

Rfactor2 had the same thing going for years and Studio 397 „fixed“ this before moving on to LMU.

 

I don't believe there's any sign of this in sim racing yet, it's algorithm based. It makes sense for it to logically head that way in future, but it's not the click of a button to implement for sure.

Because a skill level is only one variable. And AMS2 has a lot of different skill values that factor in (list).

And there can be plenty of different cars in a race - Jacky Ickx probably has a higher skill level than Jean-Louis Schlesser in Group C racing, but ultimately the Porsche would just hold up the Mercedes. Weather can affect how the cars behave, so can track conditions.

There's probably improvements that can be made, but it also makes sense that letting a qualification session take place will improve your experience further.

 

I understand the complexity behind fine-tuning AI across multiple classes and the incremental nature of improvements. However, my feedback is based on consistent issues that persist with every major update, not just a single stream. While I appreciate the improvements being made, it’s hard to overlook the fact that key problems – like poor overtaking logic, odd slipstream behavior, and getting stuck behind lapped traffic – continue to reoccur.

The AI might perform well in isolated tests or specific conditions, but in real-world races, many of us are still seeing these same shortcomings. I’m not expecting perfection or an “extra-human” experience, but if the AI can’t handle basic overtaking or judge traffic properly, it severely impacts the overall gameplay.

I understand progress takes time, but after several updates, these issues feel less like incremental progress and more like recurring obstacles. I just hope that future patches address these fundamental problems more effectively.

 

ask Crimson

 

Yeah i was thinking about it after...the do more hardware than software!

 

A good way to stop being tempted to abuse possible AI weaknesses (notably them braking a bit too soon before turns) is to play with full damage enabled as any collision might be fatal. And it is much more immersive as you have to be really careful about your actions and risky moves, instead of being tempted to pass by force with little to no consequences.

 

Well said Renato! Well Said!

 

Can we just have Road Atlanta this weekend Renato

 

Because by showing restraint on the first lap leads to a better overall race experience. Ultimately it depends on what is more important to you.

But I do agree there shouldn't be a need to, I can't understand why even if the AI are a second a lap quicker than the player they still always start slow. Perhaps it's to do with so many of them being within close proximity to each other?