So I prefer single player racing and I was wondering if AI can still be trained

No. Forever ago the training AI (which were slow) was replaced with a new AI system. We can see with the mods that it is able to quickly iterate an optimal race path in a short period (30-40 seconds)

Quote from subUwU :

So I prefer single player racing and I was wondering if AI can still be trained

I don't know how the latest version of the AI is coded. But it looks like the AI times are improving a bit race after race, to a certain extent.
If your problem is that the AI is too slow in your single player races, you can easily make it faster with the mod system.

yea but I also want to improve their spatial awareness and maybe get them to do some overtaking

For spatial awareness you won't get anything that way. The flaws of the AI remain the same, even are amplified by the increase in speed.

But with faster engines, the AI behavior changes a little. Bots are much more aggressive and attempt (and sometimes succeed ) more overtaking. Races against the AI are less boring. You are no longer alone on the track in the first laps.

It's not entirely satisfactory, I agree. It's just a way to continue playing solo without falling asleep.

Quote from subUwU :

yea but I also want to improve their spatial awareness and maybe get them to do some overtaking

get coding

If you are passionate about AI programming, I recommend reading this topic.

https://www.lfs.net/forum/thread/63807-Artificial-Intelligence-in-Racing-Simulations-Project

i do think so

You can mix some different levels of AIs on the grid. The more experienced ones will overtake the others.

actually i've tried that, and that seems to cause more issues, because the faster AI will sit behind the slower AI, causing a massive train, and in the braking zones the slower AI brakes before. Because they lack spatial awareness, the faster AI will run right into them.

Quote from subUwU :

actually i've tried that, and that seems to cause more issues, because the faster AI will sit behind the slower AI, causing a massive train, and in the braking zones the slower AI brakes before. Because they lack spatial awareness, the faster AI will run right into them.

Yeah, the AI need to be far more aggressive when there's pace differences.

I believe that new AIs in LFS are implemented using some machine learning algorithms, but don't quote me on that. There were attempts to decode the AI path files, but no one succeeded so far and devs are quite silent about their inner workings. We don't really know how they work at all.

What I saw is that once you start racing against them, they do get slightly better, but only by a small margin. This is still nothing close to racing against humans, but it does provide some sort of single-player fun. To add "spacial awareness" to AI is probably the task that needs scientific research, I mean this is an extremely complex topic and one can do many implementations from simple to more advanced. Again, this is something that LFS devs don't have time to work on, so I don't think it will be any better than what we have now.

Quote from rane_nbg :

I believe that new AIs in LFS are implemented using some machine learning algorithms, but don't quote me on that. There were attempts to decode the AI path files, but no one succeeded so far and devs are quite silent about their inner workings. We don't really know how they work at all.

What I saw is that once you start racing against them, they do get slightly better, but only by a small margin. This is still nothing close to racing against humans, but it does provide some sort of single-player fun. To add "spacial awareness" to AI is probably the task that needs scientific research, I mean this is an extremely complex topic and one can do many implementations from simple to more advanced. Again, this is something that LFS devs don't have time to work on, so I don't think it will be any better than what we have now.

The basic AI paths are "machine learned" in that they quickly iterate a bunch of attempts to find the "optimal" path but it's pretty brute force.

In terms of racing against them, they don't learn. The learning AI died like 10+ years ago when Scawen made the AI half-decent out of the box (using the presumed current "machine learning" approach).

Any perception of learning while racing is down to placebo effect or just AI gaining tyre temps and thus more speed.

Quote from gu3st :

Any perception of learning while racing is down to placebo effect or just AI gaining tyre temps and thus more speed.

You are probably right. In this case, the placebo effect is very powerful
It seems to me that there is something else. Without doubting what you say, I would like to have confirmation from Scawen on this subject

If it can help anyone, here's what I do to get decent races against the AI

The solution that works the best is to create (in a few clicks, 30 seconds) specific car versions for the AI with the original cars or some mods. I equip these AI cars with more powerful engines and slick tires. I drive a less powerful car with super road tyres. It works fine (for me, I'm not a competitive rider, I ride for fun at 2-4 seconds per lap faster than normal AI).

If I want to put AIs of different levels, I put level 5 in front of level 4 etc. to have fewer problems. After a few races, the grid gets organized and clashes are less frequent (especially at the start). It's better, but not magic.