Possibility of GT8 going Open World

[M Stark]

Idk what Kaz is planning for GT8, but for me I wouldn't want an open world game. I do want an "open world" track however like in Assetto Corsa. I think GT8 needs this feature.

 

[Mafio]

Its interesting to see the way some of you all think.

 

[AmyUD06]

Can I check what it is that you understand by this as, despite my own reinforcement, this seems to be a term you're relying on extensively.

My money is on this person thinking that each individual PS5 is running a live AI system for Sophy...which is laughable.

 

[newmedia_dev]

Can I check what it is that you understand by this as, despite my own reinforcement, this seems to be a term you're relying on extensively.

Oooo ooo ooo Mr Famine Sir, I know I know

Famine: “No Cooley, let him explain himself”
Glen: “but Mr Famine”
Famine: “I know I know, we all know but does he?”

 

[AutumnalGlow]

Well Famine, judging by the fact that you haven't actually contributed any information yourself, I can safely assume you know even less than they do. They're at least trying to explain how they think Sophy works, and you're just picking at a word and not addressing or contributing anything else.

 

[Famine]

Well Famine, judging by the fact that you haven't actually contributed any information yourself, I can safely assume you know even less than they do. They're at least trying to explain how they think Sophy works, and you're just picking at a word and not addressing or contributing anything else.

I'm trying to establish a baseline in order that we can converse on a common level - which, given the language differences, is kind of important - without wasting effort on misunderstandings but thanks for your assumptions and contribution 👍

 

[jrbabbitt]

People, I'm not the only one that sees that "open world" means "no racing" and if it is "no racing" then it's not Gran Turismo.

 

[eran0004]

Can I check what it is that you understand by this as, despite my own reinforcement, this seems to be a term you're relying on extensively.

This video explains it well:

 

[Famine]

This video explains it well:

I'm sure it does, but I'm not looking for what other people think; I'm trying to determine what it is you understand by the term so that we have a common ground on which to converse - as previously stated:

Can I check what it is that you understand by this as, despite my own reinforcement, this seems to be a term you're relying on extensively.

I'm trying to establish a baseline in order that we can converse on a common level - which, given the language differences, is kind of important - without wasting effort on misunderstandings

If you're just not interested in the discussion we can skip right past it I guess, but it's all a bit of a waste of everyone's time really.

 

[newmedia_dev]

This video explains it well:

The thing is, that Sophy is trained and then the behaviour packaged in to a self contained model on the console, it can’t learn from that point forward but it’s more than equipped to race and react to the pre described coding.

Don’t confuse how you make Sophy with how you apply it to very tightly controlled domains.

 

[Mafio]

People, I'm not the only one that sees that "open world" means "no racing" and if it is "no racing" then it's not Gran Turismo.

I think its kinda funny how people are only able to relate the idea to something they've already seen before as if its completely impossible for PD to incorporate an open world in any other way other than a way its already been done. "GTA"?!? Really????

 

[jrbabbitt]

If you have proper, legitimate racing there needs to be restrictions of track, meaning you cannot go off track, cutting corners, choosing your own route bypassing the set track limits. If you have an open world environment you cannot have that legitimate race, it's basically that simple.
If PD does implement an "open world" you will not have any legitimate point to point racing, there will not be any official menu books, nor missions because there will have to be some sort of physical route blocks to prevent cheating and makes it not open world..

 

[Mafio]

If you have proper, legitimate racing there needs to be restrictions of track, meaning you cannot go off track, cutting corners, choosing your own route bypassing the set track limits. If you have an open world environment you cannot have that legitimate race, it's basically that simple.

Believe it or not, in the universe in which we occupy, two truths can exists in a singular space at the same time. It can be an open world where one could drive a car to a track and race said car, or even another car in a regulated race. You guys cannot be this unimaginative and/or open minded. "Old" aint no excuse, either.

 

[jrbabbitt]

Believe it or not, in the universe in which we occupy, two truths can exists in a singular space at the same time. It can be an open world where one could drive a car to a track and race said car, or even another car in a regulated race. You guys cannot be this unimaginative and/or open minded. "Old" aint no excuse, either.

Thank you for pointing that out, I believed in the open world but when it came to racing it closed that openness.
I would love to have a point A to point B race but there would be many routes to take causing some restrictions. If you were even thinking to just cross-country it with an offload car it would cause more restrictions.
If there was a Cannonball type race there would be so many ways to cheat the system making it unfair to all except the cheaters.
Not fun, not racing.

 

[eran0004]

I'm sure it does, but I'm not looking for what other people think; I'm trying to determine what it is you understand by the term so that we have a common ground on which to converse - as previously stated:


If you're just not interested in the discussion we can skip right past it I guess, but it's all a bit of a waste of everyone's time really.

A neural network (or artificial neural network) is a computational model that consists of nodes and edges. There are input nodes, fed with data from sensors or with static values; output nodes, with control of various actions the AI can take, in the case of Sophy that includes braking and steering for example; and there are also one or more layers of hidden nodes.

In a feed-forward model, signals are passed from the input nodes to the first layer of hidden nodes. Each edge has a certain weight applied to it and if the sum of all incoming signals multiplied by their respective weights is greater than some threshold value, a signal from the hidden node is passed onto the next layer. Eventually the signals reach the output nodes and lead to some action taken by the agent.

The weights and thresholds are determined via the training process. The fact that you have a neural network does not mean it's learning from what it's doing. For it to learn anything you need to have a training algorithm, but that's not actually a part of the neural network.

The thing is, that Sophy is trained and then the behaviour packaged in to a self contained model on the console, it can’t learn from that point forward but it’s more than equipped to race and react to the pre described coding.

That's not quite how it works. You can't take a neural network and transform it into a set of instructions (if that's what you mean by "self contained model"). What we have on the PS5 is a neural network.

You are right that it's doesn't learn anything from racing against you. The weights and thresholds are what they are. There's no technical limitation that prohibits it from learning though, the neural network that we have is of the same type as the one they use when training it. What's missing is a training algorithm, i.e. a function to reward or penalise Sophy based on its performance and a function to change the weights and thresholds. We are unlikely to ever get such an update though, because Sophy needs training at a massive scale to make significant improvements, so it makes no sense to train it on an individual level.

Don’t confuse how you make Sophy with how you apply it to very tightly controlled domains.

Machine learning is the process by which the neural network is trained. The neural network itself is not machine learning. The fact that Sophy uses a neural network to determine its actions does not imply that it's learning from what it's doing.

 

[Mafio]

Thank you for pointing that out, I believed in the open world but when it came to racing it closed that openness.
I would love to have a point A to point B race but there would be many routes to take causing some restrictions. If you were even thinking to just cross-country it with an offload car it would cause more restrictions.
If there was a Cannonball type race there would be so many ways to cheat the system making it unfair to all except the cheaters.
Not fun, not racing.

I would leave that up to the players. Kind of like how there are lobbies dedicated to car shows, roleplaying ect. For instance imagine a freeway that goes around the entire menu/map. You and a friend can be meet at a designated point and race to another designated point. Maybe even implement waypoints. Yes, you're friend could cheat the whole way but, at that point, thats your fault for having a wack friend. jkjk

EDIT: If you REALLY want to go there, make the player pay for repairs that happen outside of the track. Even thought I hate the industry with a passion, PD could get an auto insurance partnership to sort of explain away why we don't have to pay for damages in actual races.

 

[Rockonen]

People, I'm not the only one that sees that "open world" means "no racing" and if it is "no racing" then it's not Gran Turismo.

Remember, GT is driving simulator, not racing, mkay?

 

[Famine]

A neural network (or artificial neural network) is a computational model that consists of nodes and edges. There are input nodes, fed with data from sensors or with static values; output nodes, with control of various actions the AI can take, in the case of Sophy that includes braking and steering for example; and there are also one or more layers of hidden nodes.

In a feed-forward model, signals are passed from the input nodes to the first layer of hidden nodes. Each edge has a certain weight applied to it and if the sum of all incoming signals multiplied by their respective weights is greater than some threshold value, a signal from the hidden node is passed onto the next layer. Eventually the signals reach the output nodes and lead to some action taken by the agent.

The weights and thresholds are determined via the training process. The fact that you have a neural network does not mean it's learning from what it's doing. For it to learn anything you need to have a training algorithm, but that's not actually a part of the neural network.

That's not quite how it works. You can't take a neural network and transform it into a set of instructions (if that's what you mean by "self contained model"). What we have on the PS5 is a neural network.

You are right that it's doesn't learn anything from racing against you. The weights and thresholds are what they are. There's no technical limitation that prohibits it from learning though, the neural network that we have is of the same type as the one they use when training it. What's missing is a training algorithm, i.e. a function to reward or penalise Sophy based on its performance and a function to change the weights and thresholds. We are unlikely to ever get such an update though, because Sophy needs training at a massive scale to make significant improvements, so it makes no sense to train it on an individual level.

Okay, I think I see where you're coming from now. We'll go with an "ehhhhhhhhhhhh fair enough, I guess".

I think its kinda funny how people are only able to relate the idea to something they've already seen before as if its completely impossible for PD to incorporate an open world in any other way other than a way its already been done. "GTA"?!? Really????

To be fair, there's only really a limited number of ways of doing it... ironically, I suppose

There's total sandbox, which seems closest to how Wreckreation is/will be doing it, and EventLab Island in FH5. All you get is the space, and you can add whatever you want within the defined region. Although Wreckreation seems to have at least a small, defined town as well.

Then there's a defined world sandbox, more or less what the rest of Forza Horizon 4/5 is: there's presets for certain locations like specific, non-destructible buildings or towns or forests, existing road networks, with geographical limitations (rivers, valleys, mountains), but freedom to add what you want on top.

Then we have a defined open-world, akin to earlier FH, GTA, TDUSC, and what looks like the approach for ACE. You can drive on a lot of it but everything there is fixed road network, certain inaccessible areas, and no customisation.

Finally there's the games with big maps but you can only drive on defined parts of them - though they're not really "open world" as you're hemmed into the roads (and/or dirt tracks) and everything has hard barriers and invisible walls up to the heavens to stop you from leaving. The Driver series springs to mind.


I think we'll have to wait and see how ACE handles its Eiger map - I'm expecting slightly dumbed-down physics and a load phase, so that you can't seamlessly drive around the map, to and onto the 'Ring, around the 'Ring and back out to the map without a transition - to really see how "open world" can be meshed with a racing game.

But even if it ends up being in GT eventually, I don't see GT8 going anywhere near it.

 

[newmedia_dev]

A neural network (or artificial neural network) is a computational model that consists of nodes and edges. There are input nodes, fed with data from sensors or with static values; output nodes, with control of various actions the AI can take, in the case of Sophy that includes braking and steering for example; and there are also one or more layers of hidden nodes.

In a feed-forward model, signals are passed from the input nodes to the first layer of hidden nodes. Each edge has a certain weight applied to it and if the sum of all incoming signals multiplied by their respective weights is greater than some threshold value, a signal from the hidden node is passed onto the next layer. Eventually the signals reach the output nodes and lead to some action taken by the agent.

The weights and thresholds are determined via the training process. The fact that you have a neural network does not mean it's learning from what it's doing. For it to learn anything you need to have a training algorithm, but that's not actually a part of the neural network.

That's not quite how it works. You can't take a neural network and transform it into a set of instructions (if that's what you mean by "self contained model"). What we have on the PS5 is a neural network.

You are right that it's doesn't learn anything from racing against you. The weights and thresholds are what they are. There's no technical limitation that prohibits it from learning though, the neural network that we have is of the same type as the one they use when training it. What's missing is a training algorithm, i.e. a function to reward or penalise Sophy based on its performance and a function to change the weights and thresholds. We are unlikely to ever get such an update though, because Sophy needs training at a massive scale to make significant improvements, so it makes no sense to train it on an individual level.

Machine learning is the process by which the neural network is trained. The neural network itself is not machine learning. The fact that Sophy uses a neural network to determine its actions does not imply that it's learning from what it's doing.

They use ML and then a little bit more ML but it has to be packaged.

The reason it’s PS5 only is not because it’s sky net, but because it has many more decision branches predetermined to pick from at each corner/input and the PS5 has some processing cycles to enable that.

If you understand the project and the tech, it’s a massive simulation machine that just runs multiple simulated laps and inputs.

The tracks and cars are fixed variables so the pathing and tasking is easy to describe to the agents.

This is not neural networks, it’s being able to simulate many agents to tune the optimum path, Sophy didn’t learn to drive, it already could it was used to find the optimum AI behaviour in a very very confined sandbox.

 

[jonoholland29]

Also… the budget for GTA6 is over $2,000,000,000 - that’s billion. More than it cost to build the world’s tallest building.

Have PD got the money to build a similar open world to their level of detail - even 25% the size of the GTA6 map?

 

[eran0004]

They use ML and then a little bit more ML but it has to be packaged.

That package is the neural network. There's no other way to pack it.

If you understand the project and the tech, it’s a massive simulation machine that just runs multiple simulated laps and inputs.

I don't know what you mean by massive simulation machine, but it sounds like you might be confusing the infrastructure used to train the AI at a massive scale with the neural network itself.

The tracks and cars are fixed variables so the pathing and tasking is easy to describe to the agents.

I don't understand what you mean by that.

This is not neural networks, it’s being able to simulate many agents to tune the optimum path, Sophy didn’t learn to drive, it already could it was used to find the optimum AI behaviour in a very very confined sandbox.

Sophy was trained from scratch, just like all neural networks have to be. You can't teach or instruct a neural network to do anything (at least not to any degree of success) without training it. I suppose that if your network is very small and the task you wish to perform is very simple you could make some educated guesses to what the weights and thresholds ought to be, but for those applications it would be easier, faster and better to just write an algorithm.