no, no, no, can’t do that. You changed fuel and tyres, doesn’t count.
(in case not obvious) /s
2 Likes
I don’t post too often on here but -
for the love of god do not touch the AI anymore. with this last update they can finally sorta run alongside you and do switchbacks without pitting you in your rear quarter panel
The copium is strong here. lol, nothing has changed with the AI in about 4 months. They’re still the same racing line dependent AI that you can be 100 PI lower then but as long as you don’t leave the racing line they can’t pass. They’re also scripted to make the same mistakes lap after lap (those are the ones who don’t have Max Verstappen superhuman ability to improve in correspondence to your lap time as you try to chase them down). I’ve seen AI lower their best lap over a second and a half just to make sure the gap back to me is maintained over the course of a race.
But then again, these are the same folks who haven’t changed anything about the single player events in almost 6 months other than what tracks we run at (and even then it’s the same stuff sometimes, see this month where we run the Indy layout of Brands multiple times)
Stuffs bad
3 Likes
No. I’m the one being invalidated and/or discredited. Instantly. Because my experience I guess seems so outlandish that it 200% couldn’t be possible. Based on what? Your own experience?
Well, I’m not a liar and I don’t make things up. I merely provide what my experience is. I’m very willing to take knowledge of others when they don’t first completely try to invalidate my own. I set the terms of the challenge for a reason. Specifically said less than 5 laps for reasons I already explained. The Ai behaves differently based on lap count. The longer the race, the less aggressive and superhuman they seem.
I have been met with nothing but resistance for simply explaining what I experience in this game. You’re not here with me. Neither are they. Who’re y’all to completely discredit me? I’m not saying that any of my experience is what others should be experiencing. I’m not saying that it’s my way or the highway. I’m just laying down in text my experience with the Ai in this game. You can all say that you don’t experience it the way I do. I don’t care. What I do care about is that y’all seem to attempt to make me out to be a crazy person.
Did you see my post?
Maybe someday they’ll make the AI as unpredictable and imperfect as real players.
I want AI that accidentally brake too early, too late, dive on the inside, fight with each other.
Im not sure if youre being sarcastic or not, id say this game has most of those qualities, except theyre not really programmed to do it. When they exhibit these behaviors it always seems as though theyre confused as to whats going on and what to do.
I think in a lot of ways they tried, when they first implemented drivatars to emulate real players. But imo the unpredictability is this games problem. Its just like driving a car in real life, you dont really want the drivers around you to do unpredictable things, that often leads to accidents.
In real racing there has to be a certain level of trust that you have with the other drivers. Obviously things happen and its fine in a game from time to time to add a little unpredictability, but forzas ai’s unpredictability is predictable. You can count on them to do unpredictable things at the most absurd times.
Thats why i dont think its programmed, its like theyre confused or lose track of what theyre supposed to do. Like when some ai will continually go off the track at the same point lap after lap. Maybe it was supposed to make that mistake once, but gets caught in a loop of stupidity.
I think the best we can hope for is a game that has ai with a consistent level of predictability. Like gt7s ai do a good job of this. Their aggression is unpredictable, but their actions are. Some will aggressively defend going into a corner and some dont. Some will make a move up the inside if given the chance to but in a controlled fashion. They will move over for faster traffic because they follow blue flag rules. Youre able to spot their intentions before they do something.
Forzas ai do these things in nonsensical ways. They dont defend going into a corner so much as activley block you through it changing their corner lines. If any try to overtake on the inside its a dive bomb that would never work.
I wouldnt say forza ever really had the best ai. But id say in past games they were more consistent with their habits. I dont know what happened with this game, but the ai is just really bad. They truly lack the capability to do what theyre supposed to do.
2 Likes
I’ve said my piece. I apologize for offending you and making you feel your opinion is unwelcome. Let’s not drag this thread any further OT.
You know, after wasting a ton of time recording that spa clip (that crossquared won’t acknowledge), I came away super disappointed in the AI for a different reason, but also gained some more clarity that was reaffirming more than anything
They have terminal issues that confuse them. At the end of the ER-Rad straight, if I was alongside another car (was typically going into that braking zone against P3-P5 on previous attempts), they would always, ALWAYS overshoot the braking zone and push me wide when I was on the left/outside. Like, this would be the equivalent to playing someone online who would rather bully you into the run-off by braking deep off the track themselves instead of letting you share the outside
If the AI didn’t suck, you would expect to enter that chicane going two wide preparing to hug the upcoming inside/left-handed curbing
It happened so consistently that I started experimenting. Figured I’d try an undercut. So we’d be barreling into the T5/Les Combes chicane’s braking zone going two wide (me on the left), and I’d go heavy on the brakes at the last minute to do an undercut… and it would always initially work, with the AI shooting off into the shadow realm by themselves, BUT… jezuz christ… but this dumb-as-rocks, red-mist-seeing AI would ALWAYS come hurtling back into the chicane at a near parallel angle that it would always sabotage me. I’d get T-boned every time. Made undercutting—a very common, very normal, and very FUN part of racing—a total impossibility
So what’s the point of an AI that makes mistakes if it’s doing them to sabotage the player? Rhetorically speaking, not towards you. It prevents very basic, but good racing. You have to have hyper perception to have a chance at a clean battle with them
In most of my other sims, say AC and AMS2, you have separate parameters to adjust: skill level, aggression, “natural mistakes,” etc.
And why I’m so disappointed, is that the front runner runs a GOOD pace! Like, not alien levels, but P1 (and even P2 when they’re all by themselves) will use the majority of the track, brake, and exit appropriately. When they have NOTHING in front of them for at least 100 meters, they’re great
Battling them sucks. They suck at awareness. Object detection is haphazard and triggers their idiocy: their brake checking, their inability to stay on track, their mistakes (that don’t feel natural or appropriate at all); everything that’s wrong with their AI has to do with their abysmal lack of on-track detection skills
And it’s why I think so many people have these false theories about P1 having some kind of cheat codes that makes them so much faster
“I can beat level 8 AI, I’m faster than basically the entire grid, but P1 and P2 are bugged and go too fast. There’s something broken with them.”
P1-P2 don’t have any kind of special sauce; they’re just running a high-level pace that most users probably aren’t ready for. Their pace just isn’t dumbed down by a machine gun of panic attacks like a squirrel that doesn’t know where to go when a car comes down the road
The truth is that the entire GRID should be as fast as them, and level 8’s pace is likely way faster than the majority of players that don’t know how to use all of a track or manage the circle of traction with more advanced braking and throttle techniques, so it creates this false sense of scale of the AI’s leveling pace
If the AI was non-problematic, it should be next to impossible to gain half a grid of positions in two laps at Spa (or ~5 minutes on any track for that matter). You’d be battling for a handful of seconds at most, not clawing a 10-15 second delta going last-to-first because everyone behind the top split is having an aneurysm
So the disappointment is that the AI CAN be good, but it has one massive flaw that affects it in every situation outside of a one-car hotlap vacuum
3 Likes
I would love for them to hold an AMA with the AI team. When I look at a video like this, the complexity behind what’s going on is pretty mind-boggling. And this is just to train one AI, by itself, to go fast. Start adding in variables like multiple cars and avoidance, things like tire wear, mechanical damage, when best to make a pitstop…it’s mind-blowing. (I have a very little bit of programming knowledge, but nothing anywhere near what would be required to get something like this working.)
I say this not to try to defend Forza’s AI…clearly, it’s possible to do better - one needs only to look at other games currently available on the market. It can even be argued that T10 themselves have done better in the past - perhaps their older AI was not as technically complex, but I think it’s fair to say that it was at least more enjoyable to race against than what we have now.
I’d be really interested to know things like, what kind of system was used to train the AI? Was it some sort of machine learning network like shown in the video here? Or some other method? How is it technically different than the approach taken in past Forza titles? What variables exist for the devs to alter/improve the AI’s behavior?
2 Likes
I think the only improvement made from past games is that the ai are faster. Their issues stem from a lack of spacial awareness. Imo if take into account only one ai at a time can be fast, they lack spatial awareness and some ai get stuck in a loop of stupidity(drive off track, decide not to use full throttle, one ai is hyper aggressive accelerating in to corners), i think there has to be a lack of processing bandwidth.
They are obviously still related to past forza ai, so theyre not new. Something has caused these behaviors to happen and the fact that after a year of updates there has been little improvement. This means whatever the real issue is, it cant actually be fixed. Theres no overhead left to make them function properly.
2 Likes
I fear that you’re right 
1 Like
I wonder if that’s truly the case. Or is it a matter of priorities due to a lack of manpower (due in turn to a lack of funding, due at least partially to the game’s underwhelming sales performance).
I mean, in theory at least, nothing is unfixable given enough time and money. I think the problem is that this game doesn’t have the luxury of either.
1 Like
Forzas ai has always used machine learning. Heres a snippet of an article released before this game launched.
(Greenawalt explains that the Forza series has long used machine learning to power its AI-controlled characters that player races against. In the pre-Xbox One days, this process was owed to a Bayesian machine learning system that operated on the console’s local hard drive; but after this was moved onto the cloud. There, Greenawalt and his team used the network to train the Drivatar.
“Now, with the latest Forza, we’ve taken machine learning and applied it to build time and not run time or load time; so not while it’s running or loading, but actually we’re able to do it before the game launches. Instead of having machine learning power the moment-to-moment decisions of the Drivatar, we’re having it train the Drivatar to control the car, and then we’re using an optimizer to make the lines that Drivatar follows.”
“What that’s allowed us to do is train massive amounts of data so that we can take every car, with every upgrade, and all the tuning options through the wet, through the dry and train that controller so that the AI can make the car do everything it wants it to do. Ultimately, it was about applying the machine learning we were familiar with to a different place.”
With that, Greenawalt says applying the machine learning to a different part of Drivatar has led he and his team to the fastest AI with no cheats or hacks, and to a position where they’re maximizing their use of the technology in a non-arbitrary way.)
So personally going by these statements, the focus was to make the ai faster, which one per race is and the other was to not use rubberbanding which is also true.
What it doesnt mention, which is the problem the player faces, is the ai cant actually race anyone. Why this is, is unknown, but its obvious they cant make the right decisions on the fly either fast enough or at all. A simple test of this is when you pull in front of them and they just jam on their brakes. Theres no reason to do this, they dont differentiate you driving at speed in front of them or braking in front of them, they just brake.
For awhile the ai when driving along side of them would just drive off the track, why would they do this other than them not being able to change their racing line to continue going foward.
Although improved, they still have a hard time passing both the player and the other ai. They can only do so in the easiest of situations like on a straight. If there are multiple ai going through a corner, some sort of chaos usually ensues because of the lack of spatial awareness and the speed at which they can make a decision as to what to what line to be on.
Its kind of funny and sad in way because you have all this fancy sounding tech that has ultimately delivered some of the worst ai ever in a racing game and they cant fix it. Its like they put in more work to make it bad than it would have taken to make it good.
4 Likes
My assumption is based off how ai is actually powered if you will, which is through the cpu. But its not the only thing the cpu is powering during a race. When they added the car radar, the guy who made it said they dont have much room to add things to the ui due to bandwidth constraints because they cant deviate from their performance targets.
If it was a struggle to add car radar, how are they going to make the ai better. The answer is they obviously cant, because the game doesnt utilize multiple cpu cores properly. Its just another example of how this game wasnt built from the ground up and certainly is not prepared for the future as it doesnt use modern techniques.
2 Likes
I mean, to be fair, you did say this is your assumption. And it does seem plausible given the circumstances. I wonder, though, if it’s a case of the AI not having the compute power/efficiency to race in a believable, human-like manner, or if it’s something that could be improved by changing the weight given to certain input variables.
So according to this, if the machine learning system isn’t even running while you’re running the game…what actually is going on?
I’d be interested to know more about this “optimizer” and how it operates. Like you said, it sounds like it’s given priority to finding optimal driving lines, and not how to maintain optimal speed and lines in the presence of other cars or unexpected obstacles.
I seem to remember reading somewhere that with FM23, the AI no longer uses player behavior as training data. (Maybe they stopped doing it with an earlier title, I’m not sure). I think a lot of people were glad to hear this (myself included), because the first thing that came to mind was, “I’ve played public multiplayer and seen how other humans drive…I don’t want drivatars that act like it’s Wreckfest and punt me to the shadow realm every chance they get”. But surely there must be ways to adjust for this bad behavior and weigh things more toward the highly skilled racers so that their tendencies are the ones more closely mimicked. To not use player data at all (if that is indeed what’s going on) seems like a case of throwing out the baby with the bathwater.
1 Like
Realistically all we’ll ever have are assumptions, theyre never going to say any of this. But going by things ive read in other forums as well as here throughout the years,when the ai had issues these were the types of things that were talked about.
They basically said they had their existing ai data, which would have been cloud drivatar data from past games and then focused on gaining time, getting faster.
So they basically “train” the ai on their side and this optimizer is what runs in the game. This optimizer is the likely culprit as it takes too long to choose new lines. Theyve shown a video its posted somewhere here of how the ai senses its surroundings so in theory they do account for this in some way, but like i said the speed at which this happens is delayed.
When pcars 2 released, they made a big deal of how many cars at one time the ai could detect. This is where processing comes into play, as the more cars in that circle of detection the more processing is required. In this game a one on one confrontation with the ai is often fine, its when theres multiple cars that the bigger issues reveal themselves.
This is where things like your driving up beside the ai and they just move right into you as though youre not there. They do this because they truly have no idea youre there. Sometimes doing this youll notice that they all of the sudden do see you and they get all nervous looking, like a quick left right steer, or brake check.
The fact is these ai are not situational, they dont know what to do in mpst circumstances. They dont know how to pass, they dont know what to do when you drive in front of them, they dont know to allow traffic by if their being lapped, they dont know to stick to one racing line while going through a corner or going down a straight. They just pick the line theyre going to use that lap and thats kinda it, whether youre in the way or not.
As far as the “team” fixing things, i think its like 2 people one of which is Chris Esaki, who is also part of the team fixing the penalty system. I think maybe its time to get someone else on the team that may be better suited to figuring these things out.
3 Likes
Everything you need to know about the AI can be summed up in one act
Stop your car on the racing line on a clear bend
See what the AI does
Even better, stop at pit out but leave space to go around you
See what they do
4 Likes
I posted a video on here somewhere about ai and how different games compare including sims. I believe grid legends did the best on this particular test. Forza obviously didnt, but im pretty sure past forzas faired better, yet another curious downgrade.
1 Like
I get what they’re trying to show with videos like the one you mentioned (it was an interesting watch, thanks for posting it btw). But at the same time, I take things like that with a little bit of a grain of salt.
Parking sideways on the apex of a fast corner or stopping and remaining stationary in the middle of the track are, all things considered, fairly uncommon occurrences in an actual racing scenario. Sure there can be legitimate accidents where similar things happen and cars come to rest in unexpected places, but most of these tests are deliberately designed to try to “break” the AI and to show them in the worst possible light by putting them in situations that - for the overwhelming majority of races that people run - will never occur.
I think most people would agree that GT’s Sophy AI offers some of the best wheel-to-wheel racing experiences and some really enjoyable battles. Yet it still had problems when encountering stopped vehicles in the middle of the track (it performed better than some, but still wasn’t great.) I think it comes down to developer priorities and what’s going to offer the best bang for the buck in terms of CPU usage (especially in FM23’s case, which appears to have such limited available CPU overhead.) Maybe someone can develop an AI that realistically avoids stopped vehicles like that, but does it then become a tradeoff with actual racecraft like knowing how to draft, or when to implement an up and under or when to try a divebomb…the stuff that I think actually matters the most to players?
I find it interesting that the game that fared the best in terms of avoiding the stopped vehicle was Grid Legends…the least “sim” game of all of the ones tested. I think it’s a likely assumption that it also has the least going on “under the hood” in terms of physics calculations and such - is it a stretch to then think it therefore has the most headroom in terms of CPU cycles to allot to the AI for better spatial awareness and collision avoidance? Maybe. I don’t know. Maybe I’m just talking out of my you-know-what. 
In my opinion, AMS2 has some of the best AI that I’ve raced against. Sure it’s also capable of doing some insanely stupid things, as displayed in that video. But I don’t really care about that because, on the whole, in the majority of scenarios, they act believably or at least predictably enough to be enjoyable to battle with. Obviously the holy grail would be for the AI to excel in all these areas, and maybe we’ll get there someday. But, to quote the opening monologue from Silo, “all we know is…today is not that day”.
1 Like