- 176
- Greece
Gran Turismo Sophy: Sony AI x Polyphony Digital
- Thread starter Magog
- 1,719 comments
- 194,637 views
- 8,721
What are the potential differences in training the model against the game's video output versus using the more compact game state data instead, as is "traditional" for AI controlled opponents?
What about translating learned experience between the two types of model?
Is it possible to achieve the same level of performance using either approach?
I ask this because a video based approach only works for what the player is currently seeing - i.e. as a driving assist, or B spec Bob type take over of the player's car.
Each opponent AI would require its own video feed to work in the same way, which would rule out being able to run on one machine. That is unless they can teach a model to work with the internal game state and achieve a similar pace and competitive behaviour...
AI has by far been the worse area of the GT series in my opinion so any improvement is welcome.
I'll be honest the technical talk on how it works bores me to death so I don't look into it but it looks promising and I certainly look forward to the update when it arrives.
I'll be honest the technical talk on how it works bores me to death so I don't look into it but it looks promising and I certainly look forward to the update when it arrives.
- 538
- USA
- TheLadiesMan
Most relevant questions to us, the players, are:
1. Scalability
2. Time to implementation and how much implementation (as in, special one off mode or replacement for everything in the game)
Having super human AI doesn't really help the vast majority of players. It's incredibly demoralizing and demotivating to race (or play any game) against the top top tier of players. I had good success with my early days in Sport.. but whenever I was in a race with Fraga or Hellzfire etc.. Nope. Theres the top 5 or even 1% of players.. and then there are the top 0.1% lol
And implementation because if it is just a one-off special mode, then I don't care. I'm not interested in racing against lewis hamilton (except maybe once.. he's bad, because it's a video game) or driving on the moon or driving a redbull X prototype car etc. And if isn't in the career mode then kinda what's the point. The career mode races is where I'll spend a lot of my time. Online races can be fun but it's a different game at that point. I like playing the single player stuff for fun.
Edit: @Jordan - Actually, here's a question: Let's say it is eventually updated and added to the game more broadly. Does that mean if they make any core changes to the physics or tires or anything of that sort, does the AI have to relearn everything? And does the AI have to have a different model for every car, track?
1. Scalability
2. Time to implementation and how much implementation (as in, special one off mode or replacement for everything in the game)
Having super human AI doesn't really help the vast majority of players. It's incredibly demoralizing and demotivating to race (or play any game) against the top top tier of players. I had good success with my early days in Sport.. but whenever I was in a race with Fraga or Hellzfire etc.. Nope. Theres the top 5 or even 1% of players.. and then there are the top 0.1% lol
And implementation because if it is just a one-off special mode, then I don't care. I'm not interested in racing against lewis hamilton (except maybe once.. he's bad, because it's a video game) or driving on the moon or driving a redbull X prototype car etc. And if isn't in the career mode then kinda what's the point. The career mode races is where I'll spend a lot of my time. Online races can be fun but it's a different game at that point. I like playing the single player stuff for fun.
Edit: @Jordan - Actually, here's a question: Let's say it is eventually updated and added to the game more broadly. Does that mean if they make any core changes to the physics or tires or anything of that sort, does the AI have to relearn everything? And does the AI have to have a different model for every car, track?
Last edited:
- 436
- Leafs Nation
- deviln3
Oh man the AI battle with Yamanak is amazing: 33:25 in the racing video.
AI #1 has to also defend against AI #2 during the battle with Yamanak. The way AI moves under braking would cause most human players to lose control and spin, but AI manages to stay on track, and barely loses time.
CarloPlaya on youtube put together a great highlight list in the comments.
AI #1 has to also defend against AI #2 during the battle with Yamanak. The way AI moves under braking would cause most human players to lose control and spin, but AI manages to stay on track, and barely loses time.
CarloPlaya on youtube put together a great highlight list in the comments.
- 2,427
- The failed experiment entitled “California”.
Yea, especially since I’ve never seen “Sophie” spelled “Sophy” in my friggin’ life
Last edited:
- 1,713
- Sydney
- snakebacon97
Turn 10's gonna have to rename their "Unbeatable" haha!
It's fascinating, looking forward to it and seeing how it gets implemented with difficulty settings, also hopefully it'll adapt to the player's pace with options for how much you want to be pushed. We can have settings like:
It's fascinating, looking forward to it and seeing how it gets implemented with difficulty settings, also hopefully it'll adapt to the player's pace with options for how much you want to be pushed. We can have settings like:
- Slow as the old AI, but in a more natural way and they won't pit-manoeuvre you by accident
- Matching your pace but letting you win by a bit (I'm tired, inflate my ego)
- Matching your pace and going a bit faster (make me git gud)
- As fast as those DR B drivers I could never catch in Race C
- Sophy takes the gloves off (can average me beat Sophy in a much faster car? Imagine being chased by a relentless MX-5 around the Nürburgring, could you stay ahead in an M3? Could Horst von Saurma?)
- 236
- France
Well I do not know what kind of angle would be used to blame PDi or Sony for doing this research?
I mean PDi are aware of their own AI limitations as well as any other racing game Editors, no ?
So why complain about anything ?
Remember it was only with AlphaGo very recently (comparing to Chess with IBM DeepMind) that we were able to beat Humand at Go.
The Alpho Go project turned it's head to Starcraft II not that long ago after Go and was powerfull enough to beat GrandMaster.
Please consider that even if you consider the "feats" related in this article as "baby" steps be aware that the hardest part is now already completed!
The part that require to create a set of "learning" or "teaching" frameworkd for your IA to train by itself.
As a matter of "wrongly" emulated human behaviour as such as agressivity, input lag of commands, or reactivity well those parameters are the easy part to fine tune if you pardon my expression.
You don't grasp the huge amount of ressource in computing in generatings a model for trainning. Since once it done you just have to run your IA into to the game and it will train by itself. Without you having to give any more directions. Or more precisly once you are happy with what your IA as become by itself ( as "beating the human world toplayer" without any type of what you consider "technical advantage" ie faster reaction time to events in game.
Well then it is just a matter of choice to make the IA ""less perfect" on certain aspect (following racing line, never going outside, accepting loosing a position in order to avoid a devastating cash. Those are the setting we certainly expect to be used for a "regular" game such as GT7.
I am amazed of some results. It seems the AI was able to learn advanced racing behavior like escaping a block and turn it into an opportunity (classic F1 behavior). SO I am sure many people in the racing field will want to have some of what has accomplished here.
Another result is the description of new "trajectories" for taking some turns in order to gain avantage at the end of the next straight.
Dragon Tail being used as an exemple. This here is like AlphaGo versus Lee Seedol with the "god move" no could have predicted would result to gain an advanage of.
Peace everyone
I mean PDi are aware of their own AI limitations as well as any other racing game Editors, no ?
So why complain about anything ?
Remember it was only with AlphaGo very recently (comparing to Chess with IBM DeepMind) that we were able to beat Humand at Go.
The Alpho Go project turned it's head to Starcraft II not that long ago after Go and was powerfull enough to beat GrandMaster.
Please consider that even if you consider the "feats" related in this article as "baby" steps be aware that the hardest part is now already completed!
The part that require to create a set of "learning" or "teaching" frameworkd for your IA to train by itself.
As a matter of "wrongly" emulated human behaviour as such as agressivity, input lag of commands, or reactivity well those parameters are the easy part to fine tune if you pardon my expression.
You don't grasp the huge amount of ressource in computing in generatings a model for trainning. Since once it done you just have to run your IA into to the game and it will train by itself. Without you having to give any more directions. Or more precisly once you are happy with what your IA as become by itself ( as "beating the human world toplayer" without any type of what you consider "technical advantage" ie faster reaction time to events in game.
Well then it is just a matter of choice to make the IA ""less perfect" on certain aspect (following racing line, never going outside, accepting loosing a position in order to avoid a devastating cash. Those are the setting we certainly expect to be used for a "regular" game such as GT7.
I am amazed of some results. It seems the AI was able to learn advanced racing behavior like escaping a block and turn it into an opportunity (classic F1 behavior). SO I am sure many people in the racing field will want to have some of what has accomplished here.
Another result is the description of new "trajectories" for taking some turns in order to gain avantage at the end of the next straight.
Dragon Tail being used as an exemple. This here is like AlphaGo versus Lee Seedol with the "god move" no could have predicted would result to gain an advanage of.
Peace everyone
Thanks as always Jordan!! For Sophy, I was wondering how they decided on the state action models that they went with and what kinds of parameters they include. I also wonder the possibility of setting up the station action variables so that they more reflect human perception rather than the absolute knowledge an agent would have in the virtual world.
So, if you could take any of my numbered questions, that would be awesome! Thanks again Jordan!!
1) Is the goal to create the fastest AI, AI that mimics human driving, or both?
2) What state, action, or environmental variables were the most influential in the top reinforcement learning models?
3) Do you see a future where Gran Turismo is the training grounds for real-life AI racecars?
4) Are custom models for individual players or styles of driving something in the pipeline?
SestoScudo
(Banned)
- 3,000
- Australia
- Decepticon 47
I want to see AI do mistakes if not challenging me or defending their positions. I want some gritty racing.
Lets see what Sophy will behold.
Lets see what Sophy will behold.
SestoScudo
(Banned)
- 3,000
- Australia
- Decepticon 47
- 236
- France
LOL @ Sophy Orange doing some illegals moves
Imagine Toto Wolf:
"Box Box"
Maestro Lewis "Orange" Hamilton:
"NO."
Imagine Toto Wolf:
"Box Box"
Maestro Lewis "Orange" Hamilton:
"NO."
- 28,331
- Brooklyn, NY
- KR_Viper
- I Renown I
"...via a future update."
So, next year or never — understood.
Maybe this year but we're already getting the game this year, don't be greedy.
So, next year or never — understood.
Maybe this year but we're already getting the game this year, don't be greedy.
- 87,105
- Rule 12
- GTP_Famine
It's slightly more obvious than that.
The suffix "-sophy" - as in "philosophy" - literally means "knowledge". It comes from the exact same Latin (and Greek, oddly) root as the name Sophie/Sophia (generally held to mean "wise one").
What better name for an artificial intelligence than "knowledge"?
- 1,233
- VS & 74 (France)
- GT-Alex74
Kaz talked about how the PP system calculations were based on an AI running a fictional track - is Sophy being used for that already ? If not, would there be any plans to update the PP system to use the Sophy AI within GT7 ?
Also, in the races they ran against real players, I noticed stewards gave warnings to Sophy and that the AI cars got track limit penalties as well - is it something the AI is able to learn from and adapt to currently ?
- 51,151
- Australia
- SpacedustDaddy
Imagine if Sophy was injected into ASIMO(or whatever robot) and sitting behind the wheel at a world tour event? How awkward would that be?
Edit: @Jordan
This is my question to ask Kaz: Is there potential for us to ever see Sophy sitting next to human drivers, at a future World Tour event?
Edit: @Jordan
This is my question to ask Kaz: Is there potential for us to ever see Sophy sitting next to human drivers, at a future World Tour event?
Last edited:
- 4,053
- Wisconsin
- GTP_Eggstor
- 735
- Maryland, USA
Could this one day result in B-Spec mode and GT5's underrated 'remote races' returning? I personally think this was a super fun feature in its day, very fun to develop your own AI race team for your friends to try and beat with their own... all while getting rewarded for their results. It was a mini-game, within an already large Gran Turismo game.
Last edited:
- 766
- Las Vegas, NV
- K_E_87th
- B-Spec returns for GT7
- Sophy eventually added
- Level up B-Spec Bob enough to go beyond the level cap and unlock Sophy AI
- Race against your B-Spec Bob to fine tune their Sophy parameters
- Eventually, watch them destroy GT AI with cars 150 PP less than the competition
Not like any of the aforementioned will happen.
- 192
Similar threads
