So......lets talk AI?

[Msportdan]

So firstly I want to say good job on the AI (so far) suppose im worrying that you will ruin the AI like what AC has done to its AI. But to be fair simbin have always done a competent ai. I think theyre good they overtake you (unlike cms f1) theyre quite quick even on low 90s level. I would like to see them better RF2 aI.

However maybe we could just talk about points about the AI and what you would like to see, or how they perform.

*I would like to see them defend a tight corner (hairpin) if your right behind I like them to close the door on the inside. (anticipate your out breaking move)

* Not get bunched up that sometimes happens. when accident occurs I think?!

What do you think? had a good moment with them. ?

 

[Arthur Spooner]

I've been trying the WTCC cars on Suzuka (and a bit on Hungaroring) with AI set at 98. I've noticed that the AI cars seemed to have problem with oversteer in long sweeping corners and corners where you enter with more speed, losing the rear. The first corner at Suzuka takes at least one car out every lap. The following section seems to be problematic too. There was some corners at Hungaroring that also caused some trouble. Did anyone else experience this? Haven't tried any other difficulty level yet.

But other than that I must say that the AI is behaving pretty well.

I can confirm this. AI has difficulties with oversteering of the FWD cars. But I must admit I had difficulies with the FWD oversteer too in the beginning. But since I drove the WTCC cars in Race 07 for years I was able to alter my setup so that I have better control over the rear end. I only tried the Honda so far, maybe it's different for the other ones. In Race 07 the different FWD cars also were diverse in their tail-happiness. While the Seat was very stable, the Volvo S60 was so tail happy that it even had a higher tyre wear at the rear if you didn't apply some extreme setup. The S60 was really weird... I always wondered if the S60 physics were a bug or a feature.

 

[shardshunt]

I just had a fantastic race around bathurst in the touring classics and the thing that surprised me was the mistakes the ai made were right for example one of the ai stepped of the road around the kink in conrod straight lost the rear end and missed the turn speard into another car this is a semi-common mistake irl. I would like to have seen a safty car but otherwise it was great to see.

 

[Sk3ptik0n]

I wanted to say that I like the AI so far. I'd like to see a bit more dynamic racing on their part but overall I think it's one of the best (especially on the new crop of games, which is not saying much, I am afraid).

I have 2 problems with it though:

1) Adaptive AI seems to be all over the place, clearly I am better at some tracks than others, but to beat the AI by 5 seconds in one track and then be 2 seconds behind at the next seems to be a bit excessive. Hopefully it will improve, but I'd like to know what the logic behind it is.

2) I took the time to compare the qualification results of some of the RL WTCC and while there is usually 1sec spread for the top 10 or so, overall some cars are up to 6 or 7 seconds off the leaders time. It would be nice if the game mimicked that, even a bit more if possible.

That would allow us to get into a race and always have some cars to race against, as opposed to taking the led and then lose the pack in the rear view mirror, or have the entire field so much faster that you end up racing alone. Frankly, when I heard "Adaptive AI" I figured that it worked more like if you were faster than the pack in practice by a good bit, a few cars would improve at least enough to give you a good race in quali and the race. Conversely, if the game sees you are consistently 2 or 3 seconds off the pace, it would slow down some of the cars at the back to at least give you a good race even if in the mid pack or in the rear.

Frankly, so far I have been less than impressed with this "adaptive" experiment. In fact, on certain track-car combo I went back to setting the AI manually for better racing.

 

[Arthur Spooner]

1) Adaptive AI seems to be all over the place, clearly I am better at some tracks than others, but to beat the AI by 5 seconds in one track and then be 2 seconds behind at the next seems to be a bit excessive. Hopefully it will improve, but I'd like to know what the logic behind it is.

This effect was already there before offline races were introduced. In free practice you had the adaptive AI already and it was just as you described here. I even encountered this behaviour between two sessions on the same track.

I guess the reason for the introduction of the adaptive AI was that in gMotor sims you always had very different results if you gave the AI a fixed strength. Example: in Race07 when I had the AI to be on par with me in WTCC cars on Valencia it was all off again when I had the same AI-strength set for WTCC on Monza or with Radicals on Valencia. The differences could be huge - not always but sometimes depending on chosen class/track-combo.

So an automatically adapting AI that saves your performance based on class/track-combo and iteratively adjusts AI strength could really help here. But this approach obviously needs several attempts for each class/track-combo to get it right. So if you stay within class and change only the track it will have to learn all new from the beginning. Maybe the automatism takes into account other values gathered from the same class on other tracks, but this does not necessarily fit for the new track, it might even be far off in any direction.

The S3S guys said several times that it takes at least three attempts for the adaptive AI to adapt to your performance. So if I got this right, if you will drive three races with the same class/track-combo, the adaptive AI should be quite fine for this combo and will stay so - for this combo. But to be honest I don't want to drive three races with all classes and on all tracks to have a good AI all over the game. This is why I personally think that an adaptive AI is a good idea but quite worthless if it has to adapt to every class/tack-combo again an again. Best would be if the AI wouldn't be so inconsistent between tracks in the first place.

Disclaimer: Since I don't have any real insight into the AI coding I'm doing only guesswork here - based on years of observation with try-and-error of course. Maybe Marko could enlighten us a bit more how far my guesswork is correct and where I'm wrong.

Sadly nobody was capable to make an AI like in GP4 until today. It was fast, consistent, defensive, agressive but it never crashed into the players car if it was avoidable. If you had an accident you could be sure it was your own fault...

 

[Msportdan]

i agree author and the reason why a individual difficulty per track would be better than this adaptive system ..

http://www.racedepartment.com/threa...dividual-track-difficulty.93859/#post-1826055

 

[Sk3ptik0n]

This effect was already there before offline races were introduced. In free practice you had the adaptive AI already and it was just as you described here. I even encountered this behaviour between two sessions on the same track.

Indeed, I have encountered that a lot in Game Stock Car, but not so much in GTR2 for instance. These days I am racing the Formula V in GSC and some tracks I take pole at 113% then the next track I have to dial it down to 95 to have a chance at staying with the pack, forget the pole.

I also agree with you that having to do 3 full races per track/class for it to recognize your ability seems worthless. I like racing in SP, but that's almost a full time job.

I am sure there are severe drawbacks, but a system that based the AI on your lap time in any session and then adjusted the AI performance as you progress would be more desirable and faster to set. Maybe you could choose at what level you wanted to race: back, mid-pack or front.

I'd actually prefer a system where it assures you always have some cars around to battle and rewarded your driving as it got better. I never liked the Gmotor Packs of cars lap after lap or the qualification where all cars were within 1 second.

Even in R3R I'll start a qualification session and be dead last, then I improve .5 of a second and suddenly I am in the top 5 or even get pole. I'd really like to see some more variation.

 

[Dan Allen]

Well, I've been trying to do the SP races since I last posted about a month ago, and I still think that they are a little bit too aggressive.

The last race I remember doing was:

Corvette Prototype - Laguna Seca - 100% difficulty - Auto-Clutch - Paddle Shifters

They bunch up around me and leave me no room either side, and the slightest tap still sends me flying off the track...

 

[Heppsan]

This effect was already there before offline races were introduced. In free practice you had the adaptive AI already and it was just as you described here. I even encountered this behaviour between two sessions on the same track.

I guess the reason for the introduction of the adaptive AI was that in gMotor sims you always had very different results if you gave the AI a fixed strength. Example: in Race07 when I had the AI to be on par with me in WTCC cars on Valencia it was all off again when I had the same AI-strength set for WTCC on Monza or with Radicals on Valencia. The differences could be huge - not always but sometimes depending on chosen class/track-combo.

So an automatically adapting AI that saves your performance based on class/track-combo and iteratively adjusts AI strength could really help here. But this approach obviously needs several attempts for each class/track-combo to get it right. So if you stay within class and change only the track it will have to learn all new from the beginning. Maybe the automatism takes into account other values gathered from the same class on other tracks, but this does not necessarily fit for the new track, it might even be far off in any direction.

The S3S guys said several times that it takes at least three attempts for the adaptive AI to adapt to your performance. So if I got this right, if you will drive three races with the same class/track-combo, the adaptive AI should be quite fine for this combo and will stay so - for this combo. But to be honest I don't want to drive three races with all classes and on all tracks to have a good AI all over the game. This is why I personally think that an adaptive AI is a good idea but quite worthless if it has to adapt to every class/tack-combo again an again. Best would be if the AI wouldn't be so inconsistent between tracks in the first place.

Disclaimer: Since I don't have any real insight into the AI coding I'm doing only guesswork here - based on years of observation with try-and-error of course. Maybe Marko could enlighten us a bit more how far my guesswork is correct and where I'm wrong.

Sadly nobody was capable to make an AI like in GP4 until today. It was fast, consistent, defensive, agressive but it never crashed into the players car if it was avoidable. If you had an accident you could be sure it was your own fault...

I agree that I need to change the AI strength for different car classes, mostly cause I need more practice in some classes.

But I don't have a problem doing a pen and paper championship with a set strength for all the tracks on the same class.

 

[Sk3ptik0n]

SInce this thread seems to keep going, I wanted to jolt down some notes about my latest experience with the game.


For starters, some AI seem to be unnecessarily aggressive. I cannot count how many times I have been turned around or love tapped from behind and sent crashing against the barrier. While having that happen once in a while would be OK and realistic (minus the part where a slight tap send me spinning like a top) I did a race where it happened several times in a row on T1.

In some tracks/car combo (Zolder/GT3 for sure) there are turns where the AI slows down way too much. Others were they go too fast and a bunch of them pile up against the barrier. The latter I haven't seen in the past couple of weeks, but the former still happens a lot.

In the good news department, last night I did a race at Laguna Seca with the Vintage European Sedans (forgot the name of the class) and the first race was pathetic: I was 8 seconds a lap faster than the runner up. The AI certainly had a very low esteem of my abilities. I started last intentionally and the approach to T1 was like chasing a funeral.
However, the second race was much better and I finished P4 again starting from the back. The top 6 cars were just about my pace. So that tells me it can adjust rather quickly.

In a non AI department, so OT for this post, I did a 14 lap race with the radicals at Mid Ohio and I was very surprised when around lap 10 I started losing the front tires.

I had started in P8 (I set it up manually) with Adaptive AI. This was the 3rd race I did but the first longer than 8 minutes.

At first I started chasing and passing every car and by lap 8 I had gained P1. I then put about 8 seconds between me and P2 when I realized that my car was starting to understeer a lot.

By lap 13, I had 4 or 5 cars on my tail and I had to drive defensively trying to slow them down while I turned that car the way you would a cement truck. Then at the last lap, at the carousel 3 cars passed me because I simply could not turn the car around and my line was all over the place.

I have to say that for a Radical their passes were completely unrealistic. No RL driver would have tapped me and hit me the way they did. Each car would have had about $40K of damage in their body work alone because of that. If it was WTCC it would have been different but a real radical would have crossed the finish in tatters.

 

[Glygoo]

I don't know if this is an AI issue or a track bug but this is what I found. I just had a great AI race at Zandvoort in the WTCC BMW. Just 12 opponents and difficulty at 92% but I'm gradually increasing it until I find my level. I was making some good passes and had a photo finish at the end, really good. On lap 4 I hit a bug. I was just coming out of the tight right/left chicane before the two right handlers that lead onto the main straight and the car in front literally stopped dead like it had hit a brick wall. When I checked the replay there was nothing there but the car in front had sunk into the track, very strange? Has anybody else had this problem? By the way, I turned motion blur on in the graphics settings and it gives a great sense of speed. I'll try and upload a video if I can work out how to do it.

Cheers
Glygoo

 

[Sk3ptik0n]

I have seen similar things. There are some car/track combos where the cars sort of hit the brakes after exiting a turn. If you are chasing them closely, you are guaranteed to plow into them.

On portimao I was having a great race when all of a sudden a bunch of cars decided to slow down to a crawl and then happily kept going and never did it again. I saw them do that with a car blocking the road, nut on the replay there was no car. Obviously, they were all seeing a mirage.

 

[Msportdan]

i have noticed the more cars you run the worst they become, pileups bumping knocking etc..

dunno if thats just me. vo

 

[Dan Allen]

i have noticed the more cars you run the worst they become, pileups bumping knocking etc..

dunno if thats just me. vo

Interesting idea. I may try running against fewer opponents

 

[Sean Kenney]

I still hate how they can brake so quickly compared to me....I ram them at times.:-( ADAC gt cars.

Think the AI is very good overtall

 

[Emilio]

I agree with @Ser_Renely, adaptive AI especially. It delivers a very good challenge and I hope that in the future they'll further improve the way the AI learns new trajectories and keep pushing especially when "you're behind".
Ramming and wrecking happen in the very first lap, especially at the start and during the first turns when cars battle for their spot, but if you play it like you'd do in a MP race I found the AI to be fairly aggressive

 

[JayJayRaceinho]

tried some ai races with various cars at default 100% after new update and they have a pretty fast pace and are very aggressive rear-bumping you. Most races ended in the first lap .... either in the grass or as last place. It's good the AI is fighting for their place at racebeginning so you don't overtake the half grid, but this should be looked at ....sometimes 2 cars overtake you, one from left the other from right, makes me look like a noob .... if the AI is faster then you they tend to rearbump you instead to wait for their "best situation" to overtake or atleast try to be more aggressive on straights and more careful on corners.

...and the Audio 90 GTO needs some liverys, great driving after update, but only 1 livery.

edit : ADAC-GT AI seems lot better, can also easily drive with 105%. Those cars i mentioned above where those with the old physics (saleen,canhard etc.). Still lots of tracks where AI takes corners slow and i rearbump them...