RaceRoom Racing Experience

Force Feedback

• Really quite impressive these days. Might be better than rFactor 1, original Assetto Corsa, and different but comparable in quality to Assetto Corsa Competizione. -Shovas, SimuCube 2 Pro

Adaptive AI

• Overview
  • Adaptive AI adjusts the AI difficulty level for the current car/track combo by comparing the player's average lap over the last 10 laps against AI grid's average fastest lap
• Tips
  • Adaptive AI Training Strategy
    • Run short 10min races, across various tracks, keeping to a single car class, until you are no longer solidly in last place
    • Make sure you are competent at race pace for each car/track you try by practicing first otherwise you won't be helping Adaptive AI adapt
    • After 5-10 races like this, Adaptive AI will begin to improve AI difficulty level so that they begin to feel more competitive relative to you, stop punting you so much, push/pass with more grace, and you should no longer be stuck in last place
    • Adaptive AI will use the average laptimes for untrained tracks based on trained car classes
    • Adaptive AI will use the average laptimes for untrained cars based on trained tracks
    • Adaptive AI will use the average laptimes for untrained cars and untrained tracks based on all other average laptimes for trained cars/tracks (based on my experiments below)
  • Quickly Training the Adaptive AI:
    • Run a number of short 10min races across various tracks but keeping to a single car class across to quickly train the AI
    • Skip practice and qualifying if you already know the track to get through it more quickly
    • Skipping qualifying should be fine as Adaptive AI only trains on Race session lap times
  • Don't confuse car/track unfamiliarity with Adaptive AI
    • Practice until you are competent at race pace
    • I went to a new, unfamiliar car/track and thought the Adaptive AI had reset because they were so dominant but I tried again and practiced until I had competent race pace and found out the Adaptive AI was very competitive.
• Tools
  • The Adaptive AI Primer tool can be used to 'seed' the laptimes data used by Adaptive AI which may help speed up the process of adapting
• Caveats
  • How Adaptive AI was designed to be trained:
    • Run multiple races at the same track
    • Run multiple races with the same car class
    • But, if you're in a new car class or at a new track, it will use average laptimes for the car class or track you've raced before
    • Worst case scenario you're at a new car/track combo and there's no average laptimes for either car or track...
      • Based on an experiment, it might be judging AI level based on existing average laptimes from other car/track combos
      • It's not clear what happens in this case but it's likely it completely breaks as it has no car or track to grab laptimes from to adjust Adaptive AI.
• Quotes
  • From this post
    • "The AAI works by comparing the average fastest lap of the AI grid to the players average lap over the last 10 laps."
    • "So you need to run enough laps for the grid to get spaced out a bit. otherwise they're being held up either by the player or other AI."
    • "4-5 laps is usually enough to get a decent lap-time for the AI."
    • "The player has to run at race-pace during this as well, otherwise the player average will plummet and the whole thing will be for naught."
    • "It will take a few attempts before the AI starts getting up to speed. So do one track/car combo repeatedly until it's up to speed."
    • "Then switch to the next. Rinse and repeat until you have 4-5 tracks done with that class, and by then the law of averages should mean any new track will start in roughly the right place."
    • "Yes, [Adaptive AI] only trains during race-sessions. And yes, you do have to have AI on-track for it to work."
    • "If the AI can't find info on that particular track, it'll take the average of all the tracks you've done with that particular car-class. And vice-versa for new cars on a track you've already run with a different class."
    • "The more tracks you've done, the bigger the chance the adaptive AI will hit the right spot in the first try. Which is where the 4-5 tracks needed come in."
• Review:
• https://forum.kw-studios.com/index.php?threads/adaptive-ai.14262/#post-194212

Adaptive AI Scratchpad

===
240703

R3E Adaptive AI Tracks Run with DTM 2023 Cars, 20 AI, 10min Race 240703:
- No qualifying always grids you 6th. I wonder if that means the goal of AAI is to always give you a chance at a win.
 1. Donington
 2. Mid Ohio
 3. Imola
 4. Brands Hatch Grand Prix
 5. Brno
 6. Zandvoort
 7. Zolder
 8. Hockenheimring
  ^ From being destroyed to comfortably running 10th mid pack, kind of odd
  ^ Maybe there's a minimum laps recorded count before AAI kicks in
  ^ Noticed AI were less aggressive with punting from behind and passed with more care
  ^ Noticing rubberbanding. They never seem to get more than 3-5sec away and you only really catch up if there's an accident.
 9. Hungaroring
  ^ 17th
  ^ AI continue to be less aggressive and pass with more pass
  ^ Similar rubberbanding
10. Interlagos
  ^ 18th. Was last for a while until the AI made some mistakes.
  ^ Similar rubberbanding
11. Monza
  ^ 1st, comfortable 5-3 for a while, worked my way up late, held 1st +0.7sec, almost gave it up
  ^ Forgot to switch car. Hope that didn't mess things up.
12. Moscow
  ^ 1st, Comfortably top 3
  ^ New track to me, nice track though
  ^ There's some weird quirk with isiMotor AI where once you get them to a competitive AI level they seem to run into you less and become more situationally aware of you
  
Conclusions:
- It takes more races/laps than you think while doing it but it's not a crazy amount either
- R3E is underrated. Racing with AI is very decent once dialed in.
  ^ Unpopular but fun cars and tracks are a cool feature of R3E
  
Experiment: Untrained cars at trained track 240703
- Praga R1 (Untrained) at Moscow (Trained)
- Expectations:
  - > If working as expected (ie. has track data), comfortably top 5
  - If working but not as expected, mid pack 10-20th
  - If not working at all, solid last place
- Results: 
  - 5th, comfortably top 5, but AI hot on my tail
  - Maybe some rubberbanding
  - New car, needed to practice to get up to speed
  
Experiment: Untrained cars at untrained track 240703
- WTCR 2022 Honda Civic TCR (Untrained) at Red Bull Ring (Untrained)
- Expectations:
  - If working as expected (ie. no car or track data), solid last place
  - > If working but not as expected (maybe using random track data), top to mid pack
  - If not working at all, solid last place
- Results:
  - 4th, comfortably top 5, rough start and fell to 8th but climbed back up in a few laps
  - Interesting it was apparently able to judge AI level based on average laptimes from other car/tracks because it couldn't have known my laptimes until I raced this car+track combo.
  - New car, needed to practice to get up to speed
  
Next Steps:
- Now that DTM 2023 are trained, how well will AAI apply it to other car sets on the trained tracks?
- Consider the gridding 6th (when no qualifying) for GTR2 AI tweaks
===