Official General AI Comments & Discussion Topic

Anyobe else finding a few AI issues at Le Mans, using GT3/LMDH/LMP3, namely;

• Turn 1; they tend to run a bit wide then gain magical traction and sort of "snap" back onto the racing line.
• Second Mulsanne chicane, they are able to brake much later than me.
• The first fast Right hand turn into 'Indianapolis' they seem to have far too much traction and get round that corner far quicker than me.

 

Is it just me or do the Formula V10 G1 cars have significantly worse starts than the McLaren MP4/12?

 

LMDH ERS modes don't affect power output, they change fuel efficiency.

 

Will racing AI ever actually consist of AI, so it can solve All of these issues? I'm very surprised that a console sim (gt7) is beginning to utilize true AI. We need it on the hardcore, pc sim side.

 

In Daytona, AI still are getting stuck and crashing constantly in T1 and the Bus stop.

 

Seriously? There have been so many reports with videos and screenshots, including myself...

 

I'm pretty sure the evidence is asked not to prove the issue's existence but to give a bit more detailed insights/evidence to try to debug why it may be happening. No need to be stroppy.

 

Screenshot of T1

Videos of the bus stop chicane.

 

- Car: Copa Classic B And GT classics
- Track: Velopark 2017
- AI strength: 106
- Track conditions: both wet and dry
- Report: the AI is very slow through corners 5, 6 and 7. I easily overtake both the Chevrolet and the Porsche of the GT classics and I gain about 2 seconds on the Copa Classic B cars.
Then they brake quite late into turn 8. I brake at the 100m sign, they (Copa) can brake about 20 meters later.

 

- Car: FClassic Gen 2
- Track: Watkins Glen Short
- AI strength: 110
- Track conditions: dry
- Report:The AI is very slow on sector 1, specially the esses. I am almost 8 tenths faster with the model 2 than the next car (McLarens), while I am barely on the pace through the lap (so they could be actually losing more than a second wrt a reasonable pace there)

 

- Car: Gen1 GT3 but others as well
- Track: Silverstone National
- AI strength: 100
- Track conditions: dry
- Report: AI struggle at T2 & T3, they seem to take a very wide entry into T2 and lock up and spin. This happens nearly every lap, I did gain places as a result but still not ideal
From memory this has been like this for a long time?

 

- Car: Gen2 GT3s AND LMP2s (didn't try LMDh's)
- Track: Le Mans
- AI strength: 110
- AI aggression: Medium
- Track conditions: dry
- Report:
I have been doing a couple of races a day at Le Mans this week but have been struggling with the AI whenever it gets close to me - to the point they now sometimes scare me off the road completely!

It seems at this circuit (more than others) they

(i) have 'magical' acceleration out of corners even if you brake check them and get an initially better exit; and

(ii) thus AI in your own class seem to have a speed advantage down the straights and going into corners which they try to exploit with moves that really aren't on.

I've ended up off the road a dozen times this week and kept blaming myself although I thought the AI might have been the cause. I finally took the time to look at a replay and I have documented this one example of "moves that aren't on" in this unlisted video:

I'm in the Audi and the Porsche behind me goes for a move which involves cutting the chicane, making contact and keeping his foot in it - which turns me around and puts me in the wall.

Maybe I left the door open?? I don't think I could have avoided that. What do you think?

EDIT: I've run a couple of GT3 races on low AI aggression and it seemed to help alot. They still passed me if I made and error but seemed to back out of dangerous lunges. Normally I prefer medium aggression but for whatever reason at Le Mans, low feels more realistic. Feeling a bit 'safer' I drove much better too. Win win.

 

I can second these findings. The second chicane on the mulsanne is very difficult to navigate safely when they brake like 50m of more after you need to. It also doesn't help that when starting a race the AI are basically at 100% race pace immediately where it takes 1-2 laps for our tires to come up to temp so you get absolutely mugged by the AI at the start of races while you're waiting for the tires to come up to temp and gain traction.

 

Regarding Daytona, there are two track layouts - one for sports cars and another for NASCAR. The only difference in layout is an additional bus stop chicane.

In the sports car configuration, drivers hit the tire wall at Turn 1 every single lap, whereas the NASCAR layout runs smoothly throughout but can only start with rolling starts. As for the bus stop chicane, even when I drive formula cars there, I end up spinning out. The curb design makes it clear this track isn't suited for formula racing - you can tell just from its name.

 

Most tracks in the game (perhaps Interlagos is the only exception) have at least one corner/section where the AI exceed player physical limitations in this way. Braking and/or acceleration/corner speed are simply impossible to match, then equally there is a corner, section or straight where they are slow and the overall laptime balances out. I understand the incredible amount of content in the game makes this a daunting and difficult problem to solve, and that the work is only increased with every physics revision. With the scale of the problem I find it astounding that, as stated in the developer physics thread, Reiza will not leverage the power of current AI LLMs to attempt to solve these problems. Given the large dataset that must be available from time trial ghosts, this is a problem where the solution is potentially in reach, yet the attitude of the developer's is closed minded. More content recently teased is fantastic but the issues with the AI drivers will only ever be multiplied, and leaves me concerned that there is very little chance they will ever be solved...

 

Whilst I have long been curious whether an appropriately designed machine learning approach could leverage the time trial data to get better AI lines (if the lads are interested I'd certainly be up for sending in my CV!), your implicit claim that Reiza should use the 'power of current AI LLMs' shows that you might have a somewhat incorrect idea of how they work or how someone might go about solving one of those problems.

I am 100% sure, however, that the reason it hasn't been used is not 'the attitude of the developer's is closed minded'. I think there has been a misunderstanding of what it would mean to use 'AI (like a generative model we think of as AI now)' to make AI (as the rules for the non-player characters/entities in a game. Making Sophy (an extremely expensive and highly complex endeavour) is a completely different kettle of fish to using time trial data to inform AI lines/behaviours (which itself requires certain expertise at a cost that may not justify the effort). It is possible the expertise is already available in the company, but that there are other more pressing and valuable ways to expend such efforts. I think we can all agree at least that the AI in AMS2 is some of the best in the business - and I very regularly have extremely enjoyable races against many cars and on many tracks that I just could not set up a multiplayer lobby for and get anywhere near as enjoyable a result.

 

Regardless of the class and strength, they always avoid it and brake very early on the straight instead of straigthtlining it. The curb is very forgiving due to the high speed, so I don't think that adjusting the ai line to go over them would cause them to crash

 

True, I have no idea how they work, but does anyone? Machine learning is indeed a far better term when applied to my perspective.

I'm not expecting a Sophie style AI. As you rightly point out the existing computer drivers using 'old school logic' are pretty good. However they, I believe, could potentially be optimised by leveraging the power of machine learning to "fill the gaps" in it's flaws at certain sections of the many tracks, in the many cars that AMS2 offers. The braking, acceleration, steering, speed, track conditions ect are datasets and, the ghost replays from humans are also datasets. This is what AI is good at, analysis of large datasets and pattern finding. This is my point, how this is done I can't answer, maybe you can! Current AI is already coding fast and efficiently, all I (and others) are saying is a more open minded attitude to the possibilities of leveraging machine learning to optimise the computer drivers code and pathing could result in a better experience.

Sorry for being a grumpy and somewhat intransigent old sod, all I'd like (and others no doubt) is a more balanced experience against the computer drivers in single player...