Fixed vs Adaptive AI

Discussion in 'General Discussion' started by stlutz, Sep 5, 2021.

  1. vvPULSEvv

    vvPULSEvv Member

    Joined:
    Mar 28, 2020
    Ratings:
    +7 / 0 / -0
    @anno900
    I use default setups... however, for qualifying I do remove fuel, but that is it.
    I do drive consistent laps, as Crewchief reminds me of that quite often. That said, there are AI incidents that slow me down, and I'll occasionally get invalidated laps as well. I have no clue has RR factors these laps into your 10-lap average.
    What's good is that it is dynamically changing based on your last 10 laps, so with that, if you have 10 consistent, fast laps, then I can see the following race having you mid-pack... However, in most races, I don't get 10 straight optimal laps due to race traffic, ai wrecks, etc... so perhaps with those "bad" laps factored in with the good laps, I'm always competitive? My seasons consist of 2 races per track, the first one 10 minutes +1, and the 2nd 20 minutes+1.
    Hope this helps.
     
  2. Maskerader

    Maskerader Well-Known Member

    Joined:
    Oct 6, 2019
    Ratings:
    +352 / 0 / -0
    Basically, you're penalized for getting better and rewarded for being inconsistent. Right?
     
  3. vvPULSEvv

    vvPULSEvv Member

    Joined:
    Mar 28, 2020
    Ratings:
    +7 / 0 / -0
    @Maskerader

    I suppose you could put it that way, but I do enjoy the end result, as I don't want to win every race, and I am competitive (top 5) probably 70% of the time, and top 10 90% (20-30 car field). That's a ballpark average.

    I'm curious if there's any variables/ranges to the AI as well. For instance, your current AAI level at a particular car/track combo is 100%. Within that 100%, is there a random (mini) range that the cars are assigned? I guess if so, it would be the same with fixed ai setting as well. I haven't run fixed, so I wouldn't know. For those that do run fixed, does speed/lap times vary slightly each race, or can you bank on the fastest lap times of the AI being identical each race? Just curious.
     
  4. Christian G

    Christian G Topological Agitator Beta tester

    Joined:
    Apr 8, 2015
    Ratings:
    +2,411 / 0 / -0
    There's variety in two dimensions. Each AI driver has his individual talent file which determines their strength through a combination of traits. In general this creates a range of AI lap times that spread around the user's (or the fixed) pace at about +/-0.5 seconds.

    In addition to that there's a randomness variable that acts as a form-on-the-day modifier. That means even the same AI opponent won't perform the same when repeating the same race several times.
     
    • Informative Informative x 3
  5. vvPULSEvv

    vvPULSEvv Member

    Joined:
    Mar 28, 2020
    Ratings:
    +7 / 0 / -0
    • Like Like x 1
  6. Muzarati

    Muzarati Member

    Joined:
    Jan 27, 2018
    Ratings:
    +20 / 0 / -0
    +0.5 seconds I could probably live with, but why are we getting +1-1.5 seconds? But the fact that the AAI will go faster is still the flaw, like a previous post said, you find the sweet spot and max yourself out, they're always gona go faster which defeats the whole purpose of being adaptive!
     
  7. Christian G

    Christian G Topological Agitator Beta tester

    Joined:
    Apr 8, 2015
    Ratings:
    +2,411 / 0 / -0
    We are aware that bigger spreads do occur, which is why I said the above applies 'in general' and 'about'. If you find class-track combinations where the spread is very large, you're welcome to let us know and we can try to narrow it.

    Regarding the second point: The general idea behind AAI is to create an AI that's as close to the players pace as possible while still providing a challenge. It's supposed to push you to improve your driving. This can indeed lead to a situation like you described (even tho technically the progression isn't infinite, it will cap out at the maximum AI strength, but I'm aware almost nobody will reach that point, including myself).

    We could probably change that so the AI adapts to your pace and then takes one step back, so to say, but then you'd end up with an AI you will almost always beat if you're consistent enough, which also isn't an ideal outcome.
     
    • Like Like x 1
  8. vvPULSEvv

    vvPULSEvv Member

    Joined:
    Mar 28, 2020
    Ratings:
    +7 / 0 / -0
    @Christian G

    You have the best AI in all of Sim Racing, and the AAI is icing on the cake. That's primarily what makes RaceRoom my go-to sim.
     
    • Like Like x 1
    • Agree Agree x 1
  9. Nico Kunze

    Nico Kunze Well-Known Member

    Joined:
    Sep 1, 2018
    Ratings:
    +395 / 0 / -0
    Should those reports be regarding race pace or qualifying or both? If its quali then how many laps should one wait to give every ai driver the chance to get a clean lap in?
     
  10. Hendrick

    Hendrick Well-Known Member

    Joined:
    Oct 7, 2020
    Ratings:
    +64 / 0 / -0
    Yes same track car combo
    Red Bull Ring Spielberg
    GTR3/Porsche 911 R
    Starting position 11 of 32
    15min races
    They now doing lap times far beyond my personal best. I do not know what formula/algorithm is in place for AI strength but after a few experiments I'm going back to a fixed AI.
     
  11. Maskerader

    Maskerader Well-Known Member

    Joined:
    Oct 6, 2019
    Ratings:
    +352 / 0 / -0
    Is it possible to let players set up this "step back" manually? If one wants to practice their consistency and be rewarded with wins, I see nothing wrong with it...
     
    • Agree Agree x 2
  12. Christian G

    Christian G Topological Agitator Beta tester

    Joined:
    Apr 8, 2015
    Ratings:
    +2,411 / 0 / -0
    I don't think so. You can inject a certain value but if you are consistently faster than the AIs average - which you're bound to do if you win a race - I assume the adaptation algorithm will eventually kick in and step back up once again.
    Race pace. That's what counts for the adaptation algorithm and the AI will perform better in quali.

    @Hendrick Can't say for sure without knowing all the details of course, but it sounds like the AAI might still have been in the process of getting close to your actual level. It does that in a "one step forward, two steps back" metric, so it's normal for AAI to oscillate around your actual level until it finds the best match.
     
  13. Muzarati

    Muzarati Member

    Joined:
    Jan 27, 2018
    Ratings:
    +20 / 0 / -0
    And that's the problem...:( I don't think that should happen...couple of tenths....maybe?
     
  14. Maskerader

    Maskerader Well-Known Member

    Joined:
    Oct 6, 2019
    Ratings:
    +352 / 0 / -0
    I'm having a hard time understanding how the algorithm works. Why do player level and AI level must always meet? Why it's not possible to make AAI aim for "player lap time + x seconds"?
     
  15. Muzarati

    Muzarati Member

    Joined:
    Jan 27, 2018
    Ratings:
    +20 / 0 / -0
    Why does it need to take one step back?

    ...and if you want to be pushed harder and get spanked, you can select fixed and crank it up to 110 or whatever?
     
  16. Christian G

    Christian G Topological Agitator Beta tester

    Joined:
    Apr 8, 2015
    Ratings:
    +2,411 / 0 / -0
    Guess there's a simple (mutual) misunderstanding. :)
    I'm only describing the status quo in that quote, because I thought you were asking whether it's possible right now for the user to change sth about the way it works, by editing the adaptation file or sth like that.
    I think it would be possible to change the code in such a way, but yeah, it would require exactly that, a code change and I'm not sure how eager the guys are to change this at the moment. I can make a note about your suggestion, but don't expect things to change any time soon.
    Because right now it's based on averages, your lap time average and that of all AI drivers. So as described above, we'd need to change that one way or another for it to work differently.
    Of course you could do that but it wouldn't be quite the same. One of the main advantages of AAI is that, once it's trained properly on all tracks your going to use, it can balance out slight differences in the users pace across tracks when you're running a championship. That's not possible when using a fixed strength. Secondly AAI will automatically find the best matching level, so the AI don't outrun you by a mile, sth that can happen if you just "crank it up". :)
     
    • Like Like x 1
  17. anno900

    anno900 Well-Known Member

    Joined:
    Dec 27, 2018
    Ratings:
    +47 / 0 / -0
    I fully agree to your last part and that is what I was always aiming for. But I did not have the time to train AI good enough on every combo. I have done 2 more rounds with my 4 race per track with 5 laps setup. The AAI adapts quite well. The only issue is that AI is too slow in the first 2-3 laps whilst the filed is packed, so I gain places too easily.
     
    • Agree Agree x 1
  18. Christian G

    Christian G Topological Agitator Beta tester

    Joined:
    Apr 8, 2015
    Ratings:
    +2,411 / 0 / -0
    Agreed. It's sth that bugs me as well. What I'd like to see in that regard is the AI using the entire width of the track while they're still bunched up, so it wouldn't be as easy to find an empty outside lane on which I can pass half the field.
     
    • Agree Agree x 3
  19. anno900

    anno900 Well-Known Member

    Joined:
    Dec 27, 2018
    Ratings:
    +47 / 0 / -0
    You would need more waypoints. Difficult to implement I think. Maybe there are only few parameters, which need to be adjusted in regards to how close the AI cars can get next to each other.
     
    • Agree Agree x 1
  20. Christian G

    Christian G Topological Agitator Beta tester

    Joined:
    Apr 8, 2015
    Ratings:
    +2,411 / 0 / -0
    Absolutely.
    Luckily for me my role (<-) is limited to making smart suggestions instead of figuring out how to translate them into code. :D
     
    • Like Like x 1