What makes good AI?

[Stonemonkey]

That said, however, it seems to me that there are a lot of programming tricks going on to create an illusion of reality in the way the AI "drives". It's obvious that each AI car is programmed to carry out the exact same maneuvers in every iteration of the race until there is an "interaction with other drivers" or a "mistake", even though mistakes are pre-programmed too. It all saves processor cycles.

Ive seen similar thinking about the AI and that it's behaviour is somehow scripted, I don't believe that's the case though.

Some time ago I wrote a racing game, fairly simple and not particularly good physics (look up YABASIC PS2 F1 on youtube to see it). For the human player the physics routine took input from the controller and worked out what the car should be doing, for the AI players, the AI routine looked at the track ahead of it, worked out if it should be turning/braking/accelerating and passed that as if it was button presses to the same physics routine so the AI was acting as if it was playing through a controller too, the AI also looked at the positions of cars immediately in front and behind and could simulate pressing buttons to move over to pass or block.

When I started working on the AI, one of the things I found out was that if there was no human player then the race would always run exactly the same way no matter how many times i re-ran it and that included any mistakes or passes or blocks the AI made, to counter that I gave each AI player a small random delay for their reaction to the start light going out, from that point on the AI was always just running the same code but the race ran completely differently each time. I could however store the start delay values and replay that same race.

When I introduced a human player into the mix, that added more interference and the cars would be reacting to the player, without the random start delay the cars ahead would race out the same race each time until they were interfered with by the player. For replays I found that I only had to store the human players input, as long as the start light delays for each AI car were stored the AI would still play out exactly the same race with the recorded human players input as it did when the human was playing so I didn't have to store any data from the AI cars other than the start delays for the replays.

From all I've seen in GT, it's doing pretty much the same thing just with far more advanced physics and AI than I could manage but there's no scripting or anything telling the AI what to do, it just does it's thing and given the same starting conditions it will always play out the same way until the human player interferes with it.

 

[Griffith500]

Ive seen similar thinking about the AI and that it's behaviour is somehow scripted, I don't believe that's the case though.

Some time ago I wrote a racing game, fairly simple and not particularly good physics (look up YABASIC PS2 F1 on youtube to see it). For the human player the physics routine took input from the controller and worked out what the car should be doing, for the AI players, the AI routine looked at the track ahead of it, worked out if it should be turning/braking/accelerating and passed that as if it was button presses to the same physics routine so the AI was acting as if it was playing through a controller too, the AI also looked at the positions of cars immediately in front and behind and could simulate pressing buttons to move over to pass or block.

When I started working on the AI, one of the things I found out was that if there was no human player then the race would always run exactly the same way no matter how many times i re-ran it and that included any mistakes or passes or blocks the AI made, to counter that I gave each AI player a small random delay for their reaction to the start light going out, from that point on the AI was always just running the same code but the race ran completely differently each time. I could however store the start delay values and replay that same race.

When I introduced a human player into the mix, that added more interference and the cars would be reacting to the player, without the random start delay the cars ahead would race out the same race each time until they were interfered with by the player. For replays I found that I only had to store the human players input, as long as the start light delays for each AI car were stored the AI would still play out exactly the same race with the recorded human players input as it did when the human was playing so I didn't have to store any data from the AI cars other than the start delays for the replays.

From all I've seen in GT, it's doing pretty much the same thing just with far more advanced physics and AI than I could manage but there's no scripting or anything telling the AI what to do, it just does it's thing and given the same starting conditions it will always play out the same way until the human player interferes with it.

This is really cool, thanks for sharing. 👍

However, the bits of the AI looking for opponents to block etc. technically fall under what gamers would call "scripting", although it is at a high level (in terms of the AI's "cognitive functions"). You have to anticipate each desired behaviour at that level and explicitly code that behaviour, no? I understand why that happens, of course.

However, a similar thing happens with the driving skill - you have to create the rules (based on the track geometry, physically interpreted for generality) that govern the control inputs, somehow - GT adds variability there by extending the reaction delay to other actions, namely obstacle avoidance.

There are ways for the AI to figure the actual physical driving technique out in its own, but that doesn't happen in the game (or does only rarely), rather only as it's being made, the resulting "knowledge" being shipped with the game in a data file.

I love that a small perturbation in the start delay results in a different race; combined with the "complex behaviour from (relatively) simple rules", this is the essence of "chaos".

 

[Stonemonkey]

This is really cool, thanks for sharing. 👍

However, the bits of the AI looking for opponents to block etc. technically fall under what gamers would call "scripting", although it is at a high level (in terms of the AI's "cognitive functions"). You have to anticipate each desired behaviour at that level and explicitly code that behaviour, no? I understand why that happens, of course.

However, a similar thing happens with the driving skill - you have to create the rules (based on the track geometry, physically interpreted for generality) that govern the control inputs, somehow - GT adds variability there by extending the reaction delay to other actions, namely obstacle avoidance.

There are ways for the AI to figure the actual physical driving technique out in its own, but that doesn't happen in the game (or does only rarely), rather only as it's being made, the resulting "knowledge" being shipped with the game in a data file.

I love that a small perturbation in the start delay results in a different race; combined with the "complex behaviour from (relatively) simple rules", this is the essence of "chaos".

Yeah, chaos theory came to mind when i realised what was going on

The blocking moves weren't really what I'd call scripting, from what I remember the way the AI worked was something like this

The track was defined as a centre line with the edges calculated with trig a particular distance from that centre line. Tha AI looked at the track ahead to see which way it was turning and an offset would be applied to the centre line to the opposite side of the track and the AI would steer the car towards the offset line, that way it would approach the corner from the outside, i cant remember exactly but there were similar rules that then moved the offset to the inside to try to get the car to steer towards the apex and then to exit on the outside and braking was done in a similar way by looking how much the track ahead turned and the faster the car was going, the further ahead it looked.
If there was a car in close proximity in front then the AI would look at which side of the centre line it was on and the offset would be moved to the other side of the track from that to make the pass, similarly if a car was in close proximity behind then the offset would be adjusted but this time to the same side of the centre line as the car behind to block it. Always though, the steering was controlled by the AI to try to move the car onto the offset line.
You can see all that fairly clearly going on at the start of the youtube vid when the cars first take off from the line and they keep switching sides of the track.

 

[Griffith500]

Yeah, chaos theory came to mind when i realised what was going on

The blocking moves weren't really what I'd call scripting, from what I remember the way the AI worked was something like this

The track was defined as a centre line with the edges calculated with trig a particular distance from that centre line. Tha AI looked at the track ahead to see which way it was turning and an offset would be applied to the centre line to the opposite side of the track and the AI would steer the car towards the offset line, that way it would approach the corner from the outside, i cant remember exactly but there were similar rules that then moved the offset to the inside to try to get the car to steer towards the apex and then to exit on the outside and braking was done in a similar way by looking how much the track ahead turned and the faster the car was going, the further ahead it looked.
If there was a car in close proximity in front then the AI would look at which side of the centre line it was on and the offset would be moved to the other side of the track from that to make the pass, similarly if a car was in close proximity behind then the offset would be adjusted but this time to the same side of the centre line as the car behind to block it. Always though, the steering was controlled by the AI to try to move the car onto the offset line.
You can see all that fairly clearly going on at the start of the youtube vid when the cars first take off from the line and they keep switching sides of the track.

Indeed, I see what you mean.

That offset approach is a natural way to look at it, it's how I handled grouping in crowds (total hack, but given more time...) - I do wonder if it runs into complexity problems as you build it up. Of course, the "quality" of the resulting racing line depends on the quality of the look ahead model and the tuning of the offsets.

So it's possibly clear why games tend to go with pre made racing lines.

There was another user on here who made a procedural 2D racing game, in which the AI learned the fastest way around the track using physical controls and some kind of machine learning integration (may have actually been a genetic algorithm selecting and "evolving" control behaviours). The resulting "knowledge" was transferable from track to track; it was fascinating.

 

[Stonemonkey]

Indeed, I see what you mean.

That offset approach is a natural way to look at it, it's how I handled grouping in crowds (total hack, but given more time...) - I do wonder if it runs into complexity problems as you build it up. Of course, the "quality" of the resulting racing line depends on the quality of the look ahead model and the tuning of the offsets.

So it's possibly clear why games tend to go with pre made racing lines.

There was another user on here who made a procedural 2D racing game, in which the AI learned the fastest way around the track using physical controls and some kind of machine learning integration (may have actually been a genetic algorithm selecting and "evolving" control behaviours). The resulting "knowledge" was transferable from track to track; it was fascinating.

I was trying to generate the racing line on the fly within the AI, the reason for that was the way I generated the track which was purely done with data statements within the code (the whole thing was coded using the joypad and no external files btw) and the way I was doing that wouldn't allow for easy definition of the racing line. So it had to be able to drive round any track I designed in the data, although very far from perfect it did work and was far more than I'd hoped to achieve when I first started out on the project. PD however, have slightly more in the way of resources and editing tools so it wouldn't be all that much work to define the racing line around a circuit. I would imagine though that the cars drive that in a similar manner, using the AI to determine what control inputs would be required to best follow that racing line along with some behaviour rules regarding other cars in proximity and then pass those inputs the the physics routines.

There was a lot of tweeking of various parameters to get it to be able to get round the tracks but once that was working I could do things like adjusting how far ahead on the track the AI looked or widening the dead zone that it didn't react to so I could adjust the level of the AI which could cause it to miss braking and go wide or offtrack more/less.

I suppose one thing I could've done would be to have it adjust some of the parameters automatically each lap and time itself and then leave it running overnight to find which combinations gave the fastest laps.

 

[rixtar]

Ive seen similar thinking about the AI and that it's behaviour is somehow scripted, I don't believe that's the case though.

Some time ago I wrote a racing game, fairly simple and not particularly good physics (look up YABASIC PS2 F1 on youtube to see it). For the human player the physics routine took input from the controller and worked out what the car should be doing, for the AI players, the AI routine looked at the track ahead of it, worked out if it should be turning/braking/accelerating and passed that as if it was button presses to the same physics routine so the AI was acting as if it was playing through a controller too, the AI also looked at the positions of cars immediately in front and behind and could simulate pressing buttons to move over to pass or block.

When I started working on the AI, one of the things I found out was that if there was no human player then the race would always run exactly the same way no matter how many times i re-ran it and that included any mistakes or passes or blocks the AI made, to counter that I gave each AI player a small random delay for their reaction to the start light going out, from that point on the AI was always just running the same code but the race ran completely differently each time. I could however store the start delay values and replay that same race.

When I introduced a human player into the mix, that added more interference and the cars would be reacting to the player, without the random start delay the cars ahead would race out the same race each time until they were interfered with by the player. For replays I found that I only had to store the human players input, as long as the start light delays for each AI car were stored the AI would still play out exactly the same race with the recorded human players input as it did when the human was playing so I didn't have to store any data from the AI cars other than the start delays for the replays.

From all I've seen in GT, it's doing pretty much the same thing just with far more advanced physics and AI than I could manage but there's no scripting or anything telling the AI what to do, it just does it's thing and given the same starting conditions it will always play out the same way until the human player interferes with it.

Very interesting post. You did well with YABASIC to get it running so well. I used to do some programming with Microsoft BASIC (I assume it's similar) and it was slow and cumbersome. I can't even remember what I used it for now it was so long ago.

The reason that I suggested that there might be a few shortcuts used in the programming of GT6 is the way PD themselves say it was difficult to code due to hardware limitations, especially memory. Anywhere they would be able to reduce the amount of running code would thus enhance the performance of the game.

Having thought about this a bit more after reading your posts I think that you are right. For example, the chaos introduced by having a human player driving around the track in the reverse direction, forcing all the AI cars off the track, would be impossible to achieve using anything but some level of AI. It doesn't necessarily mean that each AI car is using the full physics engine, though. Reducing the physics involved for each AI would be a good place for PD to save some hardware resources.

It's interesting that your game, without a human player, always ran the code exactly the same way. I can see that this would make it look like it had been scripted without any real AI activity at all. One thing that I'm wondering about though is: if you can change the way the AI interacts with each other by simply changing the starting times of the cars by using a small random delay, why don't PD do something similar each time a race is run in GT6? This simple change could make nearly every iteration of a race in GT6 look different and we wouldn't see the AI cars running into each other or off the track or overtaking each other at exactly the same place every time! Maybe there are other factors involved.

 

[Griffith500]

What would be the motivation for not using the physics model already tuned etc.? Kaz said that physics is not a computational load problem (with the current games), so it wouldn't be that.

Why would you create more work in managing a separate physics model, and moreover, one that should not be discernible from the existing model? Using such a model pushes your AI skills further away from where you want them to be: realistic.

The difference that you see is likely one of sensitivity. The skill (code) GT's AI has at getting to the racing line (which is provided to it) is perhaps far more effective than a variance of a few hundred milliseconds in delay can cover up. But we don't really have standing starts any more, either. Emergence is great, but it works best when you can control its bounds.

It's worth mentioning at this point that GT's AI has a reaction time built into it. Its skill is split up into tweakable components like acceleration, braking, corner entry and exit (separately) etc. These are tunable per entry to a race.

When you have an AI whose base skill provides excellent tracing of a precomputed racing line, but terrible pace in doing so (compared to a human driver following their own line), adding extra ways for that pace to fall away (via high level "random mistakes") is probably not a good idea. Especially when the only current concession to their hogging the racing line is for them to lift off and let you pass.

Hidden with all the AI stuff in GT6, you have things like "force overtake" and other such recognisable commands from a BSpec implementation. However there is a separate thing called "Raceroid", wherein there appear features not seen yet; including "force new corner AI". I think it's fair to say that the AI's inability to take corners is its major handicap; if that can be improved, it would be safer to slow it down via other means.

This is why I'm keen to see what BSpec brings to GT6. Then we may have a better idea of where PD is taking AI in the future.

 

[Stonemonkey]

Very interesting post. You did well with YABASIC to get it running so well. I used to do some programming with Microsoft BASIC (I assume it's similar) and it was slow and cumbersome. I can't even remember what I used it for now it was so long ago.

Arrgggg you have no idea, ps2 yabasic was one of the most poorly implemented versions of basic I've ever used (sony only put it on as a tax dodge to say the ps2 was a programmable computer instead of a console). It had some nice basic graphics functions but extremely slow and for some reason, the longer your code the slower it ran and because it was interpreted you have to keep the variable names short too. The pre-cursor I wrote to the F1 had wipeout type ships flying around the track no weapons , 4 ships in race and no trackside detail or anything other than the track drawn on the ground and it ran at 60fps. For the F1 I wanted more detail and even with lots of cheats in the graphics coding it took 1/60th sec just to draw the frame without any physics or AI so I reduced it to run at 30fps and used the 2nd 60th sec for the physics and AI and had a little to spare so was going to add some more to the graphics but never got round to it. IMO it was really pushing the limits of the language.

The reason that I suggested that there might be a few shortcuts used in the programming of GT6 is the way PD themselves say it was difficult to code due to hardware limitations, especially memory. Anywhere they would be able to reduce the amount of running code would thus enhance the performance of the game.

Having thought about this a bit more after reading your posts I think that you are right. For example, the chaos introduced by having a human player driving around the track in the reverse direction, forcing all the AI cars off the track, would be impossible to achieve using anything but some level of AI. It doesn't necessarily mean that each AI car is using the full physics engine, though. Reducing the physics involved for each AI would be a good place for PD to save some hardware resources.

I doubt the code for the physics will take up much memory or much time to process and the same code will be shared over all cars in a race, if you try to include functions for simplified physics you're only adding to the amount of memory already used instead of saving anything and you're not going to get consistent behaviour of the cars on track so I'd avoid going down that path.

It's interesting that your game, without a human player, always ran the code exactly the same way. I can see that this would make it look like it had been scripted without any real AI activity at all. One thing that I'm wondering about though is: if you can change the way the AI interacts with each other by simply changing the starting times of the cars by using a small random delay, why don't PD do something similar each time a race is run in GT6? This simple change could make nearly every iteration of a race in GT6 look different and we wouldn't see the AI cars running into each other or off the track or overtaking each other at exactly the same place every time! Maybe there are other factors involved.

A computer just crunches numbers, given the same start conditions those numbers will be the same every time.

I've wondered why PD haven't done something like that too, although in some cases like some of the challenges it means that every human player playing the game is on an equal footing and can learn to pick their way through a field of cars through repeated attempts, I quite like that in some of the challenges although there are times when I feel a bit of variation would be better. As well as the startline delay I used there are other things that could alter how it plays out, slight random changes to the positioning on the start line which would also work on rolling starts, as would slight random changes to initial speed.

 

[05XR8]

Then I'll ask, can fastest laps from lets say, the top 16 cars from disciplines such as, GT3, NASCAR, Super GT500 & GT300, be programmed on the real world tracks for AI to achieve? Let's say the Jimmie Jonhnson SS is programmed to set its real world fastest laps even if it randomly qualified in 10th place at Daytona. I understand about having a human in the race will change things. But can the AI learn to achieve those lap times and still navigate the field? Is it as simple as looking up a race, finding the YouTube GTR lap times from Silverstone and assigning that car ingame the task of achieving those lap times when possible?

 

[Stonemonkey]

Then I'll ask, can fastest laps from lets say, the top 16 cars from disciplines such as, GT3, NASCAR, Super GT500 & GT300, be programmed on the real world tracks for AI to achieve? Let's say the Jimmie Jonhnson SS is programmed to set its real world fastest laps even if it randomly qualified in 10th place at Daytona. I understand about having a human in the race will change things. But can the AI learn to achieve those lap times and still navigate the field? Is it as simple as looking up a race, finding the YouTube GTR lap times from Silverstone and assigning that car ingame the task of achieving those lap times when possible?

Not quite sure what you're asking, look up 'Top Gear Self-Driving BMW 330i' for a car that can follow GPS data to drive around a track though it wouldn't be able to do that and race against other cars. If it was loaded with sensors like googles self driving cars then it could probably be programmed to race against others, how good it would be or whether it's been done already I don't know.

ETA: sorry I think I got that all wrong but I'm still not sure what you're asking. You could take a recording of a real world lap, convert it to data and replay it in a game but any other player wouldnt be able to interfere with it, if it was solid (as in not a ghost) then you'd bounce off it if you hit it and it would just carry on as if nothing happened, it also wouldn't be AI, just playback of recorded data. The thing about AI is that it makes it's own decisions, it is possible that the AI could be given the lines a real driver has taken round a lap and that it follows that line as a guide but it would also have to be free to deviate from that to take action around other cars.

If you're asking if AI could learn from looking at how real drivers drive then you're looking at AI that's somewhere out there beyond the gaming world.

 

[05XR8]

Not quite sure what you're asking, look up 'Top Gear Self-Driving BMW 330i' for a car that can follow GPS data to drive around a track though it wouldn't be able to do that and race against other cars. If it was loaded with sensors like googles self driving cars then it could probably be programmed to race against others, how good it would be or whether it's been done already I don't know.

ETA: sorry I think I got that all wrong but I'm still not sure what you're asking. You could take a recording of a real world lap, convert it to data and replay it in a game but any other player wouldnt be able to interfere with it, if it was solid (as in not a ghost) then you'd bounce off it if you hit it and it would just carry on as if nothing happened, it also wouldn't be AI, just playback of recorded data. The thing about AI is that it makes it's own decisions, it is possible that the AI could be given the lines a real driver has taken round a lap and that it follows that line as a guide but it would also have to be free to deviate from that to take action around other cars.

If you're asking if AI could learn from looking at how real drivers drive then you're looking at AI that's somewhere out there beyond the gaming world.

Im saying, can the lap time from achieved by a real world driver, be entered as data for the AI to replicate.

What I'm underatanding is, the Ai must be programmed to a set amount of driving lines and learn at what times it needs to brake and accelerate with other AI and human drivers. Im only asking, can the AI be programmed with a certain lap time and achieve it wheter it is first or last in a field of cars.

 

[Harsk100]

Well, in my opinion, to be honest, we should be able to choose the following levels:

The structure like GT3
Easy
Normal
Hard
Professional
Expert


But we should also have the following thing:
Custom- The custom mode allows you to set the AI difficulty. For example, 0% is the easier one. The highest % is the hardest.

In GT6, the AI's slow down to not push our cars and so they can make an good race without contact.

 

[Stonemonkey]

Im saying, can the lap time from achieved by a real world driver, be entered as data for the AI to replicate.

What I'm underatanding is, the Ai must be programmed to a set amount of driving lines and learn at what times it needs to brake and accelerate with other AI and human drivers. Im only asking, can the AI be programmed with a certain lap time and achieve it wheter it is first or last in a field of cars.

I'm not sure AI in general gaming terms is quite what you think it is. Sure you could provide it with a realworld drivers racing line to try to follow but it wouldn't follow it exactly (unless you're just replaying positional data which wouldn't be any use for anything other than a ghost and it wouldnt have anything to do with AI) and game physics do not match realworld physics so the lines and braking/accelerationg points wouldn't match. Then you've got traffic to deal with so for example after making a pass, the car is approaching the corner on a different line at a slower speed so the braking point would be different and the line the car enters the corner would be different so the AI has to decide at some point to make it's own decisions on where to brake and turn rather than using the realworl drivers data, it's all pretty messy stuff and if it's to be capable if switching to making it's own decisions like in the example I just gave then it would be capable without the need of the realworld data at all so it seems a bit pointless to me.

 

[Stonemonkey]

Going back to what some have thought about scripted behaviour because the cars do exactly the same thing each time and I've already said that I think it isn't, but I've been trying out the Sierra rally challenges in GT6 and it seems to depend on how your times are but the cars can appear in different places. In challenge 4, typically when I come round the first corner I see the red car approaching the 2nd corner, I pass it halfway round that corner and then there's a bit of a gap to the next car. However, if I get an exceptionally quick start off the line I don't see the red car until I'm going round the 2nd corner and when I pass it, it's alongside the 2nd car, and things like this seem to be happening all over the whole course. That I would say is possibly down to some sort of scripting which decides where cars should be placed on the track ahead of you depending on your time.

 

[Mike_grpA]

I've noticed the exact same thing on sierra. I think the cars all start roughly around the same place every time, but they seem faster sometimes, and slower other times. I think maybe it randomly chooses how good each ai is at certain things each time you start a race, because watching replays of sierra, I have noticed sometimes a certain car will just sit behind a much slower car, and hardly even attempt to get past, while other times it will try much harder to get past, being way more aggressive. Of course, if you even breathe on it, it will brake and let you past, which annoys me no end about gt ai.

It seems to me, in a championship, this random ai style seems to be set at the start of the championship, and stay the same for each race, until the championship is finished. At least this is the trend I've noticed while grinding the rbx2014 standard championship. If a car is slow and holds others up in the first race, they seem to do it every race. Also, the guy who finishes second, ALWAYS finishes second lol.

 

[Stonemonkey]

I'm pretty sure it's not random, it seems like it's something to do with the time, and the positions cars start in change, the cars are added to the track in front of you as you're driving, they dont all start on the track when you start.

 

[Mike_grpA]

When I've done the sierra events, the cars have all started in the same positions, just behave differently each time. I know they spawn on track in front of you.

 

[Stonemonkey]

The difference can be seen in these 2 videos just in the first 2 cars passed

 

[05XR8]

I've also mentioned in a different thread, if you flash the lights to pass, from the start of your race, the car in front of you will speed up. I have done this to confirm on different tracks and in single player and GT Mode.

When I flashed the lights every few seconds, that car would brake later, spend less time cornering and try to get around cars in front much quicker.

When I didn't flash lights at all, I found the normal lazy AI as usual. Only when I would flash to pass, the AI would perk up and we both catch cars ahead and breeze by "unsuspecting" AI that were "sleeping".

 

[Stonemonkey]

I've also mentioned in a different thread, if you flash the lights to pass, from the start of your race, the car in front of you will speed up. I have done this to confirm on different tracks and in single player and GT Mode.

When I flashed the lights every few seconds, that car would brake later, spend less time cornering and try to get around cars in front much quicker.

When I didn't flash lights at all, I found the normal lazy AI as usual. Only when I would flash to pass, the AI would perk up and we both catch cars ahead and breeze by "unsuspecting" AI that were "sleeping".

Haha, i didn't realise about using lights and have been having to make my way around the AI wobbling around the middle of the road and all other sorts of stupid things. Thanks, I'm off to see if i can improve my score.

 

[TayeezSA]

I know this is a GT related thread, but here's some of Driveclub's improved AI shown in the form of a duel



Thoughts anyone?

In GT related stuff, I have faith that PD can do well with the AI provided they listen to the public and make good decisions

 

[Griffith500]

I've also mentioned in a different thread, if you flash the lights to pass, from the start of your race, the car in front of you will speed up. I have done this to confirm on different tracks and in single player and GT Mode.

When I flashed the lights every few seconds, that car would brake later, spend less time cornering and try to get around cars in front much quicker.

When I didn't flash lights at all, I found the normal lazy AI as usual. Only when I would flash to pass, the AI would perk up and we both catch cars ahead and breeze by "unsuspecting" AI that were "sleeping".

I found references to flashing for the AI, and assumed it was for BSpec, and, as such, yet unimplemented. I guess the flashing would have to have an effect to justify its inclusion as a BSpec "feature", though.

I could imagine that being fun in a real time, online BSpec duel: force your Bob to flash another player's Bob in front to fluster / enrage / unexpectedly positively motivate him.

 

[Stonemonkey]

Flashing light doesn't seem to have any effect in the sierra rally challenge, at least not that i can see. Ive tried at various distances and different types of flashing but not noticed the AI doing anything but the stupid wobbling around the track it always does.

 

[Griffith500]

It could be that just lurking near them in order to flash them causes them to "run hotter" in certain circumstances. Some of the AI car / driver combos are way quicker than the rest, though.

 

[GTR V-spec]

They need to increase the speed of AI vehicles, they need to attempt overtaking manoeuvres aswell as defend them. I'd like to see them learn your faster sectors on the track and try and improve on their own time to keep up. Car selection needs to improve, and standing starts. The overtaking and causing you to crash needs to stop. It's a game where no matter how many overtaking tricks you know you can't use them because the Ai is awful. I'd like to draw ai into outbraking themselves before I undercut them or dani ricciardo style deep undercut to take the line before an overtake.

 

[05XR8]

It could be that just lurking near them in order to flash them causes them to "run hotter" in certain circumstances. Some of the AI car / driver combos are way quicker than the rest, though.

Yes. I follow close, but don't dive bomb or bump draft the car I'm flashing. Try in the GT3 and 500pp race car challenges. Watch your replay data and note the lap times. When you don't flash, you'll see just how much of a difference it makes towards AI lap times.

 

[JKgo]

....................I often use horns and headlamp flashes during offline races; fun thing is, AI cars not only doesn't go any faster (at least in my experience) but actually they politely steps aside for me, in the middle of a freaking corner.
I had a huge L-O-L moment with current 550pp American seasonal at Willows - The lead car invariably is either a GT40 or a Shelby Cobra; in this particular incident, lead car was a 15th Anni. Cobra. I caught up with my Firebird, and at the second right hand turn, I blew me horns and flashed the lights (no contact, of course; I was behind by a couple car length) - AI decided that it couldn't handle the fracas and promptly slammed on the brake, swerved, went off line, and slid into the sandy stuff. All right in front of me.
Me: "Huh, well, you see somethin' new everyday...."

 

[oli440]

Good AI should keep up with you, attempt overtakes as often as possible, but also not be aggressive to the point where they blindly send you off the track. They should appear more human. They should act like they give a damn for the cars around them. They should make mistakes too! Why not throw in some of the B-spec elements into the AI drivers, so they can lose concentration in longer races, and can be pressured into mistakes by others? A bit like R-Racing.

I can't help but feel like Gran Turismo's AI drivers are cybernetic organisms supplied by Skynet...

"MY MISSION IS... TO STICK TO MY LINES LIKE A SLOT CAR AND ANNOY YOU IN THE SIERRA TIME RALLY EVENTS BY CAUSING PILE-UPS".

 

[BrodieBoi94]

I believe that Gran Turismo's A.I have improved noticeably in comparison to GT1-4 since now they actually recognize that you are there. They won't forcefully bump you unless it's a pure mistake.

I can also say that I had dramatic spin out on a high speed downhill section of matterhorn and the A.I were able to circumnavigate me flawlessly. (I was impressed)

During the spin out I thought that I would be rammed by multiple cars.

Back on track I can't honestly call GT's A.I good overall. They lack the drive and ability to pass you such as those found in Grid Autosport.

They also lack the on the edge driving ability such as those found in Forza 4. Forza's A.I really know how to push the cars' performance

These three combinations are what makes good A.I IMO.

 

[Wiski_EK9]

I was driving the ascari "endurance" race today. I was going side by side with a AI (automtic Idiot?) and then, suddenly, he rammed me into the Gras.

That Moments, and that they are to slow are the Most dissapointing things of gt's AI.

I hope that the New one will behave like a Good human.

Drive fair
Drive clean
Drive with respect
Drive as fast as the car can handle it

Its sad that the ai is just a moving Wall :/