Gran Turismo Sophy: Sony AI x Polyphony Digital

[Saargasm]

I did all the races in tsukuba I think its the first one right?? I won all 3 against sophy I really loved it it felt more real and more like racing but I think the tyre handicap needs to go or they could just upgrade sophy's tires instead of nerfing mine the shift in grip from the beginner to expert really caught me off guard as I was expecting to have the same grip and sophys would get better..... that said I really love to see her reaction smileys as you pass her and she is really agressive too she rammed me more then once but I found she took the last turn too agressively and lost speed so I took it alot tighter and in a lower gear but hell yeah lets get sophy in all races just not the tire nerf on us

100% agree with the tire change from racing to sport to comfort. I never drive with comforts. As in never. Don't know how to drift effectively and the racing line naturally changes too.

Edit: I haven't had enough time to play R3 1v1 or any of the R4's, so I didn't know you have sport and racing respectively in the last race.

 

[jonoholland29]

I agree - we need a Sophy that can make human-like mistakes. The ability to be pressured into mistakes or over-trying would be good too. Fatigue and strategy mistakes for Endurances races too

 

[Voodoovaj]

I agree - we need a Sophy that can make human-like mistakes. The ability to be pressured into mistakes or over-trying would be good too. Fatigue and strategy mistakes for Endurances races too

All this AI talk makes me miss GT5 B-spec

 

[Sturdy]

As an utter guess as to how Sophy is working in the background, it is calculating the correct input in each moment to correctly follow the optimum line. There'll be some intelligence in how it 'understands' the handling model and thus how to get the car where it wants it etc.

I would also assume that if Sophy concludes that 74% throttle and 13* of steering lock are what's needed, then that is also what it commands from the controls and what it gets.

To make it more human, just apply a random fluff factor to those inputs to simulate different levels of competence. Beginner level, the inputs commanded are a variable +/- 10% of what it has concluded is the right amount. Intermediate, the inputs are subjected to a variable adjustment within +/- 5%. Pro level it's within +/- 2% etc. etc.

Currently it's basically always getting exactly the right input for what it knows is the fastest route. I'm sure a lot of human drivers also know what the right line / input SHOULD be, but it's actually delivering that perfectly which is part of what makes driving hard, and may be a way of allowing Sophy's "perfection" to be watered down in a more organic way, that would mean it varied naturally without having to be artificially hindered or with scripted 'mistakes' baked in.

It would also be then pretty easy to add an 'under pressure' modifier that increased the range of variation, or an 'angry', 'focused', 'calm' etc. etc.

I think leave Sophy in a competitive vehicle, leave it with the perfect line targeted, and just have a variation factor on how it implements the controls to achieve that perfect line.

 

[jonoholland29]

As an utter guess as to how Sophy is working in the background, it is calculating the correct input in each moment to correctly follow the optimum line. There'll be some intelligence in how it 'understands' the handling model and thus how to get the car where it wants it etc.

I would also assume that if Sophy concludes that 74% throttle and 13* of steering lock are what's needed, then that is also what it commands from the controls and what it gets.

To make it more human, just apply a random fluff factor to those inputs to simulate different levels of competence. Beginner level, the inputs commanded are a variable +/- 10% of what it has concluded is the right amount. Intermediate, the inputs are subjected to a variable adjustment within +/- 5%. Pro level it's within +/- 2% etc. etc.

Currently it's basically always getting exactly the right input for what it knows is the fastest route. I'm sure a lot of human drivers also know what the right line / input SHOULD be, but it's actually delivering that perfectly which is part of what makes driving hard, and may be a way of allowing Sophy's "perfection" to be watered down in a more organic way, that would mean it varied naturally without having to be artificially hindered or with scripted 'mistakes' baked in.

It would also be then pretty easy to add an 'under pressure' modifier that increased the range of variation, or an 'angry', 'focused', 'calm' etc. etc.

I think leave Sophy in a competitive vehicle, leave it with the perfect line targeted, and just have a variation factor on how it implements the controls to achieve that perfect line.

Yeah exactly - it’s Knowledge v Application.

I know what makes a car go fast, it’s the doing bit that takes practice and skill.

 

[Gingerale]

As an utter guess as to how Sophy is working in the background, it is calculating the correct input in each moment to correctly follow the optimum line. There'll be some intelligence in how it 'understands' the handling model and thus how to get the car where it wants it etc.

I would also assume that if Sophy concludes that 74% throttle and 13* of steering lock are what's needed, then that is also what it commands from the controls and what it gets.

To make it more human, just apply a random fluff factor to those inputs to simulate different levels of competence. Beginner level, the inputs commanded are a variable +/- 10% of what it has concluded is the right amount. Intermediate, the inputs are subjected to a variable adjustment within +/- 5%. Pro level it's within +/- 2% etc. etc.

Currently it's basically always getting exactly the right input for what it knows is the fastest route. I'm sure a lot of human drivers also know what the right line / input SHOULD be, but it's actually delivering that perfectly which is part of what makes driving hard, and may be a way of allowing Sophy's "perfection" to be watered down in a more organic way, that would mean it varied naturally without having to be artificially hindered or with scripted 'mistakes' baked in.

It would also be then pretty easy to add an 'under pressure' modifier that increased the range of variation, or an 'angry', 'focused', 'calm' etc. etc.

I think leave Sophy in a competitive vehicle, leave it with the perfect line targeted, and just have a variation factor on how it implements the controls to achieve that perfect line.

This sounds easier than it is. Basically Sophy has 3 inputs: steer, accelate and brake. It checks every millisecond Wat the best input would be.

If you would say, pick the best input in 95% of all times, it wouldn't be able to drive a straight line.

 

[SPL170db]

Just finished the last 1v1 race at Suzuka. It got a little dicey and Sophy got a little spicy trying to pass me going into T1 and we both ran it wide, but it seems the AI has a more difficult time recovering from foul ups. Oh well, I'll take it.

I honestly thought it would be harder to 100% all of the these races but no doubt this AI is miles above and beyond the standard AI of the game.

 

[Roggers]

What do you think? Did Sophy Bleue try and force me of the track?



Have to say racing the Sophy AI is pretty good and fun! Only criticism I have off it is that the Violet car make some interesting and unrealistic moves I think? Was racing the Honda Integra races, and I left a small gap that the Violet car decided it was going to make a move for. I saw in my rear view mirror the violet car speed up mid corner, make a dive and get past, but go really deep into the corner, then drift the corner and keep the place. To me that just seemed a little too unrealistic for me.
The Bleue and the Rouge (to some degree) are really competitive and I have a few good races with these two cars. Green is way to easy, and Violet very competitive but just the issue above I think at times.

 

[dabz343]

Wow, overall...very interesting critiques of Sophy so far. I'm with the majority who see it as an improvement, but with certain rough spots that need attention. I would like PD (or Sony or both) should start to compile player feedback so that Sophy becomes not an advisory to overcome, but an enjoyable partner -- and a key selling point for GT8.

 

[Magog]

What do you think? Did Sophy Bleue try and force me of the track?



Have to say racing the Sophy AI is pretty good and fun! Only criticism I have off it is that the Violet car make some interesting and unrealistic moves I think? Was racing the Honda Integra races, and I left a small gap that the Violet car decided it was going to make a move for. I saw in my rear view mirror the violet car speed up mid corner, make a dive and get past, but go really deep into the corner, then drift the corner and keep the place. To me that just seemed a little too unrealistic for me.
The Bleue and the Rouge (to some degree) are really competitive and I have a few good races with these two cars. Green is way to easy, and Violet very competitive but just the issue above I think at times.

What's unrealistic about it? Violette has supreme car control. Have you watched the older videos of the Pros racing Sophy? They learn and emulate her lines and to an extent her driving style. Fight fire with fire. Not easy but not impossible.

 

[Imari]

It's not easy, but its not as hard as you make it. Right now Sophy makes 100% of the right decisions, so it generates the fastest output. PD needs to lower that so it makes more mistakes, the thing is the mistakes need to consistent and realistic. That is quite hard to do, but not impossible.

I see a lot of users seem to think Sophy is some sort of sentient AI (which cant be slowed down), which I think is quite a wrong impression. Right now I seem to think we can already pressure Sophy into making mistakes by following it really closeby.

Have you been reading the papers and articles from Sony.AI on how they made Sophy? It's not easy. It is hard. It's hard specifically because of the way Sophy was made. It's designed to optimise itself to be as fast as possible, and it's not clear that you can get something that drives well at speeds significantly slower than that without making radical alterations to the method.

Quotes from a GTPlanet article, dated Feb 2022.

As Kazunori Yamauchi explained to us, the machine learning process provides Sophy with more rules of behavior than human programmers could possibly devise, but that strategy comes with its own drawbacks as well.

“The AI up until now was rules-based, so it basically runs as an if-then program,” Yamauchi-san explains. “But no matter how many of these rules are added, it can’t handle conditions and environments other than those defined. Sophy, on the other hand, generates a massive amount of implicit rules that humans can’t handle, within its network layer. As a result it is able to adapt to various conditions and environments. But because these rules are implicit, it means that it’s not possible to make it learn ‘a specific behavior’ that is simple for a rule-based AI.”

It's not sentient and I haven't seen anyone other than you even suggest that it is. But it is a monolithic program - you can't just chop bits off it to cripple it. It's a black box that has been trained with the goal of being very fast. What do you do to that to slow it down?

As Sophy was debuted as a super-human driver that was capable of defeating the best Gran Turismo players in the world, questions and concerns immediately arose about its ability to adapt to less competitive human drivers.

According to Peter Wurman, Sophy can adjust by literally driving like a newer driver instead of just artificially slowing down. “This is also part of our future work,” the Sony AI America Director explained. “Our goal is to create an agent that, when in a ‘slowed’ mode, is driving like a less experienced driver, rather than being handicapped in some way, like arbitrarily speeding it up or slowing it down in violation of the physics.”

That was a year ago. They're clearly aware of the problem and had a goal that they wanted to meet, an AI that behaves like a less experienced driver. That it does not appear in this first round of Sophy releases implies that at best it's not at a playable stage yet, and at worst it still doesn't exist - for reasons that are fairly obvious when you think about how Sophy is built and trained.

So if you as a software developer want to achieve this, what do you do? If you have suggestions then I'm sure the developers would welcome them. Because at the moment you're saying it's not that hard, but you seem to be pulling that out of thin air.

It is hard, and that's why they haven't done it yet. It may seem unintuitive that it's more difficult to make this AI drive slowly than it is to make it drive like an alien, but here we are. They will almost certainly get there in the end, but it's an open question as to whether the Sophy framework they have now will be sufficient to develop a suitable slower AI or whether they'll have to add in a whole lot more new technology as well.

I've seen it make plenty of "mistakes" but almost always in an effort to be a clean driver. If you mess around in front Sophy she will do anything to not hit you if possible. When following her she will react to cover you and that can be exploited but that requires you to be faster than her in the previous corner which is no mean feat.

This. I don't think you can force Sophy to make mistakes, I think that the reward for not hitting the player is so high that when you're close it looks like it's making "mistakes" because it starts behaving erratically. It's trying to defend as well as not hit the player, and I suspect that they've tuned it so that it will ultimately prefer to crash itself rather than hit the player if possible.

It's interesting that this looks like a human player under stress, because ultimately it's a similar thing going on. The human is trying to manage a bunch of conflicting priorities and as a result does some (or all) of them poorly.

I'm talking about small errors when there is no external data in place: i.e. making errors without the player intervening in it's course.

If the player is close, it's a data point that Sophy is including in it's choices. You don't have to be on a collision course to be influencing how Sophy drives.

As an utter guess as to how Sophy is working in the background, it is calculating the correct input in each moment to correctly follow the optimum line. There'll be some intelligence in how it 'understands' the handling model and thus how to get the car where it wants it etc.

No. It's a black box. It's a neural network trained to evaluate a number of parameters and spit out car control outputs. There's no "understanding" of the handling model, or cars, or tracks, or anything a human would relate to.

The reality is that the whole point of AI like this is that nobody can point to a part of the program and say exactly why it chooses to do certain things, it's a giant equation that happens to give very good answers almost all the time.

I think leave Sophy in a competitive vehicle, leave it with the perfect line targeted, and just have a variation factor on how it implements the controls to achieve that perfect line.

Nice idea, but if you've ever played with a dodgy controller or a loose wheel, then you know that random variation on your inputs both slows you down and turns you into a completely unpredictable menace. It would probably work to slow Sophy down, but it would undo a lot of what makes her racecraft great. Instead of being able to go safely side by side with Sophy, low level players would be getting rammed off the road.

You might say that's accurate to the low level racing experience with real humans, but that's exactly why a slow but clean AI is desirable. Slow drivers don't like being rammed off by the AI any more than fast drivers do.

This sounds easier than it is. Basically Sophy has 3 inputs: steer, accelate and brake. It checks every millisecond Wat the best input would be.

No. I'm not sure we know what speed the GT7 Sophy instances are running at, but the ones referenced in Nature ran at 10Hz.

Checking every millisecond would be 1000Hz. There's no way they turned up the polling rate by a factor of 100 just to put it on a PS5. That would be incredibly inefficient.

 

[dabz343]

Sometimes it's best to remain humbled.

 

[Gingerale]

Have you been reading the papers and articles from Sony.AI on how they made Sophy? It's not easy. It is hard. It's hard specifically because of the way Sophy was made. It's designed to optimise itself to be as fast as possible, and it's not clear that you can get something that drives well at speeds significantly slower than that without making radical alterations to the method.

That was a year ago. They're clearly aware of the problem and had a goal that they wanted to meet, an AI that behaves like a less experienced driver. That it does not appear in this first round of Sophy releases implies that at best it's not at a playable stage yet, and at worst it still doesn't exist - for reasons that are fairly obvious when you think about how Sophy is built and trained.

Yes I've been following Sophy for a while. And it IS possible, because a year ago they showed the video of the famous superhuman AI which basically had the perfect race. Now, we have AI which we actually can beat 1 vs 1.
So it is possible, you're contradicting yourself.

So if you as a software developer want to achieve this, what do you do? If you have suggestions then I'm sure the developers would welcome them. Because at the moment you're saying it's not that hard, but you seem to be pulling that out of thin air.

Im not saying its pulled from thin air. Im not saying its not hard. Stop misquoting me, I literally said it was not easy.
You have to look at moment where the AI differs from from superhumans. Yo

It is hard, and that's why they haven't done it yet. It may seem unintuitive that it's more difficult to make this AI drive slowly than it is to make it drive like an alien, but here we are. They will almost certainly get there in the end, but it's an open question as to whether the Sophy framework they have now will be sufficient to develop a suitable slower AI or whether they'll have to add in a whole lot more new technology as well.

I still don't understand that you keep saying 'they haven't done it yet'. I just beat Sophy this morning?

No. I'm not sure we know what speed the GT7 Sophy instances are running at, but the ones referenced in Nature ran at 10Hz.
View attachment 1234043
Checking every millisecond would be 1000Hz. There's no way they turned up the polling rate by a factor of 100 just to put it on a PS5. That would be incredibly inefficient.

That's because Sophy is not run at a PS5 of course...

It is important to note this type of computing power is only needed to “create” Sophy, not run it. The machine learning process eventually results in “models” which can then be executed on more modest hardware.

“The learning of Sophy is parallel processed using compute resources in the cloud, but if you are just executing an already learned network, a local PS5 is more than adequate,” Kazunori Yamauchi explained. “The asymmetry of this computing power is a general characteristic of neural networks.”

That would make everyone's PS5 explode lol.

 

[Meythia]

That's because Sophy is not run at a PS5 of course...

Of course it is.
Start your race, disconnect from the internet and see Sophy is not waiting for any magical connection from outside your PS5 to recieve information of how to drive.
The AI on track is run on PS5 consoles.

 

[Imari]

Yes I've been following Sophy for a while. And it IS possible, because a year ago they showed the video of the famous superhuman AI which basically had the perfect race. Now, we have AI which we actually can beat 1 vs 1.
So it is possible, you're contradicting yourself.

I replied to your post about making Sophy drive slower, and in the greater context of an entire discussion about making a slower Sophy. Those sentences were contracted references to that subject.

And so the meaning was:
It's not easy (to build a slow version of Sophy).
It is hard (to build a slow version of Sophy).

I mistakenly assumed you could read the context. My bad, I'll try not to rely on your ability to read context in the future.

Im not saying its pulled from thin air. Im not saying its not hard. Stop misquoting me, I literally said it was not easy.
You have to look at moment where the AI differs from from superhumans. Yo

You're saying it's not hard to build a slow Sophy, and providing no evidence. You're up against several papers describing a system specifically designed to go fast, one of the scientists a year ago saying that this is their future work, and it both not being available in-game and there have been (to my knowledge) no further publications on advancements to Sophy that would indicate that they've made a slower Sophy.

Because I'm pretty sure that would be publishable if they had managed to take Sophy, an already pretty impressive piece of technology that did the rounds of the news in the appropriate circles, and generalised it make it tunable for speed.

I still don't understand that you keep saying 'they haven't done it yet'. I just beat Sophy this morning?

Bro, if you can't understand context when I've just spent the previous two paragraphs talking about building a slower Sophy and literally quoting one of the researchers talking about it, it's going to be impossible to have a conversation. I cannot spoon feed you everything. You're going to have to remember more than two sentences at a time and figure out how they relate to each other.

That's because Sophy is not run at a PS5 of course...

Do you have evidence for that to be the case? Because my understanding was that we don't know. It's an always online game, it's not trivial to just run the game offline and check.

It's possible that they got it running on PS5, and that would be one of the reasons why it's only on PS5.
If they were running it from the cloud then it should work just as well on PS4, although they might still choose to only release on PS5 for the sake of pushing hardware sales.

Someone will eventually figure it out, but until then I'm happy to assume that it's running on the PS5. Which is really neither here nor there, because when the Nature article was running it was also running on cloud hardware. At 10Hz.

If they're streaming the Sophy data to the PS5 from the cloud, a very high tick rate would make even less sense. Even if you're close to whatever server they have, there's going to be dozens of milliseconds of latency in the connection. And Polyphony are not renowned for their excellent online connections. Calculating and controlling every millisecond would be absolutely insane to do over a connection like that, and while I sometimes give Polyphony a hard time even I don't think they're that stupid.

If they're in the cloud, they're using the minimum hardware and bandwidth possible. Which means not doing 1000 calculations a second. Because that's stupid.

Where did you get that number, anyway? I assume that you made it up, but you seem inclined to actually defend it which means you must have some sort of source. Right?

That would make everyone's PS5 explode lol.

Dude, get a reading comprehension course or something. The training of the AI was done on very powerful computers, the specifications of which I have actually posted earlier in the thread. However, when Sophy is just driving and not learning, it runs on significantly less powerful hardware, the specs of which are also in this thread.

The Sophy we have is not learning. It's just driving. We don't know how much they've optimised the "driving only" agent in the last year or so, but it's absolutely conceivable that they got it to run on PS5.

This is why I asked if you read the papers and articles. You have, but the information seems to have gone in one ear and out the other. You might need to back up and educate yourself on some of the basic concepts of machine learning before you're going to be able to contribute meaningfully to any conversation here.

 

[Meythia]

It's an always online game, it's not trivial to just run the game offline and check.

That is basically what I did to verify my previous comment.
1. Start GT7
2. Start any "race together" and wait to be put on track
3. Press PS Button and navigate to settings -> network
4. Disconnect from the internet
5. Go back to GT7 (the game will tell you results can not be updated because the game is not connected)
6. Finish your race and see Sophy is not disconneted and still doing the task at hand.

Afterwards you will be notified the game has been put into offline mode and offer only the basic stuff like Music Rally and such.

 

[Monkey Man]

I've yet to see anyone beat Sophy 1-on-1 by catching up to and overtaking her.

Every single case I've seen so far has been where the driver had to take her before T1 and then hold her off.

Once she's behind, it doesn't take any super-advanced skill level to keep her there, I'm guessing because she's trying her best to race cleanly.

IOW, anyone in this thread who's said she's not-too hard to beat is dreamin' IMHO.

 

[Gingerale]

I replied to your post about making Sophy drive slower, and in the greater context of an entire discussion about making a slower Sophy. Those sentences were contracted references to that subject.

And so the meaning was:
It's not easy (to build a slow version of Sophy).
It is hard (to build a slow version of Sophy).

I mistakenly assumed you could read the context. My bad, I'll try not to rely on your ability to read context in the future.

You're saying it's not hard to build a slow Sophy, and providing no evidence. You're up against several papers describing a system specifically designed to go fast, one of the scientists a year ago saying that this is their future work, and it both not being available in-game and there have been (to my knowledge) no further publications on advancements to Sophy that would indicate that they've made a slower Sophy.

Because I'm pretty sure that would be publishable if they had managed to take Sophy, an already pretty impressive piece of technology that did the rounds of the news in the appropriate circles, and generalised it make it tunable for speed.

Bro, if you can't understand context when I've just spent the previous two paragraphs talking about building a slower Sophy and literally quoting one of the researchers talking about it, it's going to be impossible to have a conversation. I cannot spoon feed you everything. You're going to have to remember more than two sentences at a time and figure out how they relate to each other.

Do you have evidence for that to be the case? Because my understanding was that we don't know. It's an always online game, it's not trivial to just run the game offline and check.

It's possible that they got it running on PS5, and that would be one of the reasons why it's only on PS5.
If they were running it from the cloud then it should work just as well on PS4, although they might still choose to only release on PS5 for the sake of pushing hardware sales.

Someone will eventually figure it out, but until then I'm happy to assume that it's running on the PS5. Which is really neither here nor there, because when the Nature article was running it was also running on cloud hardware. At 10Hz.

If they're streaming the Sophy data to the PS5 from the cloud, a very high tick rate would make even less sense. Even if you're close to whatever server they have, there's going to be dozens of milliseconds of latency in the connection. And Polyphony are not renowned for their excellent online connections. Calculating and controlling every millisecond would be absolutely insane to do over a connection like that, and while I sometimes give Polyphony a hard time even I don't think they're that stupid.

If they're in the cloud, they're using the minimum hardware and bandwidth possible. Which means not doing 1000 calculations a second. Because that's stupid.

Where did you get that number, anyway? I assume that you made it up, but you seem inclined to actually defend it which means you must have some sort of source. Right?

Dude, get a reading comprehension course or something. The training of the AI was done on very powerful computers, the specifications of which I have actually posted earlier in the thread. However, when Sophy is just driving and not learning, it runs on significantly less powerful hardware, the specs of which are also in this thread.

The Sophy we have is not learning. It's just driving. We don't know how much they've optimised the "driving only" agent in the last year or so, but it's absolutely conceivable that they got it to run on PS5.

This is why I asked if you read the papers and articles. You have, but the information seems to have gone in one ear and out the other. You might need to back up and educate yourself on some of the basic concepts of machine learning before you're going to be able to contribute meaningfully to any conversation here.

Why are you discussing so overly aggressive? I'm just here to talk about Sophy, you're in it to win a discussion and own other posters it seems. I'm not an English native so the stuff about getting reading course is a bit overdone imo.

Again, I never said it was easy to make Sophy slower. But it is not impossible, and PD has shown in the last year that they are progressing.


For the 'run op PS5', I used the exact wording from GTPlanet, I mean that Sophy is not created on PS5 but a general model. So it doesn't have to do calculations every 100 milliseconds. This is from the Nature article.

But I'm not in for this discussion, just chill a bit, stop being so overly aggressive as you seem to think you're in a contest or something. Relax man.

 

[Sturdy]

This sounds easier than it is. Basically Sophy has 3 inputs: steer, accelate and brake. It checks every millisecond Wat the best input would be.

If you would say, pick the best input in 95% of all times, it wouldn't be able to drive a straight line.

I don't mean for it to pick the best input 95% of the time, and 5% of the time it's totally random.

I'm going to tie the response in with part of Imari's quote below.

Nice idea, but if you've ever played with a dodgy controller or a loose wheel, then you know that random variation on your inputs both slows you down and turns you into a completely unpredictable menace. It would probably work to slow Sophy down, but it would undo a lot of what makes her racecraft great. Instead of being able to go safely side by side with Sophy, low level players would be getting rammed off the road.

You might say that's accurate to the low level racing experience with real humans, but that's exactly why a slow but clean AI is desirable. Slow drivers don't like being rammed off by the AI any more than fast drivers do.

I don't mean that every conclusion the AI has come up with only has a 90% chance of being used, and 10% of the time a random number is thrown in.

For reference, Imari I do understand how a neural network works, I was using the term 'understanding of the physics' purely for brevity of discussion, I appreciate any perceived 'understanding' is an emergent behaviour of the neural network, like an ant colony 'deciding' on the location for a new nest.

What I mean is, seemingly at a 10Hz interval, Sophy takes the inputs from the game data, car location, speed, angle, tyre condition, weight, opponent proximity / position etc. etc. etc., these factors go through the black box, and out of that black box comes a conclusion in the simplistic form of 3 commanded inputs to the game, a commanded steering angle, a commanded throttle % and a commanded brake %.

Obviously I am simplifying this.

My suggestion is merely at say 'Beginner' level, a variation is applied (of random magnitude each calculation) across a range of + / - 10%. (10% may be a bit much but it makes the maths easy to demonstrate what I mean...)

For utter simplicity, the 'randomiser' is a number between 0 and 20. It is randomised as part of each 10Hz calculation, and that is used to pick the % variation (within the range +/- 10% for Beginner).

So Sophy's calculations have concluded that a steering angle of 30* is optimal, with 0% brake and 80% throttle.

The randomizer has randomly landed on 10. For our range of +/- 10%, 10 corresponds to the midpoint, so 0 variation.

Sophy inputs exactly as requested, 30* of steering, 0% brake and 80% throttle.

The next 10Hz interval, the calculation happens again. It's a constant radius corner, and Sophy's black box again concludes 30* steering, 0% brake and 80% throttle.

The randomizer has landed on 13. So we will apply a +3% variation to what Sophy has commanded.

Thus due to not being perfect, Sophy ends up inputting 30* + 3% steering, so 30.9*, 0 + 3% of braking, so 0% still (3% of 0 is still 0...) and 80 + 3% throttle, so 82.4%.

This doesn't mean Sophy randomly spears off into the trees, it means for 1 tenth of a second, she has input 1* more steering than is absolutely optimal, and a tiny bit more throttle than is optimal. For 1 tenth of a second.

The next 10Hz interval, the calculation will start again, taking all of this new position etc. into account.

That is in no way the same as a dodgy controller, or being unable to drive straight or randomly swerving all over the place. It being a percentage of the optimal input means in a straight line / low inputs it'll have almost no impact at all, and because of how the throttle and brakes work the larger sways at higher numbers won't really matter too much, as Sophy uses ABS anyway, and tiny percentage adjustments to throttle opening at high openings don't make much difference to the power output.



Hopefully that explains it better. My suggestion is to leave the black box calculating the optimal inputs every tenth of a second, and just add some fuzziness to the implementation of that optimal input to mimic the slight fallibility of human attempts at hitting exactly the right amount of brake input every time etc. etc.

 

[RedRockValley]

I will say, it is pretty ironic that people were accusing OP of "huffing copium" in the opening pages of this thread...

 

[Imari]

Hopefully that explains it better. My suggestion is to leave the black box calculating the optimal inputs every tenth of a second, and just add some fuzziness to the implementation of that optimal input to mimic the slight fallibility of human attempts at hitting exactly the right amount of brake input every time etc. etc.

I get it, I just think that you'll run into problems with racecraft and close racing. To make a Sophy slow down to the speed of the average GT player feels like it should take a fair amount of modification to the controls.

But the racecraft problem isn't insoluble. Maybe if you make it so that the fuzziness isn't entirely random when the player is within a certain distance - ie. steering fuzziness always pushes Sophy away from the player, accel/brake fuzziness always defaults to what would move Sophy away. That's probably not that hard to do when they're already tracking the relevant variables to feed into the base Sophy algorithm.

But honestly, if I was the developer I would try this method. It's relatively simple to set up and it will very quickly be obvious if it's a viable path or not. It may be that even if there's downsides they are an acceptable trade off for a Sophy that is slower.

A Slophy, if you will.

Why are you discussing so overly aggressive? I'm just here to talk about Sophy, you're in it to win a discussion and own other posters it seems. I'm not an English native so the stuff about getting reading course is a bit overdone imo.

You accused me of contradicting myself when it was your reading error that caused the problem. And then misread again. And again.

If your English isn't up to the task, then don't come in accusing other people of being contradictory. I assumed that you knew what you were talking about and meant what you said. Your English appears to be at least as good as some English speakers on this forum and you claimed to have read some pretty technical and complex articles and papers on the subject. So I assumed that your skill in English was at a fairly high level. I don't think that was unreasonable.

If your English skill is low then there's no shame in that - I have a second language but I certainly couldn't discuss anything this complex in it. But if you're not confident then you need to tell people so that they can cut you some slack. I'm not a mind reader.

Again, I never said it was easy to make Sophy slower. But it is not impossible, and PD has shown in the last year that they are progressing.

They have not shown progress on slow Sophy. We saw fast Sophy last year, and we now have fast Sophy in a bunch of different cars on PS5.

That's it. No progress on slow Sophy.

So it doesn't have to do calculations every 100 milliseconds. This is from the Nature article.

Correct. 100 milliseconds, or 10 times a second, or 10 Hz. That is what I posted.

Not every millisecond. That's a big difference, and it matters if we're talking about how Sophy is controlled.

This sounds easier than it is. Basically Sophy has 3 inputs: steer, accelate and brake. It checks every millisecond Wat the best input would be.

Before you take it personally again, it's about correcting incorrect information. You wrote something that was incorrect, even if it was a mistake. I pointed that out. You can take that as an attack if you want, or you can just say "sorry, I made a mistake".

If we're going to have a discussion about Sophy, it's important to start with accurate information about the basic facts of the system.

But I'm not in for this discussion, just chill a bit, stop being so overly aggressive as you seem to think you're in a contest or something. Relax man.

I'll make you a deal. You start using accurate facts and stop accusing me of things that aren't true, and I'll try not to come across as aggressive.

I will say, it is pretty ironic that people were accusing OP of "huffing copium" in the opening pages of this thread...

The OP in question.

It's happening, boys. The AI in GT is about to surpass our wildest dreams!

I mean, has it surpassed our wildest dreams? Probably not for most people, unless your dreams were really shooting low.

It's an interesting and impressive technology, but it's still pretty limited in how it affects the gameplay. It has potential, but I think "copium" was a fair assessment at the time given it was before release and the limited information on Sophy. It's easy to get a little too excited with things like this. Understandable with all the hype around, but still copium.

And even with an early version out now, it's still not a magic bullet for all the problems that we've had with the GT AI for the last... forever. Nor is it likely to be any time soon. It's not even "game changing" as the Sony/PD tweet would claim, it's a few good challenge events. They're some of the better events in the game, but this is honestly the sort of events that many people had hoped would be in GT7 from the very start.

That is basically what I did to verify my previous comment.
1. Start GT7
2. Start any "race together" and wait to be put on track
3. Press PS Button and navigate to settings -> network
4. Disconnect from the internet
5. Go back to GT7 (the game will tell you results can not be updated because the game is not connected)
6. Finish your race and see Sophy is not disconneted and still doing the task at hand.

Afterwards you will be notified the game has been put into offline mode and offer only the basic stuff like Music Rally and such.

Nice. That sounds like it's confirmed to be running locally then.

 

[RedRockValley]

The OP in question.

I mean, has it surpassed our wildest dreams? Probably not for most people, unless your dreams were really shooting low.

It's an interesting and impressive technology, but it's still pretty limited in how it affects the gameplay. It has potential, but I think "copium" was a fair assessment at the time given it was before release and the limited information on Sophy. It's easy to get a little too excited with things like this. Understandable with all the hype around, but still copium.

Oh, I was talking about how people were saying it wasn't going to be in the game at all and that it was just a simple experiment/marketing showcase, the complete other end of someone overhyping this basically.


Obviously it hasn't completely changed the racing game landscape, but it's definitely super cool to see GT get a much nicer AI system.

 

[Gingerale]

They have not shown progress on slow Sophy. We saw fast Sophy last year, and we now have fast Sophy in a bunch of different cars on PS5.

That's it. No progress on slow Sophy.

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

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. I'm sorry for stating every millisecond instead of every 100 milliseconds though. That was wrong of me.

 

[Sturdy]

I get it, I just think that you'll run into problems with racecraft and close racing. To make a Sophy slow down to the speed of the average GT player feels like it should take a fair amount of modification to the controls.

But the racecraft problem isn't insoluble. Maybe if you make it so that the fuzziness isn't entirely random when the player is within a certain distance - ie. steering fuzziness always pushes Sophy away from the player, accel/brake fuzziness always defaults to what would move Sophy away. That's probably not that hard to do when they're already tracking the relevant variables to feed into the base Sophy algorithm.

But honestly, if I was the developer I would try this method. It's relatively simple to set up and it will very quickly be obvious if it's a viable path or not. It may be that even if there's downsides they are an acceptable trade off for a Sophy that is slower.

A Slophy, if you will.

As a developer (but not of this sort of software) I definitely would be inclined to try something of that nature. I think by keeping the 'fuzz' proportional to the actual commanded input, the existing Sophy AI behaviour would be capable of self-correcting fairly well and it would also give you some nice fudge numbers to work with.

The randomizer for example, instead of being totally random could be based around a normal distribution, you could then use a standard deviation number as a modifier as well, to control the spread of the inaccuracies within the +/- range to perhaps allow a wider range without causing utterly random driving, but to let occasional bigger mistakes creep in even for the higher difficulty levels.

It's quite interesting to me the way it could be set-up to produce some more representatively human fallibility into Sophy's implementation of the perfect line it has derived.

 

[Foofer37]

I've yet to see anyone beat Sophy 1-on-1 by catching up to and overtaking her.

Every single case I've seen so far has been where the driver had to take her before T1 and then hold her off.

Once she's behind, it doesn't take any super-advanced skill level to keep her there, I'm guessing because she's trying her best to race cleanly.

IOW, anyone in this thread who's said she's not-too hard to beat is dreamin' IMHO.

Agree. So here's where we're at:
Kaz creates the 'Best' 'Fastest' AI in the world that you can bump in T1, get in front and easily keep it behind you the entire race. Or, you can enter a beginner and Intermediate race where you can watch them do superhuman, unrealistic things, but once you get past them they drive slow. My 2 cents is that Sophy sucks. What is this? Who cares about all this 'DARPA' crap? It's a racing game. If it's not a fun challenge and part of a fun experience then just sell it to the military. wow. Kaz has truly lost it.

 

[Voodoovaj]

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.

 

[05XR8]

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.

It’s all good because it’s learning. This is what’s presented to us and as mentioned, all data will be analysed for improvements.
Who knows, maybe we’ll have Turd SOPHY and My Buddy SOPHY mixed in future versions.

 

[Magog]

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.

That hasn't been my experience or that of many people who have raced against it. I found it far cleaner than any of the people I have raced against in online.

 

[Voodoovaj]

It’s all good because it’s learning. This is what’s presented to us and as mentioned, all data will be analysed for improvements.
Who knows, maybe we’ll have Turd SOPHY and My Buddy SOPHY mixed in future versions.

But people learning is not ok? Sure

That hasn't been my experience or that of many people who have raced against it. I found it far cleaner than any of the people I have raced against in online.

From the videos posted, I would disagree with your findings.



I'll take real people. Then I know they are doing their best, even if their best is terrible.

 

[RedRockValley]

I'll take real people. Then I know they are doing their best, even if their best is terrible.

From what I’ve seen, I quite like the more aggressive “rubbin’s racin” approach Sophy takes. Apart from a few moments I wouldn’t call her a straight up dirty driver.

Of course not everyone is gonna want to race like that, so it would be cool if when Sophy is fully integrated into GT7 she comes with an aggression slider, like Assetto Corsa AI had.