Automobilista 2 March 2024 Development Update

I agree in part but the AI need to be on the same physics as the player as well. But also like Horia pointed out there is a danger that the AI become too good.

 

I highly doubt that ML would actually produce the "randomness" and "errors" you want. How would they even "machine learn" to make realistic mistakes without effecting the normal racing experience? How would we be able to keep the same level of customizability that the AI skill system currently presents to us?

I do not want to sacrifice the level of customizability AMS 2's AI has at the moment. It's (at worst) second to Iracing in terms of customizability, which has the best AI on the market according to most simracers. It is not a coincidence that two of the best AI in simracing right now has extensive levels of customizability.

 

Yup, Machine Learning is not necessary to create good "human" like AI, GP4 did it 22 years ago and as Tom wrote a couple of posts before the F1 games after 2016 have become the benchmark of great AI, they do small lockups, overshoot corners, can spin on corner entry or exit, crash into each other when trying to overtake, have hot and cold days, look for the player slipstream, block the inside and can do the switcheroo, they are very, very good. Amazingly, GP4 did most of that as well with much lesser technology and they actually did more mistakes when it was very wet, something the F1 games dont do. Obviously the physics were simpler in GP4 so it would be theoretically "easier" to program the AI to be more dynamic with less variables at play, but its still very impressive.

I am sure the ME and AMS2 are capable of simulating all those things and they already do simulate a lot of them. It took Codemasters 6+ years of work to improve their AI and basically 10+ to get where they are now with 2023 which is the benchmark IMHO. So hopefully Reiza can put their AI among those by the time the game is properly finished, but the caveat is that we are talking about 20+ times the amount of cars to calibrate the AI, so godspeed to them.

 

not forgetting that many used to consider AMS 1 AI experience to be outstandingly good or best of all and we all hurrah'd at the idea of that quality coming to AMS 2, iinm.

 

In AMS2 I have already seen that the AI defends its line and was impressed because I only knew this behavior from RF2 so far. It really helps the immersion. Perhaps this requires the parameter in the AI file to have a high value, which is why I check every custom AI file for this value and increase it if necessary. Or maybe Reiza has put this behavior on hold for the time being for some inexplicable reason? I haven't raced against the AI for a while so I don't know for sure. But if you compare the AI in PCars 2 with that of AMS2 we are talking about huge differences in quality in favor of AMS2 (although there is still a lot of room for improvement).

 

I would like to see a video of that "defending" happen. Here is a short video of there being no AI defending in the current version of the game. After 10 minutes of this i got bored and quit again because racing AROUND the AI instead of against them is no fun. Increasing the AI difficulty doesn't improve things either because a higher difficulty does not add a missing AI feature. Also that 0:44 moment is not an "AI defending" moment either, they just have the line that is advantageous and it would have been embarrassing if they could not pass me there.
P.S. sorry for no in-game audio, screwed up the recording so i just put some music in instead.

 

Tom, I thought Renalto had replied to this concern of yours.

Cool music on the video though. Made me smile

 

it s today??

 

Definitely not

 

This unfortunately isn’t going to happen in any racing sim, barring some miracle breakthrough with CPUs in general.

 

This is a post that I've wanted to make for around a week now. My goal is to show Reiza's track record to hopefully show why I'm not only incredibly patient with Reiza, but also to show why some of the more general complaints are incredibly frustrating to read for me. This is NOT to excuse problems with the game (which there are plenty of) but to instead prove how far the sim has come, along with showing how the community elevated it. We must understand where we have been to understand where we are and where we will be hopefully going. Due to my love of taking screenshots, I have found that a story can be told through them. As I typed much of what's below me early this morning before forcing myself to actually sleep, even me, one was AMS 2's biggest cheerleaders was taken aback about just how much the sim has improved thanks to the hard work of both Reiza and the community. I hope you enjoy going on this journey with me as much as I did writing it.



This screenshot was taken on ‎June ‎6th ‎2022 as of the time I'm writing this (March 27th 2024) that was 660 days ago, the very same day that the BMW M-Hybrid V8 was unveiled. In the span of just under 2 years AMS 2 went from Formula 1: Tokyo Drift to something that was actually realistic. The transformation in that one aspect alone is quite honestly incredible, and that wasn't the only aspect that had this level of transformation. It was not long ago when cars like the Caterham's and McLaren F1 were not just "the weakest of the bunch" but flat out impossible to drive as the next picture taken on ‎November ‎5th ‎2022, 508 days ago just under a month after Max Verstappen secured his 2nd championship shows.



That was not the only area AMS 2 has seen massive improvements. There has been massive leaps when it comes to modability and customization, and both Reiza and the community share credit with this one. The start of 2023 was a very pessimistic time in sim-racing, however through all of this darkness an important fire was lit. 371 days ago, on March 22nd 2023 the KAOS Modding Team released their legendary 2023 Indycar mod just two days after Fernando Alonso scored his second podium in a row at Jeddah. On April 15th, 347 days ago and the very same day the IMSA Grand Prix at Long Beach ended in tears for Ricky Taylor after making a costly mistake that gifted the Porsche 963 its first ever win, paoloambrosio released the AMS 2 Content Manager on RD which made adding mods to AMS 2 as simple as a drag and drop.



This picture was taken 303 days ago on May 29th 2023, the day after Josef Newgarden won the Indy 500 for the first time in his decade plus career. This picture shows how both Reiza and modders made the best out of a bad situation. Reiza using the renewed line of communication with Indycar to get the iconic Indianapolis Motor Speedway, and the KAOS modding team filling in the F-USA 22 shaped hole in our hearts. And while the oval racing in AMS 2 needed (and to a point still needs) plenty of work it is undeniable that the state oval racing has been constantly improving since it was first introduced in Racin' USA part 3.

It was not only mods though, things like race helmets and driver suits became customizable. This let the amazing painters in this community add one of the most important details of open wheel racing. A culmination of this happened on November 5th 2023, 143 days ago and the same day Ryan Blaney won the NASCAR Cup Championship for the first time when a group of very talented and dedicated people called the Immersion modding group released their first mod as that name, the F1 1990 Season set. While I do not have a picture from that time, I do have one of the equally fantastic 1992 set that came out last month.



This is just scratching the surface of what both Reiza and the community have done to this sim. The community has given us skins to represent a wide array of disciplines from Formula 1 to NASCAR, along with some surprisingly great conversions from other M.E based games. Reiza has given us an incredible array of unique tracks and cars, many of which are not present in areas other than modding in other sims. tracks like the 1970 version of Spa with the chicane, the 1971 version of the Nürburgring Nordschleife and Sudschleife, 1988 versions of Estoril and Jerez, and 1991 versions of Interlagos and Barcelona. Cars like the Formula Junior, Formula Inter, Brabham BT52, and all of the Formula Hi-Tech models. Annoying bugs like open wheel cars flying in the air whenever wheels touched or detached pieces of a car not visually showing up on said car after restarting a race have been ironed out. The AI are much better at navigating slower traffic and actually staying on the track, of course it can be improved but it's a far cry from what it used to be. We also now have Safety Cars with various cars representing multiple eras along with trackside LED lights. While I cant be sure for certain, I'd bet that even AMS 2's Achilles Heel (multiplayer) has made large improvements between 2 years ago and now. ALL of this and more has happened within the past 2 years which leads me to my final picture, taken today March 27th 2024, 3 days after Max Verstappen had his first DNF since the 2022 Australian Grand Prix.



This image was supposed to just be the BMW at LeMans, but my lack of skill decided to make it much more thematic to the first screenshot in this post. 660 days ago, the very same day BMW unveiled this very car this screenshot would've been impossible. Not just because of the content shown in it, but also because I had spun the car out. If this was in 1.3 there is a good chance I wouldn't have overcorrected after hitting the kerb at the end of the Ford Chicane, but thanks to multiple updates that slowly refined AMS 2's handling model over the course of 1 year, 9 months and 21 days I got exactly what I deserved for being too liberal with the kerbs instead of getting away with it. If I said that we would not only be getting the BMW M-Hybrid V8 and LeMans but also substantial improvements on just about all fronts within the next 2 years, 660 days ago, I would've been "optimistic bombed" into oblivion (and maybe even Skyrim ). Yet all of that happened, all of AMS 2's growth over the past 2 years that I've shown in this post happened. For a team that's known for pushing updates back, all of these improvements and additions have happened within a decently short time. Especially has days go by quicker and quicker as I grow older and older.

AMS 2 becoming a substantially better sim is about as objective as an opinion can get. This doesn't mean the sim is without flaws, there are plenty of them. This is to prove that Reiza has an amazing track record of ironing out said flaws to not only make the sim better but also open up more opportunities for talented members of the community to create more realistic skins, ai sets, and even mods. With all of the evidence and with all of the progress, there is no doubt in my mind that Reiza are still continuing to work to keep ironing out issues that currently persist in the sim.

With all of the progress that has been made in the past 660 days, I can only imagine how much better the game will be after the next 660, and I will gladly be here waiting every last second because if Reiza's track record proves anything, it's gonna be worth the wait.

 

Goddamn I wish there was a Super Winner rating to give this post.

 

thank you very much. I had spent about 3 hours "last night" (I went to bed at 4 AM like any good 20-something ) and then not long after getting up spent another 2-3 finishing it up. I spent a lot of time not only doing research but also making my let's say... "artistic interpretation" of English grammer actually be decent to read the kind words absolutely mean the world to me

 

I wanted to put a HELL YEAH gif , but i have failed

 

This post is awesome!

My introduction to AMS2 was different, but ended in a similar place — with a lot of trust in Reiza. I've found AMS2 to be remarkably approachable for being new to sim racing (and using a pad). I’ve put in more than 80 hours since mid-January and am thrilled to bits that there's still a roadmap for new content, updates, etc. It feels like AMS2 still has strong legs under it and — for a four-year-old title — is terrific!

 

Ams2 is one of the newest simulations on the market. 4 years is nothing compared to other simulations. This is just the beginning of what will hopefully be a long journey

 

It much depends on the implementation, the technology is not the problem. We do not need to go too far - PC2 vs. AMS2. If PC2 did not succeed as a sim (a good game, but not a good sim), why should we expect AMS2 to be better? Well, we expect a better implementation.
As of AI strength, well... chess games are still not impossible to beat. They are implement with enough randomness (yes, pure noise on a neural network is enough!!) to ensure playability at different player's competence levels (including been unbeatable!). That's the easy part.
Speaking of chess, the best program (Stockfish, considered unbeatable at the time) was imediatly crushed by very, very young AI agents, a few years ago, using deep-learning and they didn't even needed a human to train them! They trained themselves (reinforcement learning) only requiring the book of rules as input. A few hours latter, the best ever chess player was born...

I believe you're not too much familiarized with ML algorithms spectrum - today they are the best approach to human behaviour (including errors and cognitive biases).

Reiza is doing a good job at AI, that's for sure (it was not usable at all not long ago).
ME action is perpendicular to this problem. "AI" programming is not ME related (I may be wrong).
Nevertheless, I believe ML would be the better way to train those 20+ times the amount of cars - just leave them running overnight with the new physics and the next day they're ready to ship!! That's the beauty of reinforcement learning.

Training models are computationally expensive, inference models (the ones taking decisions, the ones that need to be implemented in the game) are not - and they run the best on GPU.
It's a matter of time until every computer has enough tensor processing capability that running 100 AI on Nordschleife or Le Mans is be a breeze

 

But that's the thing, sim racing doesn't have a "book of rules" you can feed to the AI. You're massively overestimating the capabilities of AI/ML. The more open-ended and dynamic the environment that you put the AI into, the more exponentially complex the training becomes, and exponentially higher risk of unintended behaviour.

AI works best on systems with discrete data points and repeatable patterns. Like chess, just as you said. But the reason we can train AI in chess is because on each step of the game, there is a finite and comprehensible amount of variables the AI must account for, and the states it finds itself in are much more likely to repeat over many iterations of a game. TL;DR there is a pattern to chess, and we can exploit that in teaching the ai.

On the other hand though, the amount of variables processed in each fraction of a millisecond in such an open-ended and physics based environment like Automobilista is beyond incomprehensible. You wouldn't know where to start training - what to reinforce, what to punish, what's the goal, what's the generation selection criteria.

As I said before, AI is best at pushing the limit of any system it is implemented in. If the limits are concrete and static (like in chess - the AI cannot make illegal moves or abuse any of the rules of chess), everything works fine. The more dynamic the system however, the more the AI will bend the rules and break the simulation in order to achieve the intended result.

Here are two very good videos with examples AI bending the rules because the environment is too open-ended (including Polyphony's Sophy):
OpenAI Plays Hide and Seek…and Breaks The Game! (youtube.com)
Sony's Sophy racing AI beats Gran Turismo's top players (youtube.com) (especially 6:42)

 

So in your world, feedback is "complaints". Do you think the title improved due to "patience" or "feedback"? (something you want to label as complaints). I am equally frustrated by peoples constant labelling of feedback as complaints.

 

I remember seen those videos . 2 and 4 years ago? That was ages ago. The past is now measured in months! The future is measured in weeks.
Most people are oblivious to what AI will do (very, very, soon!) ... "machines will never be as good as humans...", "at least, not in my field of expertise...". My educated guess: in 3-5 years, all sims will have it.
Note: online racing could very well became a place to avoid as many online racers will be trolls, bots racing as if they were real people (reasons? I can imagine a few, but my imagination can't grasp how far it can go)