Official General AI Comments & Discussion Topic

Game version: 1.6.9.1.3145
Track and layout used: Pocono Southeast Course 1
Car: TSI Cup
Applicable settings: AI skill 105. AI aggressiveness high. AI wet driving skill scalar 1.00x. AI error rate 1x. Real Weather with 5x progression, 21/01/2023

Hello, it's me with the fringe reports again. The TSI Cup AI at Pocono Southeast Course 1 in cold weather and a damp/wet track is very slow. Attached is a screenshot from the end of a 16-lap race session, see lapped cars and the best lap times for the player vs the AI. The race was preceded by a 10-minute qualifying and they were similarly slow there.

imgur.com

 

I love to race the different F1 eras against the AI. It looks good, no question about it, but the AI just isn't what it could be.

There would be two tweaks needed:

1. Breaking: As It is right now the car behind is way better on the breaks. They close a gap of around 4 tenth in one breaking zone. And they are war too aggressive on top of that - driving into the diffusor of the car in front, sometimes even touching it.
That is way too dangerous considering how fragile F1 cars are.

2. Acceleration: it looks like the car behind gets the power down a tad too early. Normally the car in front visually pulls away out of corners. Not the case in Automobilsta 2. Might be possible that the drag effect sets in too early or in lower speeds.

In a perfect world we would see dirty air effects as well, but that might be hard to implement.

If the devs could iron out especially the unrealistic breaking behaviour we wouldn't see as many cars clustered together. And it wouldn't be as easy for slower cars to stay in the drag.

 

Agreed.

The beaking, to my mind, is part of the general "speed boost" granted to the AI behind the player. Whilst I understand why this boost is given it is overpowering. Toned down a little and applied at certain corners exit /entries (that are the accepted overtaking opportunities on the given circuit) and to all AI , not just the AI behind the player, would work vastly better.

And whilst on the subject of braking, the AI often brake gently at the start of a braking zone and harder later. Or that at least is my perception. Opposite of what is expected from high downforce cars such as F1. Their braking performance at the latter stage is better than that of the player (especially evident on the opening lap of a race) and often leads to me rear ending the AI unless I adapt my corner entry line.

 

I've also frequently bumped into the car in front, which makes me afraid to follow too closely.

Has anyone noticed that the AI in Sebring runs particularly slow? Even when set to 120%, it still can't match the lap times of regular players.
I hope we can use the records set in TAA mode as the benchmark... They've already managed to hit 1:56.

 

That is all true.

But there is a structural mistake there. The car behind is always way better on the breaks than the car in front. Between AI cars as well. If the car is 4 tenth behind at the start of a breaking zone it closes the gap to the car in front completely after the corners is finished.

That's the reason why we see the AI cars clustered together. That's why stronger cars or better drivers aren't able to create separation and break the drag.

 

That's probably also the reason the AI cars will create a brake-check sandwich, where the car in front brakes early while the one behind rear-ends you with enough velocity that it would've been impossible to make the turn in any real vehicle.

 

AI love to go for players rear tyres when coming from behind. After contact, instead of trying to get out of this situation, they stear into players car even harder, trying to go through other car.
On ovals, when they are slightly behind, they generally go for players rear completely blind, forcing it into a spin.

F_USA Gen 1-3:
in all oval tracks, AI brake on corner exits for no reason when in front of player, no matter which strength/aggression settings. It's been like that since years. A terrible experience.

 

Although I am very pleased with the raceability of the AI meaning that they race me close without wrecking me and although the mistake does it's job relatively well one thing that bothers me is how rigid their pace is.
Based on a race I did in Oulton ( and I understand AI behaviour varies from track to track ) the 26 total cars were breaking up into separate packs ( let's say pack 1 pace 1.50 , pack 2 pace 1.52 , pack 3 pace 1.54 etc. ) but within those packs there is no pace variation.
So for all intents and purposes we've got 3-4 separate trains absolutely on rails never missing an apex or a braking zone or whatever until the mistake factor kicks in and artificially makes them do something stupid.
If we could change that somehow I believe that AI racing would be much more exciting than it currently is. Because as it is , it still remains a stopgap measure for online racing. And as we all know AMS2 isn't doing terribly well on the online department.

 

It sound strange...i've not tested it the specific combo, but in Monza doesn't seem to happen, can you provide info on the AI difficulty? i'm around 90%. Could be that your gas pedal is not calibrated properly and you don't reach 100% (i said because happened to me in the past with the older pedals)

 

Like others said above me, that speed boost they gain to follow a faster car up close creates problems. It not only happens vs the player, it happens between them too.

I just threw a FV10G1 race at Buenos Aires 6S for the sake of F1 Racing Simulation nostalgia. I set real weather and got a couple strange sessions that setup a weird grid. Villeneuve managed to gain some spots, passed me for 4th and went after Irvine for 3rd (who I was chasing). A couple laps later he was past, but Eddie got glued to his back and the two drove into the sunset for the reminder of the race. Irvine had no business doing that, but he gained +1s per lap when Jacques passed him.

Just like they simulated water puddles for AI by forcing them to slow through those, something should be done with some grip loss logic due to proximity and speed for dirty air simulation. It would help countering this magic proximity boost with high performance cars, even though it might not solve the problem, which seems ingrained in the engine and Reiza has not found the loose end to pull it out. Yet.

 

I like what you said overall but keep in mind that artificial slowdowns create their own problems. For example if you do a wet race and you reduce the rain speed multiplier you may find yourself behind an AI that goes too slow as if he is dragging their brakes. This in turn prompts you to increase AI difficulty or rain pace or whatever but then you constantly fall into this tug of war trap with either too fast or too slow AI.

 

I don't believe what you describe is the result of a speed boost like the AI behind the player. It is, in my perception, down to the AI having an overpowered aerodynamic tow combined with no dirty air penalty.

 

Not saying it's a proper solution, and it's not something that will be nailed on the first go: several behaviours of AI that now work well have been subject to numerous logic code tweaking, and countless calibration passes, before feeling somewhat natural and good to observe and race against. It used to be impossible to race AI in the wet before the puddle logic was implemented because they would simply drive through puddles with no loss of speed besides the general grip of the track (which even impacted them way less than the human driver 3 years ago), and it's something that bit by bit was reigned in.

It's hard to know the true limitations of the Madness engine, and at what point you can continue to extract performance and solve ingrained bugs on it, and when you stop chasing that and just code around its limitations to find other ways to improve your final product.

 

Unless there is headwind (wind affects player only AFAIK) or there is a calibration problem with AI parameters, the tow effect on the player is the same as the AI. Otherwise, when following a train of cars they would just pull away from you and you wouldn't be able to draft them and pull outbraking moves at the end of straights. Which you can absolutely do, and sometimes it's the chance to gain 2 or even 3 spots in one swift manouver. On the reverse, if you are driving well out of slow corners into long straights, the AI will struggle to latch onto your slipstream.

What points at a "performance boost zone" phenomenon, is how arbitrary it responds to gaps, on very different types of cars. If you can break out of an guesstimate of 8-9 tenths gap in front, they will drop back bigtime in a blink of an eye; it's quite normal to pass a rival, have him on your mirror a couple of laps when he's 5 tenths back, then if you nailed a couple sections and you managed to eek it out to 1s, less than half a lap later the rival will be +2s back.

Also, this effect would not happen when racing on a very twisty track when there is no time for tow to have an effect. If it was just draft, it would not affect their improved braking distances, and it would not affect their improved apex speeds. But it does.

 

I don't think it's related to the tow, from my experience, the player's draft seems to be generally much stronger than the AI's (at least on most Formula classes, GT1s and F-USAs for example don't seem to suffer of this from my experience). It's common for me to easily pass an opponent along a straight and then, on the next lap, with my draft, the AI barely gains any speed on me, even when we're both driving the same car. Obviously it varies between classes, but it seems to be there on pretty much every gen from F-Retro all the way to F-Ultimate. Yet they are still faster after I pass them. Besides, as was already said, they seem to gain the most time on braking and cornering, not on straights.

 

A lot to unpack here!

I am not at all convinced there is any speed boost for AI to AI battles or overtaking.

Head and tail winds have no effect on the AI, agreed.

The (around) 1 second speed boost rule is very obvious and, to me, needs some extra varibles added to it. It would help make the racing more organic. Perhaps linked to laps behind the player so they would push and back off, before pushing again for the overtake. Race conditions, like near pitstops, or at the start or end of a race. Also scaled with the aggression attribute.

Generally I agree with all that you guys(?) say. However go to Le Mans 24hr with a slightly too fast AI for your skill, 1 or two clicks sufficient, and default setup for gear ratios and wings. Observe the train of AI pull away down the straights, an overpowered tow? I may be wrong, at both the modern Spa layouts the AI pull away down the Kemmel Straight. It is impossible to carry the same speed as they do over the top of Raddillon. Is this dirty air or that they are simply too fast in Radillon leading to better top speed?

Either way it's all moot, because the AI can seldom use an overspeed advantage on the car in front (save the player unless you weave and block) as they fail to pull out from the slipstream of the car in front, get too close and ease off. They're generally far to timid in overtaking with each other. They sometimes get in a good position but back out. If they showed the same aggression as they do against the player to each other, overtaking would be far better.

The most egregious behavour by the AI is in practice and quali. The AI on out and in laps drive slow on straights and hold the racing line, they are not so slow in the corners, just fail to accelerate. They do not concede to a car on a hot lap. They will even attack and attempt to overtake the player in the last corner, to then pull into the pits! A hot lap AI encountering them generally doesn't pull out to overtake on straights, just drives straight in line behind, then gives up on it's lap. This leads to trains of slow cars and they struggle to set representative laptimes, even in long sessions. It is a behaviour that needs attention asap because it ruins these sessions completely.

For the endurance classes I'd really like to see some AI (those with the high tyre management skill)
double stint the tyres, using harder compounds. Generally utilise the other compounds available rather than just the softs, not to mention where it is a must due to track temperature. Limited tyre sets was mentioned some years ago. Hopefully these differences would lead to more AI on AI overtaking.

Needless to say the AI really needs the development attention that the tyres and physics have been granted, because whilst somewhat functional it has a long way to go in order to be good, let alone great. As an offline focused sim, to really stand out amongst the other sims available, and for the career mode to be successful, this is a must.

 

100% agreed. This massive discrepancy between player tow and AI tow (especially AI on AI tow) is one of the major immersion breaking features when racing offline.

 

f

Fully agree on them blocking during practice quali runs. As for the dirty air question. Well it's simple to solve. Do the exact same thing you mentioned but with a low downforce car. GT4 perhaps. If they still demonstrate the same acceleration then it's a variable in the code that scales with the difficulty slider. If they behave normaly then yes it's dirty air. And yes dirty air will always **** the player while the AI probably ignores it as much as it used to ignore puddles.

 

Hello Reiza,
I encountered a number of AI behavioral and general issues in this video's race.
Line choices, Blue flag adherence, pit crew on race circuit, un-warranted penalty (while the pit crew and penalty was not strictly AI, I'm not really sure where it belongs, so leaving here).

Details in the summary at the end, and lap references for some (but not all) examples in the main description.

Also, worth mentioning that the VTC2 class AI seems about +10-15 AI strength compared to what it was pre 1.6.9. Might be worth adding the class to the strength calibration list, if it's not already. Oulton Park was closer to +20, so it seems quite track dependent also.

All the best
Andy