Gran Turismo 6 AI discussion

[research]

@MX5Racer69 I've raced against Big Foot in Monster Truck Madness back in the late-90's and he was plenty quick. No brake-checking, whatsoever.

I mentioned the A.I. awareness being decent, so the other 'nodes' activity is no reason for them to have such terrible braking points/throttle out. -These are the easy parts to program. This crap happens whether we're in a pack, or all alone in GT5. Just watch the replays from a lonely A.I. backmarker POV. He will be brake-checking all by his darn self.

 

[Brend]

Alright all of you guys stating that the hardware is holding PD back as far as A.I., I have to interject. A.I. "awareness" absolutely takes up substantial processing power and RAM- no doubt. But in GT5 and even moreso GT6 the A.I. seem to be pretty aware. Aware enough to have some really good racing with side-by-side battles, overtakes, and reactive dodging moves by them if you swoop in. Heck, they even seem to be VERY adept at countersteering and catching slides. That alone must use some pretty serious resources.

Now let's examine the race-driving technique. Which scenario would take more processing power/RAM:

1. Full throttle, approach turn, release throttle/early brakecheck, let off brakes, throttle, brake again, turn wheel, coast, unwind wheel, late throttle out, one more brakecheck just for fun, full-throttle.

2. Full throttle, approach turn, release throttle/strong braking at correct point, ease off brake into turn/turn wheel, ease onto throttle/unwind wheel, full throttle out.

?

#1.

 

[MX5Racer69]

@MX5Racer69 I've raced against Big Foot in Monster Truck Madness back in the late-90's and he was plenty quick. No brake-checking, whatsoever.

I mentioned the A.I. awareness being decent, so the other 'nodes' activity is no reason for them to have such terrible braking points/throttle out. -These are the easy parts to program. This crap happens whether we're in a pack, or all alone in GT5. Just watch the replays from a lonely A.I. backmarker POV. He will be brake-checking all by his darn self.

Just one AI? Yeah, that's easy to do. But I do agree with you somewhat. The cars that are driving the simple pattern should be much faster. 100% agree.

 

[Sick Lenny]

. Research said
Now let's examine the race-driving technique. Which scenario would take more processing power/RAM:

1. Full throttle, approach turn, release throttle/early brakecheck, let off brakes, throttle, brake again, turn wheel, coast, unwind wheel, late throttle out, one more brakecheck just for fun, full-throttle.

2. Full throttle, approach turn, release throttle/strong braking at correct point, ease off brake into turn/turn wheel, ease onto throttle/unwind wheel, full throttle out.

?

Technically neither would take more in-game resources. #1 would take more time to program at first (assuming this canned AI scenario), and also makes less sense.
I am not sure why PD made their AI behave like this, but I suspect that like many other weak aspects of the game (sound), it is something they are aiming for the stars for and it's taking forever because they have a small team and/or this gen hardware can't fit all they aspire to. So for now we get this "placeholder" AI.

 

[MrWednesday]

Now let's examine the race-driving technique. Which scenario would take more processing power/RAM:

1. Full throttle, approach turn, release throttle/early brakecheck, let off brakes, throttle, brake again, turn wheel, coast, unwind wheel, late throttle out, one more brakecheck just for fun, full-throttle.

2. Full throttle, approach turn, release throttle/strong braking at correct point, ease off brake into turn/turn wheel, ease onto throttle/unwind wheel, full throttle out.

?

Is this a trick question?

#2.

It requires either more processing power to do a high quality dynamic calculation of the correct racing line, or more resources to handle a "remembered" version of the correct racing line. There are fewer corrections required after turn-in, but a lot more involved in getting it right through the apex.

#1 will end up needing more hand-holding through the turn, but that arises from devoting less to getting the right line at the outset.

 

[aronh17]

GTR2's AI is miles faster, without cheating, running in miles lower specs, taking real lines....in a 2006 game made by a small company.

Is it too much to ask to have a difficulty slider that actually works?

Does GTR2 simulate weather/time, stars, detailed aerodynamics, wind, humidity, air pressure, temperature, detailed suspension and tire physics? Did it also boast Adaptive Tessellation and Morphological Anti-Aliasing? Seeing as its a 2006 PC sim, it didn't have much in the physics department, and it barely makes it into the HD games era. I'm sure GT6 also has a higher draw distance, and many other new features that the CPU has to account for. Also, once you pull all of that CPU heavy stuff in, which is possible thanks to the CELL, its probably a lot harder to get decent AI out of the CELL considering the architecture it provides and the other things using the CPU up.

Side note: PlayStation 4 will offer much more capabilities because its much more simple to code on. I'm not trying to excuse PD from everything, as they should focus on AI more, but they have had a struggle with Sony bouncing with hardware specs from PS2 to PS3. It is safe to say that better AI will come. It should be possible from a title update, but my best guess is on PS4. Also the AI in GT6, like in that video of the Elise 111R; the AI seemed to take a different line than the others. It also pulled out of its own line to let him pass, and it dodged him when he wrecked. It looks pretty good so far, but I haven't played it yet so I'm not sure.

 

[Paolosev91]

I do not agree...
GTR2 was a GREAT sim for 2006... I think that you never played it...
After GT4 I was introduced to the fantastic world of PC sims... they're 100x times harder to master and they're really challenging (also frustrating sometimes).
These are the kind of games that I wanted to take a copy and send do Polyphony Digital, sometimes I thought that they had never seen it.
Now sim racers are many more and better... but for that time GTR2 was a great, challenging game

the "licenses" of GTR2 were 10x harder than GT... I remember the one in which you had to brake at Magny Cours wet track with a Lotus... Braking was 10x harder than any GT ever made... the car was so unstable and you had to be so fxxxxxg careful with the throttle otherwise you would spin in a second.
it was also totally focused on race cars which is something that I like.

I really would like to have a GTR2 with gt6 graphics on ps3...
I miss the qualifying, the cold tires in the first lap, and many other racing things... I would pay it hundreds of $$$.
However, GT6 will be less simulation but is still really promising to me...
but I'm also still hoping that Project Cars will make it on the ps3.

 

[Imari]

Does GTR2 simulate weather/time, stars, detailed aerodynamics, wind, humidity, air pressure, temperature, detailed suspension and tire physics? Did it also boast Adaptive Tessellation and Morphological Anti-Aliasing? Seeing as its a 2006 PC sim, it didn't have much in the physics department, and it barely makes it into the HD games era. I'm sure GT6 also has a higher draw distance, and many other new features that the CPU has to account for. Also, once you pull all of that CPU heavy stuff in, which is possible thanks to the CELL, its probably a lot harder to get decent AI out of the CELL considering the architecture it provides and the other things using the CPU up.

It did simulate time and weather, since you ask.
It didn't have a Copernicus simulation, but it had stars. I doubt anyone would be able to pick the difference.
It had more detailed suspension and tyre physics than GT5 did.

I don't know if it simulated detailed aero, wind, humidity, air pressure, or temperature. We're yet to see what this actually entails in GT6 as well.

It used a different method of reducing model detail at distance, so no Adaptive Tessellation.
It had as much AA as you chose to crank up on your computer. Probably more than GT6 will have, if you've got a modern computer.

You obviously haven't played it if you think it didn't have much in the physics department. It's a capable sim, and remains one to this day.
It's a PC game, so it's HD if your computer is capable of HD.
You're sure GT6 has a higher draw distance, but that is again dependent on your computer.


GT has never had good AI. Whether that's because they choose to spend all their CPU time on graphical tweaks, or whether it's because they're simply incapable we'll probably never know. But it can be done, and as games like GTR2 show has been done by much smaller teams, on smaller budgets. PD choose not to.

 

[ECGadget]

Being a part-time AI scripter, I can tell you know, the amount of calculations needed fro AI is incredible. I know that a lot of people complain, but if you all were to become programmers and modellers, you'd be singing a different tune!

Now, I completely agree that these things need to be improved and PD have had a while to upgrade them, favouring other items instead of it. But if you think about it, it does make sense.

People have been asking for sounds and PD are somewhat improving them, but that is a department I don't know well, so I cannot comment.

People have been talking about fixing up standards and PD have done some model tweaking and more importantly texture cleanups (alpha channels, higher res etc)

PD have also gone and made the game much more awesome for those of us who love to hotlap cars and compare them, which I personally do a lot! Having tyre physics, aerodynamics based of simple CFD calculations (most likely off the Large Eddy Simulations, which take the least amount of computing power), etc etc.

As an Aerospace Engineer I can tell you that working out Aerodynamic properties of cars is a pretty tough gig, even for a computer, so hats off to PD if they've succeeded in that, while still making the game look graphically very nice (barring some standard cars of course)

Now coming back to AI:
I have a feeling the AI has been dumbed down somewhat deliberately to appeal to the masses. While I would not mind having to work very very hard to get to first place (literally sit there for an hour before a race tweaking the car to the conditions so it can compete with the rest of the field), I doubt many of us would do that, especially if the race is only going to last 10 minutes. While that may be a mistake, to change it so that all the cars would be more bunched up, aggressive, accurate etc is a mean feat. What you're proposing here is something like this (albeit this is very simplified)


While (1)

If corner:

cornermode = 1​

else:

cornermode = 0​

EndIf

​

If (cornermode == 0)

Throttle = accelerate
Steering = straighten
If (RPM == revlimit*0.95)

If (gear == maxgear)

Throttle = hold​

Else

Throttle = release
Gear = up
Throttle = accelerate​

EndIf​

EndIf​

EndIf

If (cornermode == 1)

Throttle = release
Brake = apply
Steering = turn​

EndIf

If (corner == approaching)

cornermode = 1​

EndIf​

EndWhile


This is the absolute basic of how an AI could work, and even then I'll highlight the problems with this script (if it worked)

- No idea about how sharp the turn is
- No knowledge of car (is the car automatic, what is the clutch like? Does it need to release accelerator to change gear? Which car has better brakes? Which has better grip? What about the width of the car?)
- No track awareness (What if a car spins in front? What is a racing line? What does reduced grip mean? What if I spin? Which way is forwards? Do I need to brake earlier? Can I accelerate? What happens if the car slides? What is an overtake? How do I overtake? What can I see? Where is the track edge?)
- Steering is digital (Only three states: full left, full right and center)
- Speed through corners (What's the ideal speed? Am I in the right gear? Do I accelerate? Can I brake a little in the turn? What if somene cuts me up? What is the apex? How close do I get to the wall? Do I have a runoff or is it a wall? Can I afford to push a little harder? Or my tyres screaming under the strain?)
- Condition awareness (Do I slow when its raining? What about snow? Are my tyres right for the conditions? What is my tyre condition? Night racing, do I flash my lights? What can I see? Where are the other cars? How far am I to P1? What position should I aim for? What is the state of my fuel? What is the state of my competitors? Am I faster in one part of the lap than others?)

These are just a few of many problems with that tiny script. Now just to give you an idea, if I was to fix just the very first point, it would take me at least all day to figure out how I was to code it, and I would probably end up tripling the size of the current PsuedoCode.
If I was to fix all of these, we're looking at months and months of hard work with lots of testing, reverting to old code, and thousands of lines of code. Would my AI be as good as GT6? Maybe, but would it then be compatible? Upgradable? Would it be efficient and functional or would it eat up all the system resources? There is a lot more to coding AI than just saying

BrakingPoint = Correct
Speed = Fast
Aggression = Acceptable

And unfortunately many of us on GTPlanet and in fact everywhere do not understand this. What PD have done for the game is great. It may not be what everyone is expecting, but try to see if from the perspective of one of us programmers, and you'll realise that it is difficult indeed. That being said, I am not saying that we have to be happy with what we have been given. Constructive criticism is what makes a good game great, but if and ONLY IF you know the challenges involved when you make those critical remarks. I hope that this post gives a small taster of what challeges PD face with their small team and makes us a little more appreciative of what they have acheived, while supporting them to do even more in the future, learn from mistakes they make along the way and give the fans everything they want!

👍

 

[Paolosev91]

If AI is so bad I will do like GT5: interrupt the GT mode very early (really, the worst gt mode ever since the first GT, challenging only if you choose to use low performance tyres, which is stupid) and try the online mode.
If the online mode sucks (like in GT5 if you don't organize your private races somehow), then I'll probably leave it on the bookshelf...

or play time trials! (GT best mode ever XD)

 

[jjaisli]

Does GTR2 simulate weather/time

weather no, time yes

stars

LOL, no.

detailed aerodynamics

If by detailed, you mean GT5/6 level, yes.

wind, humidity, air pressure, temperature

I *think* so but I can't remember for certain--it's been a while

detailed suspension and tire physics?

Yes.

Did it also boast Adaptive Tessellation and Morphological Anti-Aliasing?

No.

However, it's a game that was designed a few years before the PS3 was introduced and designed to run on hardware which was much less powerful and had less memory than the PS3. And, I should add, that with GTR2, you could have double the number of AI on the track at one time.

I think people make too many excuses for the AI in GT which simply don't hold any water. IMO, the AI in GT5 & probably GT6 is so slow because Sony and PD know that the vast majority of people who will play this game are casual gamers. You know, like the people who play it at game shows and take video of themselves driving off the track that make us cringe. The kind of player who will probably never get rid of the racing line and additional aids. Like it or not, unlike hard-core PC sims that cater to hard-core sim racers, GT HAS to be playable and enjoyed by a large group of people with a large range of skill sets, and that includes 10 year old kids who play it like they would Mario kart. I would suggest, it's for this reason that the AI is adjustable in Arcade mode but not in career mode. And why it's so painfully slow.

I also believe this is the reason for the ridiculous conga line style starts, which don't exist in ANY racing series in the world that I'm aware of. In a real race, in the real world, you could have a great battle with one single other car for 10 laps. And that battle could be for 10th place but it could be real edge of your seat style, exciting racing. Look at the seasonal events--pass 11 other cars in 2~5 laps. It's artificial. And it's to make it exciting. Would casual players be interested in a racing game where you fought tooth and nail for 45 minutes against 1-2 cars? I don't think so.

EDIT: Oops, I missed other replies on the next page who've already said it in more detail.

 

[Imari]

Great and informative post, but there is one bit:

I have a feeling the AI has been dumbed down somewhat deliberately to appeal to the masses. While I would not mind having to work very very hard to get to first place (literally sit there for an hour before a race tweaking the car to the conditions so it can compete with the rest of the field), I doubt many of us would do that, especially if the race is only going to last 10 minutes. While that may be a mistake, to change it so that all the cars would be more bunched up, aggressive, accurate etc is a mean feat.

This was solved decades ago with difficulty levels. Easy difficulty for those who don't want to work very hard, Nightmare for the masochistic, and a few for everyone in between.

You know this, I know this, we all know this. It's difficult to get a decently functioning AI system, I'll grant you. Very difficult. But if you've gotten that far, creating something to modulate the overall speed can't be a serious obstacle.

I don't consider it acceptable that we all have to put up with the "dropped-on-it's-head-as-a-child" version of the AI, when a difficulty selector would be relatively simple to add. It's not like difficulty settings aren't a standard feature in almost every other game ever made.

 

[ECGadget]

Great and informative post, but there is one bit:



This was solved decades ago with difficulty levels. Easy difficulty for those who don't want to work very hard, Nightmare for the masochistic, and a few for everyone in between.

You know this, I know this, we all know this. It's difficult to get a decently functioning AI system, I'll grant you. Very difficult. But if you've gotten that far, creating something to modulate the overall speed can't be a serious obstacle.

I don't consider it acceptable that we all have to put up with the "dropped-on-it's-head-as-a-child" version of the AI, when a difficulty selector would be relatively simple to add. It's not like difficulty settings aren't a standard feature in almost every other game ever made.

With this, I do completely agree with you! A difficulty level would help a lot in these circumstances. But I think that with most games, difficulty translates into Faster AI and more bashing or just faster cars. I have a feeling PD want to go a different way with it, make the AI smarter instead of just faster, but it's not working out right for them =(

 

[Cosbuster]

After playing for a couple of hour and finishing the first few races, it's better than I expected. I used an abarth 500 and the guy with the mito would give me a run for my money in pretty much every race. When I would take a corner a little badly he would not hesitate to jump on the inside and his exit speed was very decent which really showed on longer straights.

It wasn't too hard to win or anything but it did make me have to look behind my back much more often than GT5 ever did and I experienced a nervousness I haven't felt in a while.

P.S. I did have to downgrade my tyres to CM since the 500 came stock with comfort softs which are just far too grippy for such a car.

 

[WonkyFonk]

'Improving' the AI for GT6 from GT5 is NOT a difficult task in my opinion.

What was fundamentally wrong with the algorithm (in my opinion) was the 'opponent' braking while you were near or alongside. I dont believe that was the case when GT5 was released; it seemed to suddenly appear after a patch if I remember correctly. I remember it because I was doing some chase the rabbit challenge on the Nurburgring which I couldn't beat. And suddenly after the update I blitzed it and notcied the 'slowing down' behaviour in the replay.

It's almost like the caution parameter had been set to max or something, it seemed like a bug. Addressing that one issue alone would be a drastic improvement.

I understand there are other issues with the AI not being aware of you presence etc, so this would not be the only improvement possible.

 

[trustjab]

If more people put correct tire levels on these cars the racing would be closer. Sports hard and sports mediums are about what any high level sportscar has. Sport Softs are close to cut slicks, a lot like track day tires imo. Comfort soft would be 'summer tires'. If srf isn't forced and no traction control it should be pretty fun. The AI needs more aggression in overtaking. It seems they like to sit behind other slower cars and wait for a straight. Hopefully theres a fix to that.

 

[Imari]

If more people put correct tire levels on these cars the racing would be closer. Sports hard and sports mediums are about what any high level sportscar has. Sport Softs are close to cut slicks, a lot like track day tires imo. Comfort soft would be 'summer tires'. If srf isn't forced and no traction control it should be pretty fun. The AI needs more aggression in overtaking. It seems they like to sit behind other slower cars and wait for a straight. Hopefully theres a fix to that.

I agree.

Unfortunately, they've now got tyre restrictions on the events. The AI should be able to give you a decent race on the listed tyres for the race, even if those tyres are far stickier than they should be.

The other option would be to provide some incentive for players to downgrade their tyres, but racing AI who are on tyres a couple of grades stickier than yours presents some of the same problems as racing AI who have a few dozen more PP than you. You may both be close at the finish, but it won't be particularly good racing. One of you will get mugged on the straights, and the other will be swamped in the corners.

The very best racing moments for me have been when I'm against someone who is really evenly matched in every way, and struggling to find any opening to take advantage of. Just switching back and forth is more annoying than anything, and tends to encourage one player to get ahead and then block the hell out of the other.


I do think this is an open field for a game like GT to take advantage of. I'd love to see them bend the licenses more towards racecraft situations rather than "this is how you drive around a corner". Making races focused more towards head to head racing would be interesting as well, and I think it's a natural extension of the kind of extra-curricular activities that Kaz tends to participate in. I imagine he'd have a lot to contribute.

Good AI is unfortunately critical to make a game where the racing is strongly biased towards racecraft work. We're a long way away.

 

[aronh17]

It did simulate time and weather, since you ask.
It didn't have a Copernicus simulation, but it had stars. I doubt anyone would be able to pick the difference.
It had more detailed suspension and tyre physics than GT5 did.

I don't know if it simulated detailed aero, wind, humidity, air pressure, or temperature. We're yet to see what this actually entails in GT6 as well.

It used a different method of reducing model detail at distance, so no Adaptive Tessellation.
It had as much AA as you chose to crank up on your computer. Probably more than GT6 will have, if you've got a modern computer.

You obviously haven't played it if you think it didn't have much in the physics department. It's a capable sim, and remains one to this day.
It's a PC game, so it's HD if your computer is capable of HD.
You're sure GT6 has a higher draw distance, but that is again dependent on your computer.

I disagree its pushing as much physics as GT6, mainly because the recommended CPU for it is a Pentium 4, which is nothing to the CELL processor. Its as simple as GTR2 having a very limited number of tracks and cars, which makes it easier to specifically code AI for each individual vehicle and track. If you did that for GT6, you'd be looking at 1200 cars and 100 tracks with tire variables. That's a lot of work, not to mention it also has to be tested. Instead, PD opt for a more basic code that works; and can be used across the span of vehicles. It could be better if they spent more time there but they don't.

The direct remedy is online. Which PD have stated they're making a bigger push for in GT6. Its not impossible to find good lobbies and have good races. There are a lot of them, and there are clubs people make to make sure everything is clean. I can almost bet online gaming was brought up because AI wasn't good enough for gaming in general. And by that I mean the original invention of online gaming.

 

[Mad Matt]

Coming back to GTR2 for a moment:

I played it quite a lot but the graphics weren't that special compared to GT and while there were many things it did well I never rated it above GT as an overall experience. In fact it felt quite a lot like Ferrari Challenge to drive and I mean that as a compliment.

The biggest thing for me was that it only covered GT racing cars (well I played GT legends as well which widened things a bit) which drive quite differently to road cars, or even fast road cars.

So, back to AI: The AI seemed somewhat better than GT4/5 but it's difficult to say how much difference there was between my PS3 (bought a little later) and my PC at the time. My PC certainly had a lot more memory and a dedicated graphics card with a nice chunk of memory too. The PS3 architecture is quite different to x86 but my best guess would be that my high end PC was more powerful.

Anyway, all of that's irrelevant in my opinion :-)

The question here is do we want better AI? If the answer is yes then the next question is why aren't PD doing it. Some possible answers are:

1. They don't see it as important. (we may disagree)
2. They are prioritising other things above it.
3. There isn't enough resource in the PS3 to be able to improve it without sacrificing something else (which amounts to point 2).
4. They lack the ability.

We can let them know on this forum (it is said that they read it ) that it is important and that we would like them to increase the priority, even if it means sacrificing something else.

We can speculate that there is enough resource or that things could be done differently so that there could be resource freed up but I don't think that will achieve much. PD are unlikely to read this thread and say "They're right we've got a load of unused CPU cycles we could use for better AI". Even our speculation isn't that interesting (in my opinion) as we've got so little insight into how PD are using the PS3 and what kind of load it's under.

So, there you go, that's my tuppence worth on the subject.......

 

[GENERALSAVANT]

While (1)

If corner:

cornermode = 1​

else:

cornermode = 0​

EndIf

​

If (cornermode == 0)

Throttle = accelerate
Steering = straighten
If (RPM == revlimit*0.95)

If (gear == maxgear)

Throttle = hold​

Else

Throttle = release
Gear = up
Throttle = accelerate​

EndIf​

EndIf​

EndIf

If (cornermode == 1)

Throttle = release
Brake = apply
Steering = turn​

EndIf

If (corner == approaching)

cornermode = 1​

EndIf​

EndWhile

You see the matrix..

 

[avens]

More than "better" AI what I want is the option of choosing considerably faster AI, that use the same tires, cars and don't cheat. I understand it is a console game therefore it aims at a difference audience, but I want an option nothing more.

The best would be that it detects the players' skill automatically plus having the option, like many current games do, and having challenging license tests like GTR2 (gold is not guaranteed to everyone).

And yes, GT5's was very slow no matter the settings.

 

[research]

The direct remedy is online.

I couldn't disagree more strongly.

I can almost bet online gaming was brought up because AI wasn't good enough for gaming in general. ...

There's that misconception again! Have you read the thread? There are a myriad of racing games which had competent, fast, and lively A.I. -Dirt2 by codies is my favorite, but others here have listed MANY titles with great A.I. from the last 15 years.

 

[ECGadget]

You see the matrix..

Haha. I should point out that that was a very quick, off the top of my head, script in no real programming language, but the point I was trying to make is that something thrown together in 20 minutes, or even 20 days is nothing like what a full AI script can and should be. Maybe I Should show more of my matrix seeing and write a better example script xD

I think it is important that gamers understand the coding skills and complexity that goes into making a game. It would make us more patient and appreciative people!

 

[GranTurismoFive]

If more people put correct tire levels on these cars the racing would be closer. Sports hard and sports mediums are about what any high level sportscar has. Sport Softs are close to cut slicks, a lot like track day tires imo. Comfort soft would be 'summer tires'. If srf isn't forced and no traction control it should be pretty fun. The AI needs more aggression in overtaking. It seems they like to sit behind other slower cars and wait for a straight. Hopefully theres a fix to that.

How do you explain AI taking the corners so slowly?

This has nothing imo to do with tyre choice. The game should offer a challenge regardless.

I mean why should the player have to compensate for the AI's poor driving capabilities?

 

[aronh17]

I couldn't disagree more strongly.

There's that misconception again! Have you read the thread? There are a myriad of racing games which had competent, fast, and lively A.I. -Dirt2 by codies is my favorite, but others here have listed MANY titles with great A.I. from the last 15 years.

The AI in the games you play are more artificial than you think. They follow lines of code rather than make their own decisions like a person. So, you're driving against something that knows what its doing and doesn't make a realistic error. Even GRID and DIRT 2 had those AI that would make "an error" when it was really just a script with a chance of happening. I'm not implying we kill all AI, but online has made gaming way more diverse, because actual human error steps in. I'm not against playing AI at all, but online is much better in my opinion, which I know not everybody can access. I've met a decent set of people that have offered awesome times with. Discussing cars and things, having competitions, etc. Something that AI does not provide. I've played rFactor and a few other PC sims, but it still doesn't compete with an actual person. There are things that make AI better though, because they aren't meant to step out of their boundaries. If you find a decent set of guys to race with AI doesn't need to exist.

EDIT: I never meant my argument to step out of the fact that I don't really like AI, I was just trying to explain the issue with AI and what PD faces. On my side of that up there, I'm basically saying I see no reward from racing something that doesn't make any errors at all, like a human being would and does.

 

[Imari]

The AI in the games you play are more artificial than you think. They follow lines of code rather than make their own decisions like a person. So, you're driving against something that knows what its doing and doesn't make a realistic error. Even GRID and DIRT 2 had those AI that would make "an error" when it was really just a script with a chance of happening. I'm not implying we kill all AI, but online has made gaming way more diverse, because actual human error steps in. I'm not against playing AI at all, but online is much better in my opinion, which I know not everybody can access. I've met a decent set of people that have offered awesome times with. Discussing cars and things, having competitions, etc. Something that AI does not provide. I've played rFactor and a few other PC sims, but it still doesn't compete with an actual person. There are things that make AI better though, because they aren't meant to step out of their boundaries. If you find a decent set of guys to race with AI doesn't need to exist.

EDIT: I never meant my argument to step out of the fact that I don't really like AI, I was just trying to explain the issue with AI and what PD faces. On my side of that up there, I'm basically saying I see no reward from racing something that doesn't make any errors at all, like a human being would and does.

Human errors arise from a couple of factors:

Poor physical control - you're unable to make precisely the movement you want to make.
Poor decision making - you misread the situation, or simply make an incorrect decision in the heat of the moment.

Perhaps there's more, but I'd say most errors are broadly one of those two.

I see no reason why both of those can't be simulated by a machine.

For the first, just a like a human the computer knows what input it wants to do. The input then has an "error" range, depending on how "skilled" the computer is. Maybe a poor AI only gets a steering angle correct within +/- ten degrees, and takes a second to notice and have another go at correcting it (with the same chance for error). Maybe a good AI is +/- one degree, with a 0.2 second chance to notice and correct. Ditto braking, acceleration, and any other input.

For the second, a computer can evaluate multiple options. A perfect computer will always select the "best" one. Again, a random chance can be assigned to select the second option, or the third, and so on. Good AI picks optimal choices more often. Bad AI picks optimal choices less often.

Of course, it's always still just a script. But I see no reason that the script can't be made complicated and variable enough that it's indistinguishable from a real driver. I can't think of any characteristic that humans have in terms of the way they drive that couldn't be replicated by a suitably complex machine.

I imagine you could make a kind of Turing test for driving AI. Drive with a bunch of people and pick which are humans and which are AI. I can't think of any AI currently that would pass such a test consistently, but there are few that might get close on short timescales. Time will only improve those.

 

[Dynamite_Dave]

Was round my friends house earlier that drivatar in forza 5 is pretty cool

 

[Goudvis]

Being a part-time AI scripter, I can tell you know, the amount of calculations needed fro AI is incredible. I know that a lot of people complain, but if you all were to become programmers and modellers, you'd be singing a different tune!

Now, I completely agree that these things need to be improved and PD have had a while to upgrade them, favouring other items instead of it. But if you think about it, it does make sense.

People have been asking for sounds and PD are somewhat improving them, but that is a department I don't know well, so I cannot comment.

People have been talking about fixing up standards and PD have done some model tweaking and more importantly texture cleanups (alpha channels, higher res etc)

PD have also gone and made the game much more awesome for those of us who love to hotlap cars and compare them, which I personally do a lot! Having tyre physics, aerodynamics based of simple CFD calculations (most likely off the Large Eddy Simulations, which take the least amount of computing power), etc etc.

As an Aerospace Engineer I can tell you that working out Aerodynamic properties of cars is a pretty tough gig, even for a computer, so hats off to PD if they've succeeded in that, while still making the game look graphically very nice (barring some standard cars of course)

Now coming back to AI:
I have a feeling the AI has been dumbed down somewhat deliberately to appeal to the masses. While I would not mind having to work very very hard to get to first place (literally sit there for an hour before a race tweaking the car to the conditions so it can compete with the rest of the field), I doubt many of us would do that, especially if the race is only going to last 10 minutes. While that may be a mistake, to change it so that all the cars would be more bunched up, aggressive, accurate etc is a mean feat. What you're proposing here is something like this (albeit this is very simplified)


While (1)

If corner:

cornermode = 1​

else:

cornermode = 0​

EndIf

​

If (cornermode == 0)

Throttle = accelerate
Steering = straighten
If (RPM == revlimit*0.95)

If (gear == maxgear)

Throttle = hold​

Else

Throttle = release
Gear = up
Throttle = accelerate​

EndIf​

EndIf​

EndIf

If (cornermode == 1)

Throttle = release
Brake = apply
Steering = turn​

EndIf

If (corner == approaching)

cornermode = 1​

EndIf​

EndWhile


This is the absolute basic of how an AI could work, and even then I'll highlight the problems with this script (if it worked)

- No idea about how sharp the turn is
- No knowledge of car (is the car automatic, what is the clutch like? Does it need to release accelerator to change gear? Which car has better brakes? Which has better grip? What about the width of the car?)
- No track awareness (What if a car spins in front? What is a racing line? What does reduced grip mean? What if I spin? Which way is forwards? Do I need to brake earlier? Can I accelerate? What happens if the car slides? What is an overtake? How do I overtake? What can I see? Where is the track edge?)
- Steering is digital (Only three states: full left, full right and center)
- Speed through corners (What's the ideal speed? Am I in the right gear? Do I accelerate? Can I brake a little in the turn? What if somene cuts me up? What is the apex? How close do I get to the wall? Do I have a runoff or is it a wall? Can I afford to push a little harder? Or my tyres screaming under the strain?)
- Condition awareness (Do I slow when its raining? What about snow? Are my tyres right for the conditions? What is my tyre condition? Night racing, do I flash my lights? What can I see? Where are the other cars? How far am I to P1? What position should I aim for? What is the state of my fuel? What is the state of my competitors? Am I faster in one part of the lap than others?)

These are just a few of many problems with that tiny script. Now just to give you an idea, if I was to fix just the very first point, it would take me at least all day to figure out how I was to code it, and I would probably end up tripling the size of the current PsuedoCode.
If I was to fix all of these, we're looking at months and months of hard work with lots of testing, reverting to old code, and thousands of lines of code. Would my AI be as good as GT6? Maybe, but would it then be compatible? Upgradable? Would it be efficient and functional or would it eat up all the system resources? There is a lot more to coding AI than just saying

BrakingPoint = Correct
Speed = Fast
Aggression = Acceptable

And unfortunately many of us on GTPlanet and in fact everywhere do not understand this. What PD have done for the game is great. It may not be what everyone is expecting, but try to see if from the perspective of one of us programmers, and you'll realise that it is difficult indeed. That being said, I am not saying that we have to be happy with what we have been given. Constructive criticism is what makes a good game great, but if and ONLY IF you know the challenges involved when you make those critical remarks. I hope that this post gives a small taster of what challeges PD face with their small team and makes us a little more appreciative of what they have acheived, while supporting them to do even more in the future, learn from mistakes they make along the way and give the fans everything they want!

👍

There is no way AI is manually coded like this nowadays. There's no programmer tweaking large nested if-else structures. This would become way too complex and unmanageable after a while. The results would be very prone to bugs/exploits.
AI is implemented with machine learning/pattern matching. It basically replays scenarios/patterns of racing scenarios.

But your point is valid. This is very complex and people do underestimate how hard it is.



I don't think people really want the AI to be like real racing drivers. Real racing drivers will avoid crashing into barriers at all costs. People want the AI to be as competitive and risky as themselves. Maybe that's the reason for the frustrations.

 

[ECGadget]

There is no way AI is manually coded like this nowadays. There's no programmer tweaking large nested if-else structures. This would become way too complex and unmanageable after a while. The results would be very prone to bugs/exploits.
AI is implemented with machine learning/pattern matching. It basically replays scenarios/patterns of racing scenarios.

But your point is valid. This is very complex and people do underestimate how hard it is.



I don't think people really want the AI to be like real racing drivers. Real racing drivers will avoid crashing into barriers at all costs. People want the AI to be as competitive and risky as themselves. Maybe that's the reason for the frustrations.

Just to point out that that was just an example, in PseudoCode most things boil down to if-else structures! But you'd be surprised to see just how much of AI is developed using case/if-else structures!
And I don't know... I'd like the AI to be more life like, because I hate the fact that I need to drive like a complete nutter or have a powerful car to catch the front of the pack, however if I was 5th in a close race where the AI was life like, I'd be completely happy!

 

[Lawndart]

Well, i got myself a copy. Been playing for a couple o hours now.
Heres what i think so far...

Main menu, i prefer gt5. Looks like it had more events.
Car dealerships are nicely organized by continent, but heres whats already caught me by surprise.
Checking out different dealerships, went to dodge....no dodge viper!!!!! Wtf!!! So i looked arount for a bit, thinking it would be crazy if there were no vipers, i love those cars, so after a few minutes i saw the familiar SRT initials and clicked on it....all the vipers are under srt now. Why? Just makes it confusing.
Dont like the way the cars are listed in the dealership. Gt5 better.

Licence tests are now in the event categories. Which is ok with me. I preffered the gt5 one stop licence category. But whatever that doesnt really matter.

No used car dealership either. All cars are available to buy in dealerships apparently. Thank God. Hated cycling that UCD

Gameplay...step one. Forced to buy honda fit. Raced a couple of races, upgraded to honda CTR'08.
I play on my G27 wheel. The ride is smoother, a little different than the demo, feedback isnt very informative. The quality of ffb has improved, but it feels like its lost a bit of feel, felt a bit dead at times.that reminds me, i have to adjust my ffb settings.

The AI is still super easy, dont put up much of a fight. Thats a let down. Hoping they get more agressive as the game progresses

Cars look much better, the graphics are more polished. Maybe colors are a bit more vivid, giving the tracks a unrealistic look. I was hoping for better graphics this time around, but i guess we gotta wait for ps4 for that.
Graphics are smoother for sure and so is the frame rate. Have yet to try it in 3d.

Ive yet to modify any cars or even fully explore the menus. Lets hope it gets better.

First impression....id rate it 8 out of 10, compared it to GT5
It is a better game, but is it worth the price. Ya i guess...was bored with gt5 and now i can start a gt game from zero once again

The online servers are down for maintainence. I suppose they havent come online yet.
Please dont have a trading, birthday car feature that can be exploited.

How do you guys like it so far?

I really think a cheap way to deal with AI shortcomings would be a swift penalty for touching (slightly more than a tap) AI cars... Combined with a bit more awareness of the AI's surroundings. It's not a solution but more of an attempt for pd to buy more time... Too often we use the AI as a wall and then complain they are a pushover.

And I also suspect the AI uses the driving line and not their cars capability when determining breaking points and that needs to change as the driving lines are wrong, its a generic line throughout the track for any car... Over emphasizing elevation change and closing turns.

In real racing its not easy to get around a slower guy riding a perfect line...