What is smartest way to fix Howston G4 AI issues?

[Kevin Knorpp]

Edit - The following re-post from my blog is my new approach to this problem. For reference purposes, the original post has been preserved in italics at the bottom of this post.

I have decided to concentrate first on tracks which the G4 AI are reasonably able to navigate and address slowness via:

1. Driving a slower (depending on track) version of the G4 vs the AI in a faster version
2. Applying a fuel ballast to my car (more realistic than > 100% AI strength)
3. Increase AI strength
4. Set fuel = OFF to extend (if it works how I'd think) my ballast handicap vs AI
5. If all else fails, race against a different car (GTO) that is faster or has better AI

Tracks currently testing include:

• Lester - Had decent race at 115% but was still faster than AI. Next steps:
  • Save 1.2 version of track to MediaFire as was hard to find
  • Install WooChoo's skins as skins are being reused
  • Test at 117% strength
  • Test at 100% against '69 using '67 G4
  • Test at 100% against '69 using '68 G4
  • Test all of the above using fuel ballast to see how close can get to AI = 100%

Tracks to be tested include:

1. Nords - Start with 105%
2. Bugatti - Start with 95%
3. Portugal - Start with 100%
4. Sebring - Start with 105%
5. Monaco - Start with 105%
6. Mid-OH - Start with 115%
   
7. Toban - Start with 115%
8. Loch Drummond - Don't remember ; Use whatever average is after the above
9. Malaysia - Was third fastest at 100% so try 105%
10. Texas World Speedway - Try 105%
11. Lienz - Start with 105%
12. Mountain Forest - Start with 105%
13. Essington - Start with 105%

For tracks that are truly worth the testing effort and which the G4 AI are not able to navigate (Mills, Atlanta Motorsports Park), drive against a reasonably-matched car which the AI are able to navigate around the track.

Original Post for reference only:

Howston G4 AI struggle terribly and/or are very slow at several tracks, such as AMP, Mills and Toban.

What is the smartest way to fix these? For example:

• Has anyone already fixed?
  • For example, I found a car set by Whoo Choo. Does this fix anything with AI? Has anyone else fixed this already?
• What should I do first?
  • Look for something specific in the files for the vehicle, i.e. VEH or RCD problem?
    • If memory serves, the T70 Spyder mod also struggles at Toban running into the wall in the turn before last so I'm thinking is either not a vehicle issue or was not fixed by that mod.
  • Try to train the AI? And which type of training first?
  • Force AI setup at track? I think I can find how to then make default setup if I can get one to work.
    • And any tips for how to approach testing to get fastest AI setup? I.e. Try your setup first, what affects AI vs what doesn't, etc.

Anyway, you get the idea. Anything you can think of that will help me solve this faster and with less pain would be much appreciated. TIA

 

[Keith Dixon]

Be nice to see the car and AI get some official updates, I've been trying for ages to get a decent race out of it around old Spa and just gave up, still love driving it but it's just a hotlap machine for now

 

[Emery]

A few things going on in my experience:
1) AI, in general, doesn't know how to deal with an understeering or sliding car.
2) AIW for many tracks was created with a modern car, thus the racing line is too close to track edge and kerbs for vintage cars that slide around. This also affects cars with high CG, so they're prone to roll over when hitting kerbs.
3) AI needs a setup for aero & gearing on certain tracks.
4) AI struggles on green tracks.
5) The AI Grip Multiplier is too low in the TBC. If they can't hold your speed in corners with a similar setup, then the mod is weak.

So, for the G4, I'd look at these things:
a) If they're running wide, assign a setup that provides more front grip.
b) Consider making an AIW for the track if Type 2 learning does not help. Make use of time acceleration.
c) Assign a setup with higher downforce might help the G4 on some tracks, provided the model you're using has downforce to adjust.
d) Make sure the track has some rubber. Make use of time acceleration.
e) Repackage the mod with a higher AI Grip Multiplier. I don't like to see it above 1.10; the historic F3, for instance, needs to be up at 1.08 to achieve cornering speeds that a good player makes in corners which, apart from gearing, is why they're lousy at 1966 Belgium. Since you're already tinkering with sounds, I know this isn't out of your capability

Woochoo's car set and CrossPly's T70 Spyder do not adjust the Howston's AI behavior as far as I know, though I think CrossPly included a talent file.

 

[Kevin Knorpp]

@Emery, thanks again, man! Yep, I should be able to do all of these things

Only one I'm not sure of is this one:

b) Consider making an AIW for the track if Type 2 learning does not help. Make use of time acceleration.

I have never made an AIW and can look into that but how is time acceleration involved in creating an AIW. I thought AIW was where you have to go into developer mode and set corridors and maybe drive a path, etc. but how does time acceleration play into it? I'll probably find out once I look into it but was curious because it sounds like something automated if time acceleration applies which sounds great if that is the case somehow? Also, do you know of any start to finish videos for how to do AIW or does this get into figuring it out / black magic / experience kinda thing before it becomes bearable?

Many thanks again, my friend

 

[Ace King]

As Emery stated, a new specific AI fast path (aiw) for troubled car mods will solve many of these issues. This new aiw specific for the troubled car mod will provide necessary braking points and the fast path needed for the ai to race competitively!

Also, the car mod's ai stability settings might be the issue and need some tender loving care!

As an example the ai in the GT3 World's Series mod by Simtek were not competitive on most tracks so i needed to fix the ai specific settings in the car mod. For those who don't have the time or expertise you can wait for Simtek's next update since they acknowledge the ai issues and state they will fix them.

 

[Ace King]

Kevin....here is a decent aiw creation tutorial. Have fun!
http://imagespaceinc.com/downloads/rf2/modsresources/rF2_AIW_Creation_Tutorial.pdf

AI Learning - Step by Step Guide:
https://forum.studio-397.com/index.php?threads/ai-learning-step-by-step-guide.47657/

 

[Emery]

Oh, time acceleration was for the AI learning. Yeah, bad phrasing on my part.

 

[Kevin Knorpp]

Got it. Thanks, @Emery and @Ace King

AIW looks way too convoluted for me at this point in my life but the rest of it may make for a fun project, provided I maintain the right attitude and patience

I'll keep you guys posted in case you're interested and also in (faint) hopes that someone sees this and says "Stop! I've already fixed all this stuff and have new AIWs, etc. for track a, b, c..."

 

[Deleted member 130869]

These things should not be up for the end user to have to take arms and fix, but good luck K².

 

[Kevin Knorpp]

I agree but it is what it is. Only thing that rubs me the wrong way is when a title claims to have a lot of content but doesn't support it. You can't have it both ways. You either claim your content or you don't.

For example, I'd like to know what content @Marcel Offermans intends to claim as part of rF2 official content. Until he does that, rF2 will always be a "mixed bag" as everyone says.

So far, I guess all he can claim is USF2000 and future cars. By that standard, rF2 has only 1 supported car. Pretty light on content, I'd say, and even then IMO Marcel can only claim tracks which the USF2000 AI work on if he wants rF2 to ever be anything like AMS, for example, where people can just race because Reiza takes ownership and pride in quality and doesn't want sim racers to have to fiddle and in fact discourages community members from recommending file changes to other members, etc.

 

[Spinelli]

I don't mean to barge in here and start an argument or fight between RF2 and AM but, Kevin, AM isn't all perfect either. For eg.: even @ 120% AI speed, the F3s (F309, can't remember about the F301) way over-slow for the final corner @ Brasilia short - not nearly enough mid corner speed, take a bad line (especially apex to exit) at the final corner, and over-slow with too low entry-speed for turn 1. I mean it's massive, like you ram into them if you don't know about it (especially the last corner).

AI track/car-combo-consistency aside, the RF1/AM AI is not nearly as dynamic as RF2. RF2 has the best AI in the industry in my opinion. You can really tell it's down to core-engine AI updates rather than just plugging in different numbers to RF1/AM. I've seen tons and tons of AI mods (plugging in different AI numbers into PLR, AI files, etc.) for RF1 and, yes, it does some fine-tuning to change/improve certain things but you can still tell it's "RF1" AI (just like, no matter what number-plugging you do to a car's physics files, it's only "fine-tuning" unless you update the actual core physics engine).

In my opinion the problem is this: The way the AI works in the game - regardless if it's RF2, AM, RF1, etc. - is that a guy has to manually tell the game the line to take around a track. So either all cars use the same line (bad) or all cars have different lines but then the developers have to create a different line for every car for every track (tons of work - bad). The fix would be a complete revamp of the entire core-AI "engine" and have them somehow dynamically learn the track. Maybe use a separate CPU core of the PC (the game only really uses 1 or 2 cores) and devote it to this AI learning. RF1/RF2/AM does have a basic AI learning option but you have to manually set it in a text file and I heard it's not always perfect and doesn't solve all issues. I believe it will work for smaller problems (like my AM F309 example above) but, if I remember correctly, there are limitations to running it (only 1 car can use it? only practice/test sessions?...not sure).

The "core AI engine" should get a "revolution" rather than an "evolution". In other words, it should get started from scratch rather than just updated but ISI unfortunately just keep on with their engine updating back to the 90s and updating it (which is why parts of the physics, the sound, the AI, etc. still have - at moments - similarities to all previous ISI games from the 1990s with Sports Car GT through to today with RF2 and AM).

 

[Ace King]

Been doing AiW work since 2005 (rF1) to present (rF2) and i agree with Spinelli that in my opinion rF2's AI is no doubt the very best of all the past and present sims to date!!

It's not possible in this present time to have a "universal" AI that would be equally fast and talented on all tracks simply because car physics, braking points and fast lines are not equally the same with all car mods. As a solution for rF1 i made track packs with new fine-tuned AIW's (fast lines) specific for each of my favorite car mods.

In rF2 they improved the AIW editor to allow you to make several "alternate fast lines" in a track's AIW for specific car mods that were troublesome. The AI will now use this alternate racing line when they are driving the car in question. This also saves HD space clutter compared with the older methods (track packs) though there are some drawbacks in rF2 which limit the number of alternate fast lines.

 

[Emery]

Ovals are where rF2 AI are currently at their worst. ISI described it as a work in progress and Studio 397 is likely going to improve it in the upcoming release or the following one. I haven't touched oval AI since they first appeared because it was so awful, but as I recall the problems mostly revolved around their ability to obey flag rules and a little bit around passing.

Some of the passing problem also affects roadracing since the AI don't know how to use the draft with intelligence. Drive Brianza '66 with a large historic F3 field and you'll understand. I suspect if drafting is solved for ovals it will also be solved for roadracing.

 

[Kevin Knorpp]

I agree but it is what it is. Only thing that rubs me the wrong way is when a title claims to have a lot of content but doesn't support it. You can't have it both ways. You either claim your content or you don't.

For example, I'd like to know what content @Marcel Offermans intends to claim as part of rF2 official content. Until he does that, rF2 will always be a "mixed bag" as everyone says.

So far, I guess all he can claim is USF2000 and future cars. By that standard, rF2 has only 1 supported car. Pretty light on content, I'd say, and even then IMO Marcel can only claim tracks which the USF2000 AI work on if he wants rF2 to ever be anything like AMS, for example, where people can just race because Reiza takes ownership and pride in quality and doesn't want sim racers to have to fiddle and in fact discourages community members from recommending file changes to other members, etc.

I'll just quote myself to add to / clarify it in case you guys were responding to this. If not, then disregard.

There are people in the community that understand the nuances of rFactor 2 yet paint it with a broad brush by saying things "like rF2 has the best AI".

rF2 doesn't have any AI. Tracks and cars have AI. rF2 is just an engine and without quality content the AI is crap.

AI is crap if people don't know how to setup the SW.

AI is crap if people "drive the wrong combos".

AI is crap if vehicle and talent files are missing, incorrect or don't match.

Why so much emphasis on the sim? I only evaluate cars and tracks. If a sim does not enable a the type of racing and quality of AI that I need with a combo then the combo is crap.

There are many combos that work for my purposes in AMS.

There are combos that I think I can get to work for me in rF2 with enough patience and effort.

Very few, if any, (and admittedly unorthodox ; although that is a reason many decided to drive offline) combos that I am interested work out of the box.

People (not here) can get so touchy whenever anyone questions the superiority and perfection of rF2 when it's not about rF2, its about any given combo especially unless / until S397 decides to do something about it.

Reiza clearly tries to make the best of the rF1 AI and make it consistent and their tastes for what is fun happens to match up to what I think is fun also. But yeah of course their AI is limited in what is capable but the extent to which AMS pushes it is much farther and more consistent.

 

[JoeNathanYouTube]

Kevin, you can't say that people are painting rF2's AI with a broad brush and then later do the exact same thing about Reiza. You even did it in the same exact post.

And in my honest opinion, rF2's default content AI is better than what Reiza has done.

 

[William Wester]

Yes, the rF2 AI is not perfect for every track and car combo - it's not automatic, not a given. With that said, for me, the overall experience of AI racing in rF2 is much better than any other sim I race. I have to believe it's easier to create good AIW with the rF2 engine than some of the other sims. As an example, it must be more difficult to create good AIW with the AC engine as, for me, I've run info a lot of mods with terrible AI. That last statement doesn't mean anything as I personally don't know what it takes, just my experience with mods in the sims I race.

 

[Hash_V8]

Ive commented on your vids, all of those vids (if thats you) you are running on a 100 per cent green track, I would almost say it was a mistake on ISI part to make 100per cent green defualt, but most tracks have presets if you dont want to bother making your own, how I said on your vids.

 

[Kevin Knorpp]

Kevin, you can't say that people are painting rF2's AI with a broad brush and then later do the exact same thing about Reiza. You even did it in the same exact post.

And in my honest opinion, rF2's default content AI is better than what Reiza has done.

Of course I can because consistent quality is one of the most common, if not the most common, complaints about rF2 and one of the most common, if not the most common, compliments of AMS.

Ive commented on your vids, all of those vids (if thats you) you are running on a 100 per cent green track, I would almost say it was a mistake on ISI part to make 100per cent green defualt, but most tracks have presets if you dont want to bother making your own, how I said on your vids.

Why do you keep saying this? I ignored your YouTube comment because everyone can see that I show you the settings I use in the video which is medium rubber every time. Can you not see that? I literally walk through every setting I have every time and my settings include medium rubber

 

[JoeNathanYouTube]

So its okay to be a hypocrite? AMS AI is absolutely broken at Turn One at Interlagos. I can actually upload a video of broken AI from AMS offline racing. It's not very pretty. It's not that "polished".

I mean, it's not the first time you've contradicted yourself in the same post/video, so I suppose I'll leave it at that...

 

[Hash_V8]

Of course I can because consistent quality is one of the most common, if not the most common, complaints about rF2 and one of the most common, if not the most common, compliments of AMS.



Why do you keep saying this? I ignored your YouTube comment because everyone can see that I show you the settings I use in the video which is medium rubber every time. Can you not see that? I literally walk through every setting I have every time and my settings include medium rubber

sorry missed the settings, but visually on your vids it looks green, spotless, doesn't look like medium at all, just watched your number 4 AIlol vid, and TBH im still struggling to see your problem with AI, it aint perfect, but its far from what you are describing it as,medium settings seem more like my light if that, put real road on 5x, be sure to qualify, this helps sort AI out too, especially if they have talent files, and if having cars run wide on circuit normally means to higher aggression setting, try 25 or even lower.