Gran Turismo Sophy: Sony AI x Polyphony Digital

[Imari]

No, we saw an unbeatable Sophy here:

Redirecting to https://sonyresearch.github.io/gt_sophy_public

This version beat every pro player in every race and time trial.

Buddy, we just had this discussion about accurate information. You couldn't go one post without saying something that was obviously wrong.

It did not beat every pro player in every race and time trial. Some human players finished ahead of some Sophy players. It performed very well and a Sophy was the winner in each of those races, but it was not unbeatable.

Last chance - can you stop exaggerating and use accurate information? Or is this just a waste of time having a discussion with someone who is making up facts to suit their argument?


You'll also note that the Sophy that wins started on pole position in every single one of those races. In a relatively short race between two drivers of near equal skill, track position means a lot. If the humans had started on pole, would they have been able to defend and win every race? Maybe. It's a big advantage to be only giving to one side.

Also, you might want to watch the last race. The Sophy that wins went off at the first major corner. It had to catch back up from 7th. It does so because everyone else is also crashing and getting penalties. That doesn't sound like an unbeatable AI to me, it sounds like an AI that happened to be the least bad of everyone on track. It was fast, but it could have been beaten.

And now we are racing 1 vs 1 Sophy and are able to beat it.

You're able to beat it by getting ahead early and blocking the hell out of it, but very few people are beating it on pure pace.

Compared to what you see in those videos Sophy has almost certainly been retrained to avoid the player at almost any cost, because people would absolutely complain about Sophy defending that hard. It's not noticeably slower than it was, it's just much more timid with regards to contact. Which means that when the player is close and particularly when the player is ahead, Sophy is at a huge disadvantage. It has to play by rules that the player does not.

People are beating Sophy not because they're faster than it, but because they're able to exploit flaws in the AI. I could beat Lewis Hamilton too if I was allowed to block him and ram him off the road and he wasn't allowed to make any aggressive moves. That doesn't mean that I'm a faster driver than Lewis Hamilton.

Oh and I'm absolutely not insecure about my English, I'm reacting to the fact you're saying I should take English classes which is a weird ad hominem IMO.

It's an ad hominem to point out the solution to a problem that you're very clearly having with communication? A problem that you also admit exists? Okay, buddy.

If this is the level of language you're comfortable using to accuse me of things, you get absolutely no allowances made for language skill. I'd say that you need to look up what an ad hominem is, but you'd probably take that as an ad hominem as well. But if someone replies and it's clear that they haven't understood the question, it's not an ad hominem to point that out.

What would be stupid would be to reply in good faith to things like "Sophy can't be running on PS5, the training would make a PS5 explode" instead of pointing out that there has been some miscomprehension in the information that led the person to come to that conclusion.

What I find fascinating about this is that Sophy is, 100%, driving dirty by any measure of any thread of online racing. If it were human, there would definitely be a several page long thread about these tactics. People avoid online because of it.

But now it's all cool.

I don't understand.

Sophy might drive dirty sometimes, and you definitely have to drive dirty to beat it, at least the 1v1. But the GT AI has been so garbage for so long that this counts as a major step up.

When you've been drinking your own urine for years, a glass of slightly cloudy water seems magical.

 

[dabz343]

Buddy, we just had this discussion about accurate information. You couldn't go one post without saying something that was obviously wrong

Last chance - can you stop exaggerating and use accurate information? Or is this just a waste of time having a discussion with someone who is making up facts to suit their argument?

Have you recently walked outside and dealt with people in this tone?

When you've been drinking your own urine for years, a glass of slightly cloudy water seems magical.

You are your worst enemy.

 

[Gingerale]

Buddy, we just had this discussion about accurate information. You couldn't go one post without saying something that was obviously wrong.

It did not beat every pro player in every race and time trial. Some human players finished ahead of some Sophy players. It performed very well and a Sophy was the winner in each of those races, but it was not unbeatable.

Last chance - can you stop exaggerating and use accurate information? Or is this just a waste of time having a discussion with someone who is making up facts to suit their argument?

Because it were different versions obviously.
Check out this video, this was the stage they finally managed to create a superhuman AI. This version is not the version we are racing now.

You'll also note that the Sophy that wins started on pole position in every single one of those races. In a relatively short race between two drivers of near equal skill, track position means a lot. If the humans had started on pole, would they have been able to defend and win every race? Maybe. It's a big advantage to be only giving to one side.

This is simply not true. Sophy crashed in the first corner of the La Sarthe race, they proceeded to win. It beat the pro players and it was not a slow race. That version of Sophy was constantly able to beat about every human player about every time.

Also, you might want to watch the last race. The Sophy that wins went off at the first major corner. It had to catch back up from 7th. It does so because everyone else is also crashing and getting penalties. That doesn't sound like an unbeatable AI to me, it sounds like an AI that happened to be the least bad of everyone on track. It was fast, but it could have been beaten.

No, it was able to do that because it was 1.4 seconds faster than the human players per lap. That is a lot.

You're able to beat it by getting ahead early and blocking the hell out of it, but very few people are beating it on pure pace.

This is not true, Sophy is definitely not running the perfect racing lines everytime. Check the death chicane video , it passed the wall with a millimeter every time. Right now it clearly is not running the perfect laps.

Compared to what you see in those videos Sophy has almost certainly been retrained to avoid the player at almost any cost, because people would absolutely complain about Sophy defending that hard. It's not noticeably slower than it was, it's just much more timid with regards to contact. Which means that when the player is close and particularly when the player is ahead, Sophy is at a huge disadvantage. It has to play by rules that the player does not.

So there are different versions? And it is able to be programmed to be slower?
They even tell it in the videos the later version wasn't scared to backdown.

People are beating Sophy not because they're faster than it, but because they're able to exploit flaws in the AI. I could beat Lewis Hamilton too if I was allowed to block him and ram him off the road and he wasn't allowed to make any aggressive moves. That doesn't mean that I'm a faster driver than Lewis Hamilton

Lol this is the completely opposite of what I have been saying. If you use these semantics than act like it 'buddy', youre posting like youre in a fight. You're not, it's a verbal discussion. I'm sure you could do without and have a talk about Sophy without this behavior which make you look like a desperate guy wanting to win every argument. Read your posts and think is this what I want to come over like. Just a tip
Your posts are filled with fallacies

 

[ShiftingGears]

There's two sophy threads but the main issue is that PD have spoken about scalability of Sophy for months but there's no evidence for this in the testing that this update is supposedly providing. There's no grid starts outside the 1v1 either - what's the point?

The only variance in Sophy's difficulty is either by the car it pilots or the tyres the car has, while the difficulty for the player is adjusted by the tyres that are on the player car.

Without adjustability, Sophy isn't a sustainable solution for the GT AI.

 

[Griffith500]

The major difficulty with all machine learning algorithms is obtaining enough high-quality, appropriate data with which to train the model. To the point that synthetic data generation is now a major field of active research.


If only there were some way to test Sophy against thousands, if not millions of real human players of varying ability under controlled conditions.

 

[dabz343]

The knowledge funnel -- Mystery, Heuristic then algorithm. Premature conclusions always lead to mistakes.

Sry, couldn't resist.

 

[ScottPuss20]

The major difficulty with all machine learning algorithms is obtaining enough high-quality, appropriate data with which to train the model. To the point that synthetic data generation is now a major field of active research.


If only there were some way to test Sophy against thousands, if not millions of real human players of varying ability under controlled conditions.

Perhaps the information they get from the current "seasonal" event will help PD to do this somewhat.

 

[Duke_of_hazard]

Race 4 Expert. It's virtually impossible to win unless you have at least a decent wheel set up. It is possible to win with the regular PS5 controller, but you have to be dirty and cut a few corners. Almost made me rage quit. Also, they've tweaked the physics in these races to give your car a truck load of understeer, which makes the races even more annoying.

 

[ScottPuss20]

Race 4 Expert. It's virtually impossible to win unless you have at least a decent wheel set up. It is possible to win with the regular PS5 controller, but you have to be dirty and cut a few corners. Almost made me rage quit. Also, they've tweaked the physics in these races to give your car a truck load of understeer, which makes the races even more annoying.

All you have to do is defend your position and not leave any gaps for Sophy to go through. You don't need to resort to dirty tactics to win.

 

[ShiftingGears]

The major difficulty with all machine learning algorithms is obtaining enough high-quality, appropriate data with which to train the model. To the point that synthetic data generation is now a major field of active research.


If only there were some way to test Sophy against thousands, if not millions of real human players of varying ability under controlled conditions.

Yes but I don't see what specifically is being trained to be improved here...the difficulty, or rather, the raw pace - needs to be adjustable, I don't really see this providing good data because people are just going to re-start as soon as they stop defending long enough to let Sophy through.

 

[evldave333]

I think Sophy will really come into its optimal zone when used for the license tests / circuit experience in the next GT title perhaps as a ghost to chase down or in a chase the rabbit then overtake challenge or vice versa staying in front being hunted. I miss the early part of beating the licenses, perhaps Sophy could be used for some further events in a similar vein to the Lewis Hamilton challenge in GTS ideally with generous credit pay-outs.

 

[Magog]

I think Sophy will really come into its optimal zone when used for the license tests / circuit experience in the next GT title perhaps as a ghost to chase down or in a chase the rabbit then overtake challenge or vice versa staying in front being hunted. I miss the early part of beating the licenses, perhaps Sophy could be used for some further events in a similar vein to the Lewis Hamilton challenge in GTS ideally with generous credit pay-outs.

It will definitely be integrated into the game long before GT8. I don't think PD will make another GT this Gen and there is still a long way to go for PS5.

 

[evldave333]

It will definitely be integrated into the game long before GT8. I don't think PD will make another GT this Gen and there is still a long way to go for PS5.

Indeed looking forward to seeing how it evolves in this game first. I wouldn't be surprised if they don't milk this cash cow again late in the PS5's cycle would be nice to see it optimised fully (particularly the tracks which are not as detailed as the vehicles).

 

[Griffith500]

Yes but I don't see what specifically is being trained to be improved here...the difficulty, or rather, the raw pace - needs to be adjustable, I don't really see this providing good data because people are just going to re-start as soon as they stop defending long enough to let Sophy through.

Literally all the input and output of every simulation tick can be uploaded to PDI's servers after every attempt. It won't be a lot of data individually.

Then you define and measure outcomes based on that data and feed it to the model as though you were running the tests locally. In fact, recording instances where players rage quit or reattempted several times might be a key indicator to pay attention to.

The model is already nothing like what it was when Sophy was exhibited publicly previously. It's far more courteous and it clearly runs faster than 10 Hz also. Probably because it's a new model that wasn't trained on video streams, like the original was, but internal physics data instead. Well actually it could be a different representation of those initial training sessions also, with outcomes tweaked.


As I said, obtaining the data is the bottle neck. They can retrain the model thousands of times over on the same data they already have (e.g. with different emphasis on different behaviours) with the amount of processing power Sony could make available at the drop of a hat.

If Sophy's outcomes scores and weightings aren't significantly different, perhaps because no new observations or realisations have been made yet, and if it's always put against the same kinds of situations, then it will still broadly behave the same.


This data from the limited time events here will likely give Sony and PD some new insights and point to how exactly they can retrain the model to include scalability (for the player) from the outset. Sophy isn't just learning to navigate a course any more, nor to match and exceed the pace of the very fastest sim racers. Sophy's new opponent is very different, so the results are likely to be different too.

There'll probably be a subsequent round of training data collection to test any insights gleaned from the current one.

 

[evldave333]

Literally all the input and output of every simulation tick can be uploaded to PDI's servers after every attempt. It won't be a lot of data individually.

Then you define and measure outcomes based on that data and feed it to the model as though you were running the tests locally. In fact, recording instances where players rage quit or reattempted several times might be a key indicator to pay attention to.

The model is already nothing like what it was when Sophy was exhibited publicly previously. It's far more courteous and it clearly runs faster than 10 Hz also. Probably because it's a new model that wasn't trained on video streams, like the original was, but internal physics data instead. Well actually it could be a different representation of those initial training sessions also, with outcomes tweaked.


As I said, obtaining the data is the bottle neck. They can retrain the model thousands of times over on the same data they already have (e.g. with different emphasis on different behaviours) with the amount of processing power Sony could make available at the drop of a hat.

If Sophy's outcomes scores and weightings aren't significantly different, perhaps because no new observations or realisations have been made yet, and if it's always put against the same kinds of situations, then it will still broadly behave the same.


This data from the limited time events here will likely give Sony and PD some new insights and point to how exactly they can retrain the model to include scalability (for the player) from the outset. Sophy isn't just learning to navigate a course any more, nor to match and exceed the pace of the very fastest sim racers. Sophy's new opponent is very different, so the results are likely to be different too.

There'll probably be a subsequent round of training data collection to test any insights gleaned from the current one.

Just engagement data alone will be of interest to them, myself once I saw how perfect the driving was I was put off playing against it for now at least. I bet there might be a steep drop off in engagement as the difficulty increases, conversely that could be the grail for the hardcore drivers the data will reveal all.

 

[Egghaus88]

So I found the races pretty easy, but the 1v1 battle is a real challenge. Anyone saying otherwise is either an alien, lying, or crashed Sophy intentionally. It is true that if you get the jump at the start, you just have to drive defensively, but the battles I had with this AI rank up there with the best battles I've had against human opponents in Sport mode. Personally I found Sophy would try and take me on the outside or non preferred line constantly. For that I say this is headed in the right direction, it for sure needs to fall somewhere between Expert and 1v1 level of difficulty.

 

[Voodoovaj]

Sophy might drive dirty sometimes, and you definitely have to drive dirty to beat it, at least the 1v1. But the GT AI has been so garbage for so long that this counts as a major step up.

When you've been drinking your own urine for years, a glass of slightly cloudy water seems magical.

This is a 100% fair point.

But, I still hold to the belief that if they spent the money to hire a competent game designer, they could have done more with the base AI that wouldn't have required all this effort. Not to mention that it would be through the entire game.

I believe that Sophy is more about removing workforce than improving the game, because after all this time, I feel they have no desire to actually improve the single player experience in terms of competition. Likely, too many people love the way it is.

 

[jonoholland29]

This is a 100% fair point.

But, I still hold to the belief that if they spent the money to hire a competent game designer, they could have done more with the base AI that wouldn't have required all this effort. Not to mention that it would be through the entire game.

I believe that Sophy is more about removing workforce than improving the game, because after all this time, I feel they have no desire to actually improve the single player experience in terms of competition. Likely, too many people love the way it is.

They could have. But Sophy is a Sony Project with a purely Sony budget. GT7 is a SIE/Polyphony Project (albeit under the global Sony Brand) so the money is coming from different places/budgets.

 

[Magog]

One thing I've noticed facing Sophy which also applies to the standard AI the slower the car you're facing relative to your own speed the worse the AI appears to be. The bigger the performance gap between vehicles the bigger the difference between braking points and how tightly you can turn. If you're approaching a braking zone 20% faster than your opponent they are going to react to it far differently and seemingly poorly. If you're within 5% that's when you can fight fairly. They really do need to let you approach the career races the same way you approach online races with qualifying being at the very least an option.

 

[regnar]

They really need to include in their testing of Sophy grid starts or bunched up double-file rolling starts, the current Sophy races are still using the same formula giving the bot a head start and the player having to chase them down. Other games pull grid starts off and this type of start will show off the AI weaknesses and where to improve. I hope if Sophy really gets good by the time it released, then reason to bring to qualifying.

 

[Magog]

They really need to include in their testing of Sophy grid starts or bunched up double-file rolling starts, the current Sophy races are still using the same formula giving the bot a head start and the player having to chase them down. Other games pull grid starts off and this type of start will show off the AI weaknesses and where to improve. I hope if Sophy really gets good by the time it released, then reason to bring to qualifying.

The one on one race is a grid start. You start right next to each other...

 

[regnar]

The one on one race is a grid start. You start right next to each other...

In a normal race you would have multiple cars to interact with, there would be traffic, there would be many more things thrown at the Sophy as opposed to a one on one race, it's bunched up cars at the start that Sophy would have to overcome.

 

[aguia]

Drivers like Sophie or worse, in the lobby they abound.

 

[Magog]

In a normal race you would have multiple cars to interact with, there would be traffic, there would be many more things thrown at the Sophy as opposed to a one on one race, it's bunched up cars at the start that Sophy would have to overcome.

There are grid starts with the standard AI in the game and they handle it just fine. In the chili races all the cars are bunched up nicely and behave themselves. With Sophy's improved speed and tactics I'm sure she can too. This was just a test.

 

[05XR8]

I haven’t checked telemetry of SOPHY. Normal AI have all aids working. Is SOPHY using all aids as well?

 

[Magog]

I haven’t checked telemetry of SOPHY. Normal AI have all aids working. Is SOPHY using all aids as well?

Just ABS

 

[happycorey]

If they could figure out how to make it adjustable the ability to add AI to online races/lobbies would be amazing

 

[Griffith500]

I haven’t checked telemetry of SOPHY. Normal AI have all aids working. Is SOPHY using all aids as well?

Sophy will already have far better control in all situations than any of the synthetic aids in the game would provide.

Mostly because it's very involved to change settings on the fly, the existing AI didn't do that (and it would be very difficult to program that behaviour effectively anyway) and Sophy will naturally adjust inputs to achieve a desired outcome without any special attention needed.

For the same reason, Sophy shouldn't even need ABS, but the way it works in a GT game it would confer an unfair advantage if it was disabled.

 

[Magog]

Sophy will already have far better control in all situations than any of the synthetic aids in the game would provide.

Mostly because it's very involved to change settings on the fly, the existing AI didn't do that (and it would be very difficult to program that behaviour effectively anyway) and Sophy will naturally adjust inputs to achieve a desired outcome without any special attention needed.

For the same reason, Sophy shouldn't even need ABS, but the way it works in a GT game it would confer an unfair advantage if it was disabled.

How would having ABS off give Sophy an advantage?

 

[05XR8]

How would having ABS off give Sophy an advantage?

From what I recall of fast players in GT Sport, ABS Off gave better adjustability, as ABS On increased brake distance. I guess it would work the same if SOPHY is optimising inputs.