Gran Turismo Sophy: Sony AI x Polyphony Digital

[hawkeyez]

What doesn't? Where are you trying to upload it to?

I was going to upload my AI deliberately cutting me off down a straight for no reason but my iPad doesn’t even give me the option to choose it.

I was going to upload it to here but I don’t think it’s an option. No big deal.

 

[RACECAR]

 

[sems4arsenal]

Great interview. He highlighted a key point that players believe it's easy to tone the AI down.

Essentially Sophy keeps doing laps until it becomes theoretically fastest within the constraints given. How do you slow that? Do you decrease the iterations? By how much? Even if you give it a target delta to achieve and stop it improving - you can get results that are not human like. It's a very difficult task and I really don't think the solution will present itself anytime soon.

 

[hawkeyez]

Great interview. He highlighted a key point that players believe it's easy to tone the AI down.

Essentially Sophy keeps doing laps until it becomes theoretically fastest within the constraints given. How do you slow that? Do you decrease the iterations? By how much? Even if you give it a target delta to achieve and stop it improving - you can get results that are not human like. It's a very difficult task and I really don't think the solution will present itself anytime soon.

Could they not have it meter it’s speed by lap times? If I’m running 2:00:364 on Watkins Glen and it can run 1:58:582 then let it just drive a little slower?

I’ve had really good races with people, I’ve noticed I’m faster but want a fun race so I just hold back a tiny bit and we run neck and neck for a lot of laps. Sometimes they win sometimes I win, but they’ve said “wow what an awesome race!!” Like they’ve just had the best race of their life. Obviously they didn’t know I was holding back, or vice versa, I’m sure I’ve been on the other end of that not knowing they were slow on the accelerator or a tad early on the brake or lifting a hair on the straights to keep it close.

They give us sector times, so it can measure and analyze our sectors and adjust accordingly. Now to fix the fact that yes you might have close racing but ultimately win every single time… They need to randomize an amount by how much it varies it’s lap times. If I’m running 2 minutes flat, it could randomly run anywhere from 2:01 to 1:59 or something like that.

Thats the easy part. The tricky part is what do you do if somebody wipes out? You don’t want it adjusting then. So they’d have to be able to have it recognize a slower lap time from somebody wiping out. I’m usually fairly consistent so I can run within a second either way of my average. So they can wipe out and see how much it takes on average to get back on the track so anything slower than 10 seconds or something like that it recognizes as a wipeout.

Maybe if you wipe out, it randomizes a wipe out within the time left, only whether it wipes out or not is randomized as well. Maybe you spin, lose 20 seconds, and maybe three laps later it does the same and you get your chance back. But it’s a coin toss if it actually will.

If they’re applying it to say 10 other cars, have it randomize which car is going to run close to your lap times, might be the leader, mid pack or at the back which could be interesting and happens in real races, the guy at the back slowly grabbing a position or two one lap at a time.

I know all easier said than done but maybe they’ll see this and get some ideas? That would all have to be single player, you couldn’t do that online with 5-10 other drivers. For online, I wish they’d just let us throw 3 or 4 crap AI on the track with a difficulty slider. Lol. Just for those nights where it’s only two or three of you.

 

[Famine]

Could they not have it meter it’s speed by lap times? If I’m running 2:00:364 on Watkins Glen and it can run 1:58:582 then let it just drive a little slower?

Broadly, no.

It's been trained to go as fast as it can (and rewarded for doing so), as on-track as it can, as not-into-other-cars as it can, and to finish as highly placed as it can. It doesn't know how to do anything apart from that; the code they put in the game is the product of a training regime that has resulted in the "superhuman" Sophy. It has no concept of "hey let's slow down a bit to make this more fun for the guy I'm pasting". Or indeed "fun".

And that's the challenge here: making a Sophy that's fast enough - without losing the "fairness", and cleanliness, and competitiveness - to suit humans who aren't as fast as possible. You can't simply push out less well-trained versions, because they lose all the other aspects as well as the speed; you can't fudge its speed with traditional rubber-banding methods (like turning the engine down or up) because that's what they're trying to get away from - in the Race Together demo on GT7 they just gave it worse cars, which is also not an ideal solution (as I mention in the video); players don't want the fakery.

That's without digging into training a Sophy that can adapt to weather, tyre wear, fuel loads, dirty air, nitrous/DRS/overtake, tyre strategy, etc.

 

[hawkeyez]

Broadly, no.

It's been trained to go as fast as it can (and rewarded for doing so), as on-track as it can, as not-into-other-cars as it can, and to finish as highly placed as it can. It doesn't know how to do anything apart from that; the code they put in the game is the product of a training regime that has resulted in the "superhuman" Sophy. It has no concept of "hey let's slow down a bit to make this more fun for the guy I'm pasting". Or indeed "fun".

And that's the challenge here: making a Sophy that's fast enough - without losing the "fairness", and cleanliness, and competitiveness - to suit humans who aren't as fast as possible. You can't simply push out less well-trained versions, because they lose all the other aspects as well as the speed; you can't fudge its speed with traditional rubber-banding methods (like turning the engine down or up) because that's what they're trying to get away from - in the Race Together demo on GT7 they just gave it worse cars, which is also not an ideal solution (as I mention in the video); players don't want the fakery.

That's without digging into training a Sophy that can adapt to weather, tyre wear, fuel loads, dirty air, nitrous/DRS/overtake, tyre strategy, etc.

Hmmm, maybe same car/tires/tune races (different liveries) but allowing us to adjust its power output in place of a difficulty slider? we’d still have to find the percentage that works for us but whatever. Sophy seemed a bit better than current AI for not driving into you anyway.

The biggest issue with the current AI is it’s like it doesn’t know you’re there. They can show us how close an opponent is to us but they can’t have the AI distance themselves a little using that info from radar?

I can somewhat put up with the early braking but when I accelerate out of a corner with a faster exit and it just full out tries to take my line despite the fact it’s on the preferred line to begin with? Why are you steering into me when you’re already on the line you need for the next corner? Why bother with Sophy AI when that’s your biggest problem?

Anyway I get it. I agree with the training it is pretty much impossible. It only knows the best and how to get even better. I just asked Chat GP or whatever it’s called to answer a math question slowly like a human would, it answers it in a split second. I said no, answer it very slowly, split second again. I said wait five seconds before you to answer this math question. Split second again despite saying it will answer more slowly. Lol

 

[Yard_Sale]

Hmmm, maybe same car/tires/tune races (different liveries) but allowing us to adjust its power output in place of a difficulty slider? we’d still have to find the percentage that works for us but whatever. Sophy seemed a bit better than current AI for not driving into you anyway.

The biggest issue with the current AI is it’s like it doesn’t know you’re there. They can show us how close an opponent is to us but they can’t have the AI distance themselves a little using that info from radar?

I can somewhat put up with the early braking but when I accelerate out of a corner with a faster exit and it just full out tries to take my line despite the fact it’s on the preferred line to begin with? Why are you steering into me when you’re already on the line you need for the next corner? Why bother with Sophy AI when that’s your biggest problem?

Anyway I get it. I agree with the training it is pretty much impossible. It only knows the best and how to get even better. I just asked Chat GP or whatever it’s called to answer a math question slowly like a human would, it answers it in a split second. I said no, answer it very slowly, split second again. I said wait five seconds before you to answer this math question. Split second again despite saying it will answer more slowly. Lol

To sort of build off this, one thing they could do (being mindful that I know jack-squat about any of this stuff) is exactly what they did with the original Sophy races.....which was managing its pace by way of BOP. For a more realistic experience, they could make a randomized algorithm that changes the Sophy opponents BOP on a lap by lap, or sector by sector basis.... essentially simulating real racing when you have good and bad sectors. While I think they're probably at least one Sophy generation away from being able to create an opponent that knows how to manage their tires and fuel - one way around this would just to be to just pre-determine the amount of tire wear per Sophy opponent for different stages of the race. I'm sure they already do this to an extent in normal AI races.

I always felt that the most practical implementation for Sophy into the game on a broader scale, while being inside the confines of the current hardware/software parameters.... would be too implement a few Sophy opponents amongst the AI that PD has already proved they can turn WAAAAAAY UP - as demonstrated in the Human Comedy missions, clubman+ races, and the recent super formula races.

Maybe its just me, but throughout those Sophy races, I still felt like the AI in ACC was much more realistic to race against. If any of you guys have the ability to play that game, I strongly recommend it. The scaleable AI in both skill and aggressiveness is pretty impressive





edit: I'm also sure they could incorporate chassis changes into this "dynamic-randomized-BOP" such as toe angles and such, that might be able to coax Sophy into some time-costing mistakes. But then for that to even present and somewhat realistic, said BOP would have to be randomized on a sector by sector basis.

edit 2: Just took one of the doggies out for a walk and got to thinking about it more. You could accomplish everything that you would need to simulate on how Sophy handles worn tires, fuel loads, ect. (not sure about weather) by working in the parameters of the current tuning menu adjustments. You could manipulate fuel loads by way of ballast. You could manipulate worn tires by kicking out toe angles. You could artificially punish them taking the cheater lines by dynamically changing the suspension stiffness and rebound so it gets sent flying when it cuts corners. Not sure if tire wear can influence the physics parameters that Sophy functions under, but if it does.... I wonder if they could program in a system that rewards the risk of her trying to stay out on worn tires.

Interesting interview to say the least.

 

[Techlet]

I wonder if they're able to adjust Sophy's reaction speed if that's even a thing? Slow down its steering and pedal inputs or just its decision making in general to slow it down on track, mimicking a more inexperienced driver.

 

[hawkeyez]

To sort of build off this, one thing they could do (being mindful that I know jack-squat about any of this stuff) is exactly what they did with the original Sophy races.....which was managing its pace by way of BOP. For a more realistic experience, they could make a randomized algorithm that changes the Sophy opponents BOP on a lap by lap, or sector by sector basis.... essentially simulating real racing when you have good and bad sectors. While I think they're probably at least one Sophy generation away from being able to create an opponent that knows how to manage their tires and fuel - one way around this would just to be to just pre-determine the amount of tire wear per Sophy opponent for different stages of the race. I'm sure they already do this to an extent in normal AI races.

I always felt that the most practical implementation for Sophy into the game on a broader scale, while being inside the confines of the current hardware/software parameters.... would be too implement a few Sophy opponents amongst the AI that PD has already proved they can turn WAAAAAAY UP - as demonstrated in the Human Comedy missions, clubman+ races, and the recent super formula races.

Maybe its just me, but throughout those Sophy races, I still felt like the AI in ACC was much more realistic to race against. If any of you guys have the ability to play that game, I strongly recommend it. The scaleable AI in both skill and aggressiveness is pretty impressive





edit: I'm also sure they could incorporate chassis changes into this "dynamic-randomized-BOP" such as toe angles and such, that might be able to coax Sophy into some time-costing mistakes. But then for that to even present and somewhat realistic, said BOP would have to be randomized on a sector by sector basis.

For me the biggest thing is that their current AI is just not good. Figure out a way to make the races fun, automatic scaling to how your race is going and you’ve nailed it. I want fun races whether I’m on top of my game or not.

I have played ACC, PC1 and PC2, GT Legends and many more. Gran Turismo has never been great. PC2 you could dial it in and get good racing, but as you got better, you’d have to do races to find out, ok better up it a couple percent on this track with this car now. Maybe you had an off day, you’d turn it back down. I’m sure they could have it analyze lap times and adjust on the fly. Imagine not having to do research ever track and every car combo to get it dialled in.

Program it to give the illusion it’s real and keep it fun which means not crashing into me. The only thing I felt I noticed with Sophy was it was more tuned in to where I actually was. More reactive to my driving. Current AI just bonks into you while you’re both going straight for some reason. Arrgghhh I did nothing wrong and there goes my clean race bonus!! AGAIN!!! Haha

Fix current AI so it doesn’t drive right into you, scale power output on the fly accordingly to match current sector times / lap times within a certain percentage. Fun racing is close racing, not train track single file around the course. And in GT7’s case, it’s single file several car lengths apart passing one opponent at a time while dancing on the edge of losing your clean race bonus.

But you’re right, ACC, a lot of their games have much better AI. If they could throw PC2‘s AI in, drop the clean race bonus and just payout more I’d be satisfied. It’s not rocket science. Lol

 

[Yard_Sale]

Someone correct me if I'm wrong...

But despite the car list that GT1 had, you were only driving a dozen or so unique cars, just dressed in different skins....right?

I think its important to manage expectations. One thing they could do to get around Sophy learning every single car, is just to find a handful cars that are versatile enough via BOP adjustments to cover a large swath of vehicles..... and have Sophy learn those. Then just dress said cars in different skins and program in sounds that are unique to each vehicle.

What does it matter to us as a player, if Sophy is actually driving a 95' Porsche Carrera or not? I think the only thing would we care about, is if said car that Sophy was driving, looked and sounded the part.

 

[hawkeyez]

Someone correct me if I'm wrong...

But despite the car list that GT1 had, you were only driving a dozen or so unique cars, just dressed in different skins....right?

I think its important to manage expectations. One thing they could do to get around Sophy learning every single car, is just to find a handful cars that are versatile enough via BOP adjustments to cover a large swath of vehicles..... and have Sophy learn those. Then just dress said cars in different skins and program in sounds that are unique to each vehicle.

What does it matter to us as a player, if Sophy is actually driving a 95' Porsche Carrera or not? I think the only thing would we care about, is if said car that Sophy was driving, looked and sounded the part.

I don’t even care if Sophy drives the same car I guess. Same style maybe? Like what would normally be competitive with whatever we chose. If she’s always in the dodge challenger when I take the Camaro or Mustang who cares. But it’s a very good idea to just make it appear like it’s something else.

 

[24HuresdeLoL]

Could they not probably implement a slider correlating to say driver level? Say if someone is a D-Rated player, could they not just take all data from all D-Rated players and implement on that? Each level is a higher performance, so get a C-Level, it'll match up to a tenth or past a tenth , B-Rated same thing, and so on?

 

[happycorey]

I think the problem still exists and what they need is an entirely new algorithm that satisfies all of those variables in a dynamic fashion. Sophy is a 1 track 1 car pony and Sophy is not able to deviate from that without significant compromise of it's function.

 

[BlackHGT]

Perhaps it is due to the wrong task in the development of the AI. It seems that the task is:

• get around the track as quickly as possible
• Avoid collisions

But the human aspects are missing. Which I expect from an AI in a racing game.
Lack of concentration, small deviations in the braking point, turn-in point, turn-in angle, etc.... The emotion of getting angry when you make a few mistakes and the driver in front drives away, which can lead to more mistakes. Or the pressure that rises when you catch up but slowly realise that there are only a few laps left, and you try to speed up a bit... and like most things it gets worse.

Which in my opinion would be the right task:
An AI that is not as fast as possible, but an AI that is as human as possible...

 

[KAlex122]

I found the public test some month ago handled the perfect pace problem well by giving players one or two notches better tires. On hard mode you had for example Sport Soft tires while Sophy used Sport Medium. So Sophy had a harder time getting corners right which made the reaction and racing pretty believable without Sophy being held back on the straights.

 

[pornomazinger]

I found the public test some month ago handled the perfect pace problem well by giving players one or two notches better tires. On hard mode you had for example Sport Soft tires while Sophy used Sport Medium. So Sophy had a harder time getting corners right which made the reaction and racing pretty believable without Sophy being held back on the straights.

but my friend, Sophy NEEDS harder tyres, as it drifts in all turns.....🤣
they just vanish in a couple of laps 🤣

 

[hawkeyez]

I found the public test some month ago handled the perfect pace problem well by giving players one or two notches better tires. On hard mode you had for example Sport Soft tires while Sophy used Sport Medium. So Sophy had a harder time getting corners right which made the reaction and racing pretty believable without Sophy being held back on the straights.

Mentally it’s not as satisfactory saying I kicked Sophy’s butt running grippier tires!! Haha.

Other games in my opinion have much more immersive AI and that’s without it being chase the rabbit, you’d think PD would have an advantage there essentially only needing to have one on one reactions all the time. So maybe they should look into how they’re doing it and learn something rather than releasing what essentially feels like the same game each release.

There’s been some nice changes and not so nice changes, for me mostly whats changed is the handling. First booting up GT7 when it went live, I said, ehhh it’s like the same old game it’s always been but it looks great and the car physics feels a lot better. The actual ”game” part of it, like championships, AI chase the rabbit style racing, grind for credits, outrageous car prices, not being able to share tunes etc etc etc the list is long, it is the same old same old garbage. The same frustrations with AI IMO are still there they always have been, it feels copied and pasted.

So I got excited when they announced Sophy but I got my hopes up. Maybe it’s just me, all I noticed was, hey Sophy reacts a little better to my driving. It’s too bad althe AI can’t just do that.

 

[evldave333]

Im wondering as they are really breaking new ground with Sophy are our expectations too high, the subtle nuances we may want could be possible with some kind of real time processing rather than a short hand algorithm which seem to be about all that's possible for now. Its fascinating to watch the developments and im looking forward to seeing what is produced eventually. There are many parallels with driverless cars. Maybe a gargantuan next generation (quantum) level of processing will make real time AI for every vehicle on a grid and they will behave a bit like a Starling murmuration adjusting collectively and individually in real time.
For now there are so many rich pickings to be had, such as perfect / cheat driving lines to learn, and clean overtaking (yay!).

 

[Yard_Sale]

I’m almost positive that Sony is using this research for many different applications other than Gran Turismo, or video games in general.

It just might be that Gran Turismo might just happen to be the test bed, because they’d get the biggest swath of real life and real time data acquisition, and accolades by testing with the gamer crowd.

Perhaps even they’re developing this technology with own driverless cars somewhere in the future, in mind.

Who knows

 

[rlx]

Thanks for the interview and the video.

It's probably worth correcting one of your misconceptions about Sophy. While the vanilla AI observes specific rules about specific race conditions and situations, Sophy does not follow a more sophisticated or complicated set of machine-generated "if-then" rules. It is being trained using reinforcement learning, meaning that it races against copies of itself and starts with zero knowledge, other than the positions of itself and its opponents on the track. It is then rewarded for fast lap times (i.e. winning races), and penalized for track excursions, collisions, etc. The result of this training process is a neural network that can "solve" the following mathematical relation: given this set of environment variables (position, orientation and velocity of cars and coordinates of the surrounding track boundaries), what are the optimal numbers for brake/throttle (from -1 to 1) and left/right steering (from -1 to 1).

Given this architecture, Sophy should be able to learn to deal with weather and race strategy - by adding more environment variables, like the amount of water on the track, weather forecast, tyre type and status, percentage of fuel in the tank and number of remaining laps. Given these inputs, Sophy would eventually learn to avoid sliding, pit at the right moment, fuel save and choose the best tyre compound. (The vanilla AI doesn't just lack intelligence in overtakes or defending positions, but is also pretty stupid in terms of strategy - like making a 40 second pit stop just before the end of an endurance race to switch to the optimal tyre compound that will save 10 seconds on the final remaining lap. Sophy should be able to handle such situations correctly.)

It's worth noting that while it's relatively easy to train a version of Sophy that drives more (or less) aggressively by lowering (or raising) the penalty for collisions during training, it's harder to create a slower version of Sophy, since you can't just easily de-emphasize the reward for fast lap times. The most obvious way would still be to give it a technical handicap (slower car, slower tyres, higher fuel consumption). And while Sophy's knowledge should generalize from one track to almost any other (it knows how to deal with straights, left turns, right turns, tight turns, wide turns, banked turns etc), it won't generalize from one car to others, since the result of throttle/brake and steering input can vary pretty substantially from car to car. And given that the car physics in GT7 are quite elaborate, there's not much hope that you could feed the geometry/weight distribution of the car, transmission/suspension/downforce settings, type of brakes and engine characteristics to Sophy and hope for it to figure it all out.

EDIT: Here's a link to Sony AI's paper - worth reading if you're interested in the tech behind Sophy. https://www.researchgate.net/public...ismo_drivers_with_deep_reinforcement_learning

 

[bduddy]

I would think you could train a "slower" Sophy by just rewarding it hitting a certain lap time, rather than going as fast as possible, but I'm sure the Sony AI team has thought of that. I suspect that Sophy might find very strange and non-human like ways to hit those lap times (stopping right before the finish line?) that wouldn't lead to a fun racing experience, and it would be even harder than the current solution to replicate across different cars and tracks. I do think it's -possible- to figure it out, but ironically, it's a lot harder than telling Sophy to "just go fast".

 

[Yard_Sale]

Thanks for the interview and the video.

It's probably worth correcting one of your misconceptions about Sophy. While the vanilla AI observes specific rules about specific race conditions and situations, Sophy does not follow a more sophisticated or complicated set of machine-generated "if-then" rules. It is being trained using reinforcement learning, meaning that it races against copies of itself and starts with zero knowledge, other than the positions of itself and its opponents on the track. It is then rewarded for fast lap times (i.e. winning races), and penalized for track excursions, collisions, etc. The result of this training process is a neural network that can "solve" the following mathematical relation: giving this set of environment variables (position, orientation and velocity of cars and coordinates of the surrounding track boundaries), what are the optimal numbers for brake/throttle (from -1 to 1) and left/right steering (from -1 to 1).

Given this architecture, Sophy should be able to learn to deal with weather and race strategy - by adding more environment variables, like the amount of water on the track, weather forecast, tyre type and status, percentage of fuel in the tank and number of remaining laps. Given these inputs, Sophy would eventually learn to avoid sliding, pit at the right moment, fuel save and choose the best tyre compound. (The vanilla AI doesn't just lack intelligence in overtakes or defending positions, but is also pretty stupid in terms of strategy - like making a 40 second pit stop just before the end of an endurance race to switch to the optimal tyre compound that will save 10 seconds on the final remaining lap. Sophy should be able to handle such situations correctly.)

It's worth noting that while it's relatively easy to train a version of Sophy that drives more (or less) aggressively by lowering (or raising) the penalty for collisions during training, it's harder to create a slower version of Sophy, since you can't just easily de-emphasize the reward for fast lap times. The most obvious way would still be to give it a technical handicap (slower car, slower tyres, higher fuel consumption). And while Sophy's knowledge should generalize from one track to almost any other (it knows how to deal with straights, left turns, right turns, tight turns, wide turns, banked turns etc), it won't generalize from one car to others, since the result of throttle/brake and steering input can vary pretty substantially from car to car. And given that the car physics in GT7 are quite elaborate, there's not much hope that you could feed the geometry/weight distribution of the car, transmission/suspension/downforce settings, type of brakes and engine characteristics to Sophy and hope for it to figure it all out.

EDIT: Here's a link to Sony AI's paper - worth reading if you're interested in the tech behind Sophy. https://www.researchgate.net/public...ismo_drivers_with_deep_reinforcement_learning

Interesting read. Are you part of the team developing this?

 

[rlx]

I suspect that Sophy might find very strange and non-human like ways to hit those lap times (stopping right before the finish line?)

Ironically, stopping right before the finish line, when leading in an endurance race, to let the timer run out rather than run another lap, is what humans do routinely in Gran Turismo. If Sophy learns to do that, I guess it could be considered human-level intelligence

(Or if it could handle the rare occasion where crossing the line and running a final lap on empty tank wins the race, while entering the pits loses the race. This would be the obvious choice for many humans, but requires some general intelligence, beyond just racing.)

Interesting read. Are you part of the team developing this?

Oh no, not all all. I've just been working with neural networks for a while, mostly as a hobbyist, so deep learning for games AI is something I'm generally curious about.

PS: And while it's popular to dump on the current GT AI, especially now that we've gotten a glimpse of Sophy, I'm actually still quite impressed by it, given that it's a "dumb", rule-based system. Of course, if you're a halfway decent driver, it's simply not competitive, some of its strategic blunders can be infuriating, and the rubber-banding and other cheats can be so obvious that they break immersion. But I still quite admire the fact that in heavy rain, some but not all cars slide off the track, or that some but not all cars tend to crash at Schwedenkreuz, or that some cars bump into you for what looks like lack of skill, while others successfully avoid you. I once managed to make Sophy crash, which was quite satisfying, and I still think it would be a plus if from time to time, it would make unforced errors or sub-optimal moves.

 

[hawkeyez]

Well I loaded up my custom race on professional with muscle cars I’ve tuned up to match a good Camaro tune I have. They’re all different PP ratings but I run pretty close to the same lap times with them all. 5 laps. Was in third place by the end of lap 1. It was a fight to try and catch 1st position. And I experienced the rubber banding. I was thinking I’m never going to catch that one. Suddenly I make up so much time like it’s nothing, pass it, all of a sudden it’s passing me back. I will say it was a good race, but only for lap 5.

I would like to know if that rubber banding is an effect of not comparing it’s lap times to mine often enough? I was running pretty consistent. When I’m on my game sometimes my laps are within 0:00:100 of one another. So it’s not like I was messing with its calculations if it even has any? But to confirm, the last lap was awesome, lap 2-4 wasn’t boring as ever. And why did only one car be able to keep up? Something weird going on there.

 

[Yard_Sale]

A surprisingly hard AI race is an arcade race in group 3 cars at Interlagos. The AI is pretty damn good there. Same with the WTC 800 race at the ring

 

[Auditore]

Well i would go the simplest route,add weight to the Sophy cars.
Instead of manipulating the AI to be worse you just BOP the cars with extra weight.
That's something PD can already do

 

[equalitus]

Broadly, no.

It's been trained to go as fast as it can (and rewarded for doing so), as on-track as it can, as not-into-other-cars as it can, and to finish as highly placed as it can. It doesn't know how to do anything apart from that; the code they put in the game is the product of a training regime that has resulted in the "superhuman" Sophy. It has no concept of "hey let's slow down a bit to make this more fun for the guy I'm pasting". Or indeed "fun".

And that's the challenge here: making a Sophy that's fast enough - without losing the "fairness", and cleanliness, and competitiveness - to suit humans who aren't as fast as possible. You can't simply push out less well-trained versions, because they lose all the other aspects as well as the speed; you can't fudge its speed with traditional rubber-banding methods (like turning the engine down or up) because that's what they're trying to get away from - in the Race Together demo on GT7 they just gave it worse cars, which is also not an ideal solution (as I mention in the video); players don't want the fakery.

That's without digging into training a Sophy that can adapt to weather, tyre wear, fuel loads, dirty air, nitrous/DRS/overtake, tyre strategy, etc.

Players should absolutely be able to choose slower car [types] and/or tyre type for the AI cars. It's recurrent that even that GT7 have more options than GT6&GTS everything should be able to be selected. And any GT AI should be able to be adjusted between easy&hard in single player. A pity that we didn't get to do that ourselves in GT5/GT6 without hacking the game data.

Someone correct me if I'm wrong...

But despite the car list that GT1 had, you were only driving a dozen or so unique cars, just dressed in different skins....right?

I think its important to manage expectations. One thing they could do to get around Sophy learning every single car, is just to find a handful cars that are versatile enough via BOP adjustments to cover a large swath of vehicles..... and have Sophy learn those. Then just dress said cars in different skins and program in sounds that are unique to each vehicle.

What does it matter to us as a player, if Sophy is actually driving a 95' Porsche Carrera or not? I think the only thing would we care about, is if said car that Sophy was driving, looked and sounded the part.

There should be info on the cars that Sophy is experienced in, so that we can choose those cars if we want to.

 

[Famine]

Players should absolutely be able to choose slower car [types] and/or tyre type for the AI cars. It's recurrent that even that GT7 have more options than GT6&GTS everything should be able to be selected.

Yes, they should be able to do that if they want but that shouldn't be the only option.

A lot of people don't like racing against artificially hobbled cars for very good reason: it's artificial. If you're entering a GT500 race in a GT500 car, having the AI plopped into GT300s feels phony. They want to race GT500 against GT500, with the same tyre options and the same power.

At the moment, Sophy can't do that; Sophy can only go as fast as it can and that means you either have a superhuman, top 0.01%er trouncing you in the same car, or it has to have significantly worse cars to allow for 5-50%ers a chance to race against it.

And any GT AI should be able to be adjusted between easy&hard in single player.

Yeah, I mean, that's what I said - and the issue with Sophy is that, at least for now, it can't be. It's an issue Sony AI is trying to solve but I covered the reasons why it's an issue and why it's hard to solve in my post: Sophy is trained to be superhuman in speed, fairness, cleanliness, and competitiveness, and you can't train a new instance of it to be less of one without it becoming less of the others too. I've had this conversation with them (in person).

Hopefully we'll see a new in-game "Race Together" event before too long, as Sony AI is always developing it and looking at what happened in the first one. It'll be interesting to see how it's come along.

 

[BallPtPenTheif]

I hope they're not wasting time trying to figure out how to load an entire grid of Sophy cars. We really don't need it. We just need 1 sophy car ahead of us and 1 behind us. Then you can dynamically drop the AI into leading and trailing cars as the player works through the pack. Sophie worked great during the trial races, and I don't see why we can't be having Sophy races sooner rather than later.

Doing this, you really wouldn't even have to throttle Sophy's abilities that much, as the lower-powered cars and dummy AI (ahead of the pack), would rubberband the player into the mix naturally, making passing and defending the real challenge.

 

[MroczMR]

About adding Sophy to GT7 campaing. I know they said they are struggling to make it scalable.... yes but to be honest, is it really need to be scalable now? I would add Sophy as a 4th extreme difficulty mode, add qualifying to the campaing races and I think most GT fans would be incredbile satisfied with this.
Especially that most Campaing races gives you a big freedom with selecting a car for the event, you can easily scale most of the events by choosing more powerfull car, I think this could be the best way to add sophy without artificially "slowing it down"