Please PD, about AI

[Flowering00]

Can we please just have more AI on track, both in Arcade and Simulation mode? Would be pretty amazing having ~50-60 and even more cars on Le Mans, Daytona and other famous endurance racing circuits.

Just simply make it an option only on ps5 version of the game. Because I don't believe that ~9 years CPU on ps4 can handle it.

As well as perhaps remove raytracing in a replay if there is such amount of AI as even with 20 cars in heavy rain at night/overcast ps5 can't really handle that much shininess.

It'll allow for an ultimate itteration of endurance racing in GT series. Compare to what's given in previous installments.

Such as Nurburgring 24 Hours Endurance Race.
Multiclass, which can be fully implemented (maybe at half of it) from official said race in real life.
-Gr.3 (SP9)
-Gr.4 (SP10)
-And other cars, which I still have no clue why PD haven't implemented them yet, as they even sponsor that race, such cars as Porsche 911 Cup Car, KTM X-Bow GTX, Newer GT3 cars.

As well as Daytona and Lemans 24 Hours races can be in Gr.1/Gr.3 Multiclass respectively. Also classic itterations of them, which can finally utilize all of these legendary cars, such as Porsche 917K '70, Ford Mark IV '67 and etc.

Notably, PS4 version can have 24 cars on the grid.

This can solve issues for quite amount of time, such as:
-Grind races/Earning Credits
-Use of more cars in events, especially legendary ones
-Use of latest hardware
-Finally Endurance Racing

Also I think length of the races can be adjusted, player can choose of minimum 2.4 hours to full 24 hours.

For example one race can give 10,000,000 Cr. in full-cycle race and shortest one gives 10x less of that.

Also honestly just please make an option to change livery on AI in "choose car" option in Custom Races, so you wouldn't have to buy same cars.

 

[Kaiserben]

AI can't even manage basic racing rule like blue flag... Let alone the 60 cars...

 

[BangsMcCoy]

With as bad as AI currently is in GT7 I can imagine how that would go.

 

[Barracuda413]

I would much rather have fewer "good" AI, then more "bad" AI,
it seems like GT7 just cut and pasted the AI from an XBOX360 forza game

 

[Flowering00]

AI can't even manage basic racing rule like blue flag... Let alone the 60 cars...

Really? I honestly seen some AI passing Subaru 360 with ease to be honest.

 

[Scaff]

Nurburgring 24 Hours Endurance Race.
Multiclass, which can be fully implemented (maybe at half of it) from official said race in real life.

That race has 200 cars starting each year across more that 20 classes! GT7s not capable of getting even close to half of that.

 

[Kaiserben]

Really? I honestly seen some AI passing Subaru 360 with ease to be honest.

This is really bad. Worse than bad, it's non-existent. In fact they don't care at all of the blue flag : they try to put you out / get on your line when you lap them. I would not imagine what could it be on an endurance race where you will lap a lot of GR3 car with your GR1. Imagine losing the race due to a poor AI, make you crash in the straight full speed, on purpose.

 

[Artremis]

Even on PS5 the game can have framerate drops with a maxed out grid already, I'd like to see how a 60 car grid would run like.

 

[Superhero Wally]

AI can't even manage basic racing rule like blue flag... Let alone the 60 cars...

To be fair. It seems like they have at least made an adjustment (update1.21) on the blue flag, and the AI, seems to do a much better job of letting you by when your lapping them now.

 

[Talon16]

 

[kraftwerkd]

It's not only graphics to worry about, that's a more intense CPU load because it now has to calculate physics for 3x as many cars. Even Assetto Corsa Competizione starts to chug with 30 cars, mind you there are more parameters in the physics system in that game. But it's a rough comparison.

I doubt we're there yet.

 

[fokuzx2]

please PD make GT7 a racing game and not a "how many cars can i pass in 3 laps" game. to me it seems that the AI is still from the PS2 era...

 

[Luksz92]

The AI is one of the weakest parts of the game - sticks to a line, and doesn't adapt at all. Will literally ram you to get back onto the line. Not to mention the elastic band jumping ahead 🤦‍♂️ Would want this addressed long before considering larger grids.

 

[fokuzx2]

GT has always been a game that lived from the singleplayer experience. i can not believe no effort has been put into the AI for the last 15 years or so

 

[BrunUK]

It's not only graphics to worry about, that's a more intense CPU load because it now has to calculate physics for 3x as many cars. Even Assetto Corsa Competizione starts to chug with 30 cars, mind you there are more parameters in the physics system in that game. But it's a rough comparison.

I demonstrated years ago that the FPS hit with larger fields in ACC is almost entirely down to the graphics side of things rather than anything to do with physics or AI. On the PC at least, dunno if the console versions have the option necessary to check it in the same manner.

To confirm, all you have to do is set up a race with maximum opponents and change the 'visible opponents' setting. Just checked again now and get ~100 FPS with opponent visbility max, and ~160 FPS with only a single other car visbile. Physics and AI calculations in either case should be identical, the other cars just aren't being drawn. Starting the same race with only one opponent gets 180 FPS, so the framerate difference based on the actual number of opponents isn't very significant.

 

[JP76]

The AI is one of the weakest parts of the game - sticks to a line, and doesn't adapt at all. Will literally ram you to get back onto the line. Not to mention the elastic band jumping ahead 🤦‍♂️ Would want this addressed long before considering larger grids.

Correct.
The 'AI' in this game is just a blind mobile obstacle course.

That Sophy needs vast improvement too if they ever plan on releasing that.
The lap I saw of it do around the Nordschleife was a ridiculous mess, it spent almost half the lap driving off the track.
We get penalised and DR destroyed for slight and ridiculous 'track limits', especially when you gain no time in Sport mode, so currently 'Sophy' is a slap in the face and a piss take.

 

[Pizzapants]

ah yes instead of 15 AI that all somehow have 5 times my car's mass, glued to their racing line, braking at the apex of turns and swerving into my line when i'm lapping them, i can have 50 to 60 AI doing that! can't wait! sounds super fun!

 

[Nebuc72]

Correct.
The 'AI' in this game is just a blind mobile obstacle course.

That Sophy needs vast improvement too if they ever plan on releasing that.
The lap I saw of it do around the Nordschleife was a ridiculous mess, it spent almost half the lap driving off the track.
We get penalised and DR destroyed for slight and ridiculous 'track limits', especially when you gain no time in Sport mode, so currently 'Sophy' is a slap in the face and a piss take.

It has to be said that Sophy only exploits what the game allows her to. The game doesn't punish you for driving on the grass in those instances on the Nordschleife, and as it was faster for her to do so - that's what she does.

That's a failure of track limits, not of the AI. You would be faster doing the same thing.

 

[Scaff]

That's a failure of track limits, not of the AI. You would be faster doing the same thing.

And physics, a massive failure of physics, don't forget that either.

 

[Pizzapants]

It has to be said that Sophy only exploits what the game allows her to. The game doesn't punish you for driving on the grass in those instances on the Nordschleife, and as it was faster for her to do so - that's what she does.

That's a failure of track limits, not of the AI. You would be faster doing the same thing.

And physics, a massive failure of physics, don't forget that either.

any trained AI will fully and completely exploit any and all glitches, cheats, limitations, or other opportunities available to it to achieve its programmed goals. If you hate GT7's current AI, you're going to absolutely despise any machine learning AI. It won't drive anything like what you want it to. It will cheat its ass off, and drive dirtier than you ever thought possible.

 

[KAlex122]

I am no expert and i don't know anything about the actual state of Sophy but i would assume it's not that hard to tone it down.

More realistic track penalties and a no touch policy against other cars and thats it....

 

[Scaff]

any trained AI will fully and completely exploit any and all glitches, cheats, limitations, or other opportunities available to it to achieve its programmed goals.

Programmed goals are the key point here.

If you hate GT7's current AI,

Odd as GT7's current AI has nothing to do with Sophy

you're going to absolutely despise any machine learning AI. It won't drive anything like what you want it to. It will cheat its ass off, and drive dirtier than you ever thought possible.

That depends entirely on the goals it's set to achieve, and how it's then scaled so it's certainly not an absolute at all.

 

[Nebuc72]

And physics, a massive failure of physics, don't forget that either.

Very true, though the Nords video was on Sport and I'd like to think it wouldn't work as well in 7, as whilst the game does get a lot wrong - the driving physics are pretty solid.

Big step above Sport in any case (imo).

 

[Scaff]

Very true, though the Nords video was on Sport and I'd like to think it wouldn't work as well in 7, as whilst the game does get a lot wrong - the driving physics are pretty solid.

Sort of, still a lot to do

Big step above Sport in any case (imo).

Yes

 

[Pizzapants]

.

 

[Imari]

any trained AI will fully and completely exploit any and all glitches, cheats, limitations, or other opportunities available to it to achieve its programmed goals. If you hate GT7's current AI, you're going to absolutely despise any machine learning AI. It won't drive anything like what you want it to. It will cheat its ass off, and drive dirtier than you ever thought possible.

I mean, that's the difficulty behind using ML AIs. If it was as easy as just pointing a bunch of neural networks at a problem and letting them optimise the **** out of it, we'd be done already. The problem is correctly constraining the learning to produce the desired outcome, ie. something that actually behaves somewhat like a human driving a car, and that's not as simple as it sounds.