Pathetic AI: Formula GT At Nurburgring GP

  • Thread starter Thread starter Troux
  • 37 comments
  • 5,354 views
T

Troux

Premium
Messages
663
I first ran the FGT championship in B Spec, and I noticed that on every lap, Bob would royally screw up the last tight S-curve before the hairpin that leads to the straightaway. The guy was level 30 or so, and he would go WAY off course and do a full 360 coming out of it. EVERY lap, ranging from hot to mild temperament. I thought, "He's been so good past level 20, but I guess he's still got some stupid in him." I simply hit 'Pace Down' right before that S bend and he'd go through it like a normal person. No big deal, he'll get better, I hope...

So last night I finally got around to running the FGT in A Spec. It's not an easy corner, but it's still just an S Bend. SLOW in, easy through, find the corner, and fast out. This bit of common sense BLEW AWAY the competition. I was now noticing that the entire AI grid was doing exactly what my Bob was doing at that last curve. I would be halfway through the straight and could see all these little blue arrows clumped up and spinning in circles. Once again, this was EVERY lap, for 20 laps. God forbid you're stuck behind them at that corner, you can't see through the tire smoke and sand to dodge the sideways stationary cars or the cars that ram into you getting back on course.

PD made a course that is simply incompatible with its own AI. 'Pathetic' was the word that came to mind.
 
The AI would struggle on a drag strip.
 
My bob blew that corner maybe every other lap, i made sure to pace him down before that corner, and pace him on on the front straight. He won gold, and my bob was like level 27, won 1st in that race in every race in the series but monaco came in 3rd.
 
Same for me in A-Spec. Total AI carnage at then end of each of the 20 laps.

They should at least remove the race or perhaps GT5 was cobbled together so late they couldn't or just had very limited testing. It's quite embarrassing for PD.

There's so many things missing and awkwardly worked in GT5, I feel like I would've preferred them to delay it further.
 
Now, there's no proof for what I'm gonna say here. I reached this conclusion through observation and I may be completely wrong, so if anyone actually has some knowledge on the subject and find anything wrong, please correct me.

For pretty much every racing game out there I've played, the AI appears to be pre-programed. What I mean is, someone basically "told" the AI: "You brake here, you accelerate here, you turn like this here..." and so on (not actually how it works, but you get the idea). Then they give the AI some random characteristics like braking a bit earlier or latter, taking a different line to overtake or avoid someone, etc. In some games this pre-programmed characteristic is so apparent that the AI appears to be playing by different rules. I've even seen games where the AI do things which I tried to imitate only to find out they are impossible to do, as if the AI and I are driving in two different physics.

However, as far as I can tell, PD has an AI that actually knows how to drive. What I mean by this is, if you put it in any course with any car, the AI will find it's way through. It's not pre-programed by anyone, it actually judges it's car and the corner, find a line and a speed, and go for it. As a human would. Obviously the AI can't do such a thing as well as a human, that's why it's slow through corners and miss some completely. One of them being the one you mentioned. Obviously the car also play an important role, so while the AI can do it in pretty much any car, in such a different car like the FGT it completely misses the corner.

Again, I'm not sure that's true, and I simplified it a lot to keep the post easily understandable and relatively small. Basically, as far as I know, PD has an awesome AI which is pretty much unprecedented, but it unfortunately still needs a lot of work.
 
Dravonic
Basically, as far as I know, PD has an awesome AI which is pretty much unprecedented, but it unfortunately still needs a lot of work.
Click to expand...

:lol:its no different than the AI failing to get around Seattle on GT2. i have seen nothing to suggest the AI learns the track on its own rather than being on predefined "rails"
 
Well your b-spec driver is better than the standard ai in some ways. But i would not go sofar to say it is a milestone in ai :P.

It is supposedly true that the more your selected bob drives a track/car the better they get at that track and with that car. But i have put total noob bobs on some tracks and they seem to have little problem at all even in a car they never driven before. Nurburg is a good case where the ai even if they never been on the track does really well with it. But some tracks they struggle with.
 
Dravonic is onto something, though I don't know the level of automation present. It's clearly there, since they wouldn't be able to drive the generated courses otherwise.

One thing they do lack, though, is how to judge elevation changes mid-corner. They go too fast over crested corners and too slow through dipped ones.

There's a lot of theory out there and work on finding the best line on a circuit, most of which is actually fairly "simple" mechanics, although finding a "solution" is very tricky.

I still think they could use some work in the way they enter corners, though, and be better able to judge corners according to their context (corners / straights that follow / precede) Their inability there is very evident in the early levels of B-Spec'ing. They won't overtake on mild kinks in straights, for example, even though it's perfectly safe to do so because it's in the middle of a straight!
 
MadduckUK
:lol:its no different than the AI failing to get around Seattle on GT2. i have seen nothing to suggest the AI learns the track on its own rather than being on predefined "rails"
Click to expand...

If the AI were on predefined "rails", then they would theoretically have no difficulty whatsoever in navigating a track.
 
I noticed this to, every single lap without fail at least two of them would spin out here. I was in A-spec and fell back to 12th early as I don't know that track well and made some mistakes, by the end of the race I was in first with a huge lead thanks to that.
 
Dravonic
Now, there's no proof for what I'm gonna say here. I reached this conclusion through observation and I may be completely wrong, so if anyone actually has some knowledge on the subject and find anything wrong, please correct me.

For pretty much every racing game out there I've played, the AI appears to be pre-programed. What I mean is, someone basically "told" the AI: "You brake here, you accelerate here, you turn like this here..." and so on (not actually how it works, but you get the idea). Then they give the AI some random characteristics like braking a bit earlier or latter, taking a different line to overtake or avoid someone, etc. In some games this pre-programmed characteristic is so apparent that the AI appears to be playing by different rules. I've even seen games where the AI do things which I tried to imitate only to find out they are impossible to do, as if the AI and I are driving in two different physics.

However, as far as I can tell, PD has an AI that actually knows how to drive. What I mean by this is, if you put it in any course with any car, the AI will find it's way through. It's not pre-programed by anyone, it actually judges it's car and the corner, find a line and a speed, and go for it. As a human would. Obviously the AI can't do such a thing as well as a human, that's why it's slow through corners and miss some completely. One of them being the one you mentioned. Obviously the car also play an important role, so while the AI can do it in pretty much any car, in such a different car like the FGT it completely misses the corner.

Again, I'm not sure that's true, and I simplified it a lot to keep the post easily understandable and relatively small. Basically, as far as I know, PD has an awesome AI which is pretty much unprecedented, but it unfortunately still needs a lot of work.
Click to expand...

I don't think the AI in GT5 is that advanced.
I've got a good idea what they are doing, as I am a software developer.
I'll try to put it into laymans terms as best I can...

So you have a base AI program.
For a given track, the AI knows th general information about the track: where it is straight, where there are bends and the bends characteristics, and a base line speed that it follows around the course.
Then, these baselines figures are skewed based on your Bob and the car.

Let's say that for our test track we have the following scenario:
1.) Take straight away at 100mph
2.) Beginning braking for corner 100 feet before entrance
3.) Enter corner at 55mph
4.) Turn at a 20 degree angle towards the apex
5.) Exit corner at 40mph
6.) Increase speed on next straight away to 80mph

Now, let's assume that was based on a Bob level of 20 (median value).
If our Bob is a level 10, then the basic AI remains the same, just the values get adjusted.
1.) Take straight away at 90mph
2.) Beginning braking for corner 120 feet before entrance
3.) Enter corner at 45mph
4.) Turn at a 24 degree angle towards the apex
5.) Exit corner at 30mph
6.) Increase speed on next straight away to 70mph

In both of those cases Bob was driving a car that meets median specs. Without getting into all the tuning options let's just say the median is 300hp and 2000kg.

Now let's say we take our level 10 Bob and put him in a car with 400hp and 1800kg. Now his numbers might look like this:
1.) Take straight away at 95mph
2.) Beginning braking for corner 110 feet before entrance
3.) Enter corner at 50mph
4.) Turn at a 22 degree angle towards the apex
5.) Exit corner at 35mph
6.) Increase speed on next straight away to 75mph

Of course, this is a simplified view, but I believe this is the basic formula that I am seeing carried out by the AI. Throw in the occasional random "screw up this corner", "take this corner perfectly", commands to mix things up, have some basics avoidance intelligence (really basic in GT5!) and you have the basis for a simple scalable AI system.

Hope that made sense and helps a little.

:cheers:
 
Last edited:
The Ai does have problems with some corners. I was doing the Ferrari race at Monza (well I was in a Ferrari and it was Monza) on A spec. Come upon the pack at the first chicane where they managed to make a complete mess of it. I avoided them coming in to the corner well down the pack and leaving it in 1st place (I think... I did save the replay). AI and some corners are not compatible. I remember B spec on GT4 just could not negotiate the New York track regardless of what I did.
 
I witnessed this tonight. I had a terrible start to the race as I spun out on the FIRST corner so I was thinking about damage limitation and the steady grind up to a respectable place but when it came to that chicane and saw the carnage, it really made my day. Was an easy race from there and every lap seeing cars spin out just made me laugh.
 
^That makes plenty of sense, and seems to be exactly what the AI in this game does. The mid-corner braking clearly shows that they are explicitly toning down potential, and the way the driving line is detailed shows us that the game knows how the "sequence" a track, even dynamically.

Watching my Bob, and the computer AI make the same mistake lap after lap pretty much proves outright that the AI is not evolving or learning a track any better. Also, the way the AI deals with contact of all sorts is plenty of evidence to me that they are operating on their own set of physics. They still have that classic GT feel of weighing several times more than your car.

Jeremy Ricci
Simply pace him down and he doesn't blow the corner, it's that easy.
Click to expand...

Did you just respond to the thread title without reading a word in the thread or what???
 
Just checked. It was a 2 lap special event at Monza (no idea where, that is what the replay says). I was in a 430 Scuderia starting 16th. I was following an Enzo to the first corner where the other 14 cars got into a tangle. The Enzo got through OK and I ended up in 2nd place which was something of a relief when I have to rely on my own overtaking ability.
 
Spagetti69
The AI would struggle on a drag strip.
Click to expand...

They would. They'd hit the brakes before passing, ruining all momentum from their draft. So yes, they really would.
 
BKViper
I don't think the AI in GT5 is that advanced.
I've got a good idea what they are doing, as I am a software developer.
I'll try to put it into laymans terms as best I can...

...

Hope that made sense and helps a little.

:cheers:
Click to expand...

I think you might have done yourself a disservice by over-simplifying!

My theory is that, since the game knows where corners are (it counts corners in course creator tracks) the track is divided up into these segments, and each corner has a width, radius (internal and external), camber, elevation change (which seems to be ignored in course creator tracks) that is easily obtainable. Each car has a "curve" representing the maximum speed it can attain for a given radius corner.

Now, all that remains is to work out the driving line itself, which depends on straights and other corners, to work out where the optimal entry and exit points are for each corner. Then it serves to work out the best way to brake into the corner and how to accelerate out and steer to the next exit / entry point for the next segment, which turns out to be straight-forward, too.

Obviously, the output here can be tweaked for the library circuits, but not for the course creator ones (hence the tendency to get certain types of corners wrong!)

There's a nice old article on the matter, and it really is very simple (mechanically, I mean; obviously the specifics are tricky, on a car-to-car, corner-to-corner basis). It's part 17 of the Physics of Racing, which takes a "back of the envelope", "first principles" approach to racing. The maths can be "heavy", but the principles are, as I said, straight-forward.

The clever part is how the cars drive off-line, although low level B-Spec gives that one away, too...
 
BKViper
*snip*
Click to expand...

I believe the AI you describe is the kind of AI most games have. Basically, you set rules, which are track and car dependent. You work an equation to scale them so they work fairly well with similar cars. But this is where the problems start to show up. You'll probably need different rules for different types of cars since your scaling equation will probably not work with them, maybe you need different equations too, maybe depending on the kind of drivetrain you'll need different rules, etc... The rules are basically the "telling the AI what to do".

In a game with 1000 cars from very different categories and quite a few tracks, this can easily be blown out of proportion. They would waste a lot of time making those rules. They need a different way since this one is probably not feasible. And what about the automatically generated tracks and rally stages? The AI wouldn't be able to race those since there would be no rules for them.

What I believe they did is what Griffith500 described. Basically, you try to teach the AI how to drive. This way, it will work everywhere, even at the automatically generated courses. However, making an AI that can drive isn't easy. With so many different cars and tracks it's no wonder it will struggle at some point.

But again, this is just my theory.
 
I think Monaco is too hard for this game's AI as well. In A spec I had to restart the race 15 times before I made it through the first corner without getting destroyed by the AI, then when I finally accomplished that, I got to the chicane and found someone parked in front of it. I slowed down, nobody behind me did.

Then I tried it in B spec and it was exactly the same thing. Big pileup in the chicane, my guy slows down, gets annihilated by everyone behind. Went from 4th to 12th I believe.
 
I see you guys are giving the AI the benefit of the doubt in terms of adaptability and 'learning' capacity primarily based on the course creator. I think the very restraining limits on the course creator should speak for just how limited the AI's capacity is. Also, the AI seems pretty clueless when it comes to elevation changes, based on how I've seen them charge full speed at jumps (Cape Ring is a good example, but take a look at the "fun" course creator pictures and videos).

The way AI acts at chicanes and S-bends makes me think their programming is getting jumbled (in much the same way they do!), as if approaching a corner while navigating a corner is too convoluted a task for them.
 
I have just finished the Formula GT Grand Prix (A-Spec). I overlapped all 11 opponents, because they crashed every lap in the chicane. I luckily saved the replay. :)
 
I'm getting this dumb AI in that corner as well. Every lap, atleast 3 of the cars in the top half hit the gas too early, jump the curb, and shoot off into the grass. On one occasion, 2 of them spun back on to the track, and blocked everything. All 11 AI cars clumped up. When the AI spins, they wait and wait and wait for a break in the traffic to pull back on. This is the case when they are stranded on the track, right on the racing line too. They don't move.
 
That is why that race is easy. Also, it seems the AI ahead of you will spin out for sure. I mean I lost the lead 3 times in that race, went back to the pack and every time managed to get back ahead.
 
xNEVER-ONEx
That is why that race is easy. Also, it seems the AI ahead of you will spin out for sure. I mean I lost the lead 3 times in that race, went back to the pack and every time managed to get back ahead.
Click to expand...

Yeah, I made many mistakes as well due to greedy on speed.

Another funny thing is: when an AI by whatever reason is closing me up while I'm leading, then comes the chicane and I'll will say goodbye to it 8-9 of 10 times, and laughing at that race map seeing about half of them spinning 360 out to grass...:dopey:
 
paytonlow
Yeah, I made many mistakes as well due to greedy on speed.

Another funny thing is: when an AI by whatever reason is closing me up while I'm leading, then comes the chicane and I'll will say goodbye to it 8-9 of 10 times, and laughing at that race map seeing about half of them spinning 360 out to grass...:dopey:
Click to expand...

I think the most interesting part in the FGT races is the boosting. I mean notice how the same couple of cars with be right behind you no matter how many times they spin out.
 
Anyone else notice the difference from the A spec AI and B spec AI?

In b spec if my bob is running at half pace he is fine, anymore pushing from me he decides he is gonna take a stroll in the grass. However in b spec, his fellow drivers almost never pull the same feat is in a spec. Last race I did I think maybe one driver went in the grass and spun out because my bob forced him into the grass.
 
Back