AI LMP Cars Very Mistake Prone

[Peter Kerényi]

Hello!

I would like to spend more time with rFactor 2, mostly in single player. I like to set up GTE/LMP2/DPi practice sessions. From my experience, LMP2 cars are extremely prone to making errors, spinning, when close to a slower class. I have AI Mistakes at 0 in the player.json file. I do most of my sessions on official tracks with official cars (came here from Sebring). Aggression about 25, limiter at 5. It is a pretty fresh install of rFactor 2.

I guess enabling AI learning would not solve this, right? As it is more racing related than driving related. Any tips are appreciated... Thanks.

Edit: Decreasing the limiter to 1 seemed to improve it? Have to test it more.

 

[nounoubleu]

I also noticed LMP2 like to spin around GT on certain circuits. Sebring is a good example.
As every year in this season, I spent recently some hours on Daytona. This modded track is not optmized for AI.
I decided to give a chance to Autocalibrate, following method explained in the official forum.
- I selected a list of 43 AI
- in player json, I put autocalibrate at 2
- no fuel consumption, no tire wear, etc
- I let them run some laps (at least 24)
- I clicked on next session, then left.
In player/car/setup, I could see some "wzdom" something files have been generated for each car.
Since then, I have the feeling they're cleaner, LMP really handle properly.

So, perhaps I'm totally wrong but I think it's worth to give it a trial.

PS: I set up aggression at 40 and limiter must be around 10-20

 

[Peter Kerényi]

I also noticed LMP2 like to spin around GT on certain circuits. Sebring is a good example.
As every year in this season, I spent recently some hours on Daytona. This modded track is not optmized for AI.
I decided to give a chance to Autocalibrate, following method explained in the official forum.
- I selected a list of 43 AI
- in player json, I put autocalibrate at 2
- no fuel consumption, no tire wear, etc
- I let them run some laps (at least 24)
- I clicked on next session, then left.
In player/car/setup, I could see some "wzdom" something files have been generated for each car.
Since then, I have the feeling they're cleaner, LMP really handle properly.

So, perhaps I'm totally wrong but I think it's worth to give it a trial.

PS: I set up aggression at 40 and limiter must be around 10-20

It has to be done for every track separately, right? Can I put multiple classes there as well, or only LMP2s and DPi?

 

[MarcG]

I also noticed LMP2 like to spin around GT on certain circuits. Sebring is a good example.
As every year in this season, I spent recently some hours on Daytona. This modded track is not optmized for AI.
I decided to give a chance to Autocalibrate, following method explained in the official forum.
- I selected a list of 43 AI
- in player json, I put autocalibrate at 2
- no fuel consumption, no tire wear, etc
- I let them run some laps (at least 24)
- I clicked on next session, then left.
In player/car/setup, I could see some "wzdom" something files have been generated for each car.
Since then, I have the feeling they're cleaner, LMP really handle properly.

So, perhaps I'm totally wrong but I think it's worth to give it a trial.

PS: I set up aggression at 40 and limiter must be around 10-20

FYI for Auto-Calibrate="2" you need to let them do a couple Hundred Laps for it to be of any benefit, 24 is nowhere near enough. What they do at that number is to Pit and then go out again, do another 24 and rinse & repeat. This Builds up their "wisdom" so that after 200+ Laps they've got a "memory" with which to take from

 

[MarcG]

It has to be done for every track separately, right? Can I put multiple classes there as well, or only LMP2s and DPi?

yes it has to be done separately for each track, personally I would keep the same Car Class but if you're stuck for time you could include other Classes but they may interact too much and potentially cause "mis-learnings" but I cannot confirm that as I've never done it.

The Steps here should still work although I've not tried with the new UI:

* AI Learning - Step by Step Guide *

I have created this Guide for the JSON entry of "Autocalibrate", this basically helps the AI learn a given track and save their Learning Process for...

forum.studio-397.com

 

[Peter Kerényi]

yes it has to be done separately for each track, personally I would keep the same Car Class but if you're stuck for time you could include other Classes but they may interact too much and potentially cause "mis-learnings" but I cannot confirm that as I've never done it.

The Steps here should still work although I've not tried with the new UI:

* AI Learning - Step by Step Guide *

I have created this Guide for the JSON entry of "Autocalibrate", this basically helps the AI learn a given track and save their Learning Process for...

forum.studio-397.com

Seems like a lot of downtime to be honest. Too bad we don't have a "repo" for it, it could be easily shared (or better, the developer could make it in the first place). I guess you have a lot of experience with it, or at least more than most people here. In your experience, does it help with multi-class racing?

 

[MarcG]

Seems like a lot of downtime to be honest. Too bad we don't have a "repo" for it, it could be easily shared (or better, the developer could make it in the first place). I guess you have a lot of experience with it, or at least more than most people here. In your experience, does it help with multi-class racing?

I've only ever used AI Learning if the AI have trouble with a certain corner or section of track, regardless of Class Racing it could help improve the AI in general on that Track.
The thing with AI Learning is it's not going to fix those badly designed AIW tracks with poor FAST Paths, so it really is only about helping the AI to "tweak" their lap for it to get something close to what you do and if they still struggle then you're stuck without editing the AIW yourself.

Yeah it can take a little bit of time, handily Time Accelerate (as I mention in the Guide) does speed up the process somewhat especially for the "2" Learning. Handily most Official Tracks come with quality AIW Paths so generally speaking AI Learning is not needed for them (thereare some exceptions from time to time). Mark Pearce over on the Official Forums has done a lot of good work improving various tracks AIW Paths and making them available, so again there shouldn't be the need to do any AI Learning on those tracks. Trouble is when there's so many Car & Track combinations it's very hard to create an AIW Path that suits all types of Cars, you can go down the rabbit hole of doing individual Paths but then we're talking weeks of work!

The Learning INI & WIS files are all shareable but there's never been a Resource (AFAIK) for collecting all them all, which is a shame as I said there's so many different Car & Track combos it does get quite daunting. Back in RF1 days I did a complete V8Supercar Season editing and tweaking AIW on all the Calenders Tracks, loads of AI Learning, it took absolutely ages but was worth it in the end!

 

[nounoubleu]

Seems like a lot of downtime to be honest. Too bad we don't have a "repo" for it, it could be easily shared (or better, the developer could make it in the first place). I guess you have a lot of experience with it, or at least more than most people here. In your experience, does it help with multi-class racing?

I put everybobdy , DPi, LMP2, LMP3, GTE and GT3, I let them run while I was doing something else.
Maybe they run in total 2 or 3 sessions of 24laps and I noticed a significant improvement.
So, for a given track it's a limited down time and useful investment.
On this Daytona, I always had the feeling they were runing on one line, LMP2 were always hesitating to pass GT on the banking but were loving to ram me in corners. Now they look more alive and more smart.
However, after reading posts above, I realize it was probably not enough

 

[Peter Kerényi]

I've only ever used AI Learning if the AI have trouble with a certain corner or section of track, regardless of Class Racing it could help improve the AI in general on that Track.
The thing with AI Learning is it's not going to fix those badly designed AIW tracks with poor FAST Paths, so it really is only about helping the AI to "tweak" their lap for it to get something close to what you do and if they still struggle then you're stuck without editing the AIW yourself.

Yeah it can take a little bit of time, handily Time Accelerate (as I mention in the Guide) does speed up the process somewhat especially for the "2" Learning. Handily most Official Tracks come with quality AIW Paths so generally speaking AI Learning is not needed for them (thereare some exceptions from time to time). Mark Pearce over on the Official Forums has done a lot of good work improving various tracks AIW Paths and making them available, so again there shouldn't be the need to do any AI Learning on those tracks. Trouble is when there's so many Car & Track combinations it's very hard to create an AIW Path that suits all types of Cars, you can go down the rabbit hole of doing individual Paths but then we're talking weeks of work!

The Learning INI & WIS files are all shareable but there's never been a Resource (AFAIK) for collecting all them all, which is a shame as I said there's so many different Car & Track combos it does get quite daunting. Back in RF1 days I did a complete V8Supercar Season editing and tweaking AIW on all the Calenders Tracks, loads of AI Learning, it took absolutely ages but was worth it in the end!

In that case it seems like it would be useless for me. My only problem is that AI spin a lot when close to each other, especially prototypes. I have this problem on Sebring, Portland. Nurburgring GP is kind of okay. Road Atlanta, while one of the best looking mod tracks, is absolutely terrible from this perspective.

 

[MarcG]

In that case it seems like it would be useless for me. My only problem is that AI spin a lot when close to each other, especially prototypes. I have this problem on Sebring, Portland. Nurburgring GP is kind of okay. Road Atlanta, while one of the best looking mod tracks, is absolutely terrible from this perspective.

I've not noticed any issues on Sebring, another "trick" is to backup and move your player.json then let the Sim create a fresh one. This has been known to strangely "fix" AI issues in the past, but I've never been able to find out why exactly.

Also a side note, the AI need Rubber, if you're racing them on a Green Track they will slip & slide just like a Human driver would so this could exaggerate their spinning.

 

[Peter Kerényi]

I've not noticed any issues on Sebring, another "trick" is to backup and move your player.json then let the Sim create a fresh one. This has been known to strangely "fix" AI issues in the past, but I've never been able to find out why exactly.

Also a side note, the AI need Rubber, if you're racing them on a Green Track they will slip & slide just like a Human driver would so this could exaggerate their spinning.

I literally reinstalled rFactor 2 like... 4-5 days ago. I will check what the track state is/was. In Sebring they tend to spin around e.g. turn 11 trying to overtake the GT cars around the outside, clipping the grass.

 

[MarcG]

I literally reinstalled rFactor 2 like... 4-5 days ago. I will check what the track state is/was. In Sebring they tend to spin around e.g. turn 11 trying to overtake the GT cars around the outside, clipping the grass.

Could try lowering the Aggression, that may help them wait a little longer for a better passing opportunity.

 

[Peter Kerényi]

@MarcG Quick video about an AI race I did at Sebring. Clumsy. Aggression at 13%, Limiter at 0%. I was in the first 2 clips in the Ligier, white Porsche is included because Mr. Crew Chief mentioned him a lot so I checked him in the replay.

Edit: So basically it is often not even the LMP2 passing aggressively, but the GT acting strangely, going out of the bounds of the track, causing trouble. In this case maybe that calibration would actually help? Still, strange behaviour during rejoining, that needs to be tweaked by the devs for sure. And however I would like to solve it and I will try to solve it if possible, I find this stuff kind of unacceptable for paid content. It should work well out of the box.

 

[MarcG]

@MarcG Quick video about an AI race I did at Sebring. Clumsy. Aggression at 13%, Limiter at 0%. I was in the first 2 clips in the Ligier, white Porsche is included because Mr. Crew Chief mentioned him a lot so I checked him in the replay.

Edit: So basically it is often not even the LMP2 passing aggressively, but the GT acting strangely, going out of the bounds of the track, causing trouble. In this case maybe that calibration would actually help? Still, strange behaviour during rejoining, that needs to be tweaked by the devs for sure. And however I would like to solve it and I will try to solve it if possible, I find this stuff kind of unacceptable for paid content. It should work well out of the box.

Ok I'll go through the clips 1-by-1 but I'll be playing the role of "Race Steward";

1. Can't tell anything from that angle, it's a Spin/Crash - this happens in Real Life.
2. The Porsche runs wide and clearly gets two wheels onto the dirt, this destabilises him and he spins - I see this in Real Life all the time.
3. Ok what the AI do when being Lapped is to move out of the way/off the Racing Line & come off the throttle slightly to allow you to come past (this is good), he clearly gets onto the dirt which IMO is 50/50 Real Life & Sim Issue (which is bad). He shouldn't of gone that far over of course, but in order to "fix" this it could be a mix of AIW and/or Code changes which ain't gonna happen anytime soon. Personally I'd just put it down to a Racing Incident.
   As for the Rejoining, the code is there whereby he's waiting for a Gap, but this Code does need tweaking I agree, it works better in some situations than others.
4. Yeah that looks like an AIW issue with the Path too close to the tire barrier, under normal circumstances they're just fine at that point but like above with the Blue Flag he's trying to get out the way, obviously doesn't work well in that section.
5. Like #1 can't tell anything from that angle, seen crashes at that corner in Real Life.
6. Ok again a case of drifting too wide to allow faster cars to come through, I mean I've seen this in Real Life whereby the backmarker gets out the way too far and ends up spinning. In this instance it's not ideal, however tweaking of the Blue Flag Code and/or AIW Paths could help prevent this.
   

I think what you need to accept is two things here; 1) The AI will make mistakes - just like Real Life - Don't look at every single AI Incident as a "Bug". 2) Don't take them too seriously, once you get used to their quirks they're great fun to race against and IMO a lot cleaner than a lot of Online Lobbies!

For sure the AI in RF2 need a bit of work we all know that (those of us that have been here for years at least!), S397 have said in the past that they will look into AI one day and we all live in hope that day comes soon. In the meantime check out other Sims AI, iRacing/ACC/AMS2/RaceRoom all have strange moments of AI behaviour, none of them are Perfect in any sense of the word and each is better than the others in another way (RaceRoom being my personal fave).

So take the AI for what they are, in certain situations they may be rough, in others they may be great. Try that Combo at Spa, funnily enough I raced the GT3 vs LMP2 there today and it was great fun. I then did the same in AMS2 and (unfortunately) the racing was very "meh", see my point about about no perfect AI

 

[Peter Kerényi]

Ok I'll go through the clips 1-by-1 but I'll be playing the role of "Race Steward";

1. Can't tell anything from that angle, it's a Spin/Crash - this happens in Real Life.
2. The Porsche runs wide and clearly gets two wheels onto the dirt, this destabilises him and he spins - I see this in Real Life all the time.
3. Ok what the AI do when being Lapped is to move out of the way/off the Racing Line & come off the throttle slightly to allow you to come past (this is good), he clearly gets onto the dirt which IMO is 50/50 Real Life & Sim Issue (which is bad). He shouldn't of gone that far over of course, but in order to "fix" this it could be a mix of AIW and/or Code changes which ain't gonna happen anytime soon. Personally I'd just put it down to a Racing Incident.
   As for the Rejoining, the code is there whereby he's waiting for a Gap, but this Code does need tweaking I agree, it works better in some situations than others.
4. Yeah that looks like an AIW issue with the Path too close to the tire barrier, under normal circumstances they're just fine at that point but like above with the Blue Flag he's trying to get out the way, obviously doesn't work well in that section.
5. Like #1 can't tell anything from that angle, seen crashes at that corner in Real Life.
6. Ok again a case of drifting too wide to allow faster cars to come through, I mean I've seen this in Real Life whereby the backmarker gets out the way too far and ends up spinning. In this instance it's not ideal, however tweaking of the Blue Flag Code and/or AIW Paths could help prevent this.

I think what you need to accept is two things here; 1) The AI will make mistakes - just like Real Life - Don't look at every single AI Incident as a "Bug". 2) Don't take them too seriously, once you get used to their quirks they're great fun to race against and IMO a lot cleaner than a lot of Online Lobbies!

For sure the AI in RF2 need a bit of work we all know that (those of us that have been here for years at least!), S397 have said in the past that they will look into AI one day and we all live in hope that day comes soon. In the meantime check out other Sims AI, iRacing/ACC/AMS2/RaceRoom all have strange moments of AI behaviour, none of them are Perfect in any sense of the word and each is better than the others in another way (RaceRoom being my personal fave).

So take the AI for what they are, in certain situations they may be rough, in others they may be great. Try that Combo at Spa, funnily enough I raced the GT3 vs LMP2 there today and it was great fun. I then did the same in AMS2 and (unfortunately) the racing was very "meh", see my point about about no perfect AI

Yes, I understand mistakes happen IRL, and it is good to have mistakes in sims. These were just a couple of random incidents. When we start lapping I hear the Crew Chief saying yellow flag in sector X every 10-20s, with me gaining a lot of positions, it is not realistic anymore. I am all in for immersion, and this breaks it a lot. Just did a race in ACC, single class, it wasn't that spectacular, but it made me kind of believe I am there. I saw one mistake from the AI in front of me, AI rejoined safely. Slight problem is that I really want to race modern prototypes in multiclass... And almost no sim provides that. rF2 has the content, but the AI is just silly. AC has no good quality LMP2s and the AI is hit or miss. Honestly, Project CARS 2 is the closest, sadly. RaceRoom multiclass AI is even worse, faster class goes through very slowly. AMS2 AI is improving, but provides no soul and racing yet. :/

 

[MarcG]

Try a different Track like Spa as I mentioned, I get what you're saying I know how frustrating AI can be and RF2 currently can be very hit 'n miss in certain combinations, but when it Hits it's great against them and currently I'd say much better in Single Class races than Multi Class.

 

[Peter Kerényi]

Try a different Track like Spa as I mentioned, I get what you're saying I know how frustrating AI can be and RF2 currently can be very hit 'n miss in certain combinations, but when it Hits it's great against them and currently I'd say much better in Single Class races than Multi Class.

Will do. I guess if there's nothing to tweak to improve this I can only hope Studio 397 will work on it. Anyway, thanks for spending time discussing

 

[Durge Driven]

Did 5 10 lap sprints after letting 36 cars qualify in accelerated mode
One proto finished down list
In races was 2/3 incidents a race
in 5 races no car failed to finish and all Proto lapped GTE once

In 2019 ( 33 cars ) 5 Proto failed to finish and every real race has spinners
I don't know what people expect if you want no spinners is fantasy

It is ISI engine not S397 they have done nothing to improve AI since then
ISI always said that AI was not important on their list of things

edit: The trouble is Proto lapped 4-5 seconds slower then real while GTE did 1:58 / 1:59 which is spot on

115% and 40% agro

Like I said before lowering AI does not mean they will be better but raising it may actually let cars pass cleaner as higher aggression can make them more decisive

 

[Durge Driven]

50 laps, GTE finished 5 laps down so Proto needs to be faster right ?
All finished including fuel stops

edit: was about 10 yellow
make sure you turn off flag rules, it won't affect blue flags GTE will still get out of way

BTW
I also had many GTE put 2 wheels in dirt
Maybe lower strength and agro causes them to spin easier because
they jump off throttle too fast

After thought, if you never knew you can save races and resume from replay
Good for endurance, pity other studios can't

Steam Community :: Guide :: rFactor 2 Quick Guide: Using "Resume From Replay" as a Save Game

This Quick Guide shows how to use “Resume From Replay” as a save game in rFactor 2. This allows you to stop a session (in Practice, Qualifying, or Race) and resume your session at a later time....

steamcommunity.com

 

[Peter Kerényi]

Thanks @Durge Driven for the advice, I will try again with increased aggression. Strength is something around 105 for me at Sebring. I turned flag rules to black only already because in the first try we had FCY after turn 1 already (an LMP2 failed to qualify, probably causing trouble at the start). What was your AI limiter setting during your testing?