Research with the AI

Discussion in 'General Discussion' started by J0E, Nov 12, 2020.

  1. J0E

    J0E Registered

    Joined:
    Sep 26, 2019
    Messages:
    156
    Likes Received:
    61
    Let's start with the question first and then the rationale behind asking it:
    Is there any value to using the AI to test different setups?
    Background: I did a 10-lap race at LeMans last night and noticed that some of the AI pitted on lap 3, lap 9, etc. After the race, I loaded up LogAnalyzer and saw that several of the cars had started with 25% fuel instead of 100%. I read this is possible if you don't go through the full sessions or if you restart a race midway or skipped parade lap and all that, but this was a clean race weekend with none of that stuff.
    One of the suggestions was to force the AI to use my setup (I think I have to set the tires to "hard" before doing this else the AI will pit on lap 1 to change to hards). Anyway, in the past when I have forced the AI to use my setup, I have noticed a couple of seconds improvement in their practice laps. That led me to wonder if forcing the AI to use a specific setup and looking at the results of their laptimes would be an effective and meaningful way of narrowing down setup changes.
    Please, before you start typing about personal driver style and all that, I'm not talking about using the AI on fixed setups to find the IDEAL setup. But if you were to make changes and look at the relative changes before and after the changes, you might be able to, for example, find out that lowering the car past a certain level starts to have a negative effect (e.g. if the car is now bottoming out). Having the AI test the changes might be more meaningful since it would remove the human fault element and inconsistency. From LogAnal reports on consistency, the AI maintain a fairly consistent performance, whereas mine is always worse than theirs. My inconsistency and poor driving style could invalidate my personal testing and I would have to drive scores of laps with each change to flatten the curve caused by that. The AI isn't going to vary that much, and I think it would be more efficient for them to do the testing.
    I'm sure I'm not the first person to think about this and it may have been addressed in the past (a search of "can I use the AI to test car setups in RFactor 2 on a Thursday when I'm supposed to be working instead" doesn't turn up any meaningful matches except for links to the HR department).
     
  2. davehenrie

    davehenrie Registered

    Joined:
    Jul 6, 2016
    Messages:
    7,453
    Likes Received:
    4,369
    You found one of the AI limitations in the hard tire only example. Another is the AI do NOT use 100% of the human car's physics. So any testing to benefit you would have to consider the AI are not as fully constrained as you would be. A third problem is the fuel discrepancies. Someone recently found if his chose a timed race, say 2 hours, many of the AI would run out of fuel and the log analyzer showed the same random light fuel load. If he chose LAPS instead of time, all the cars started with full tanks.
     
  3. J0E

    J0E Registered

    Joined:
    Sep 26, 2019
    Messages:
    156
    Likes Received:
    61
    The fuel one is interesting because I set the race for 20 laps and saw the fuel problem in LogAnal. I have to think about the human constraints part. I'm not sure it matters because I'm not looking for the AI to find an ideal setup that I could use. I'm thinking the AI could tell me that it's better or worse to make a change up or down. For example, is this a track where it's better to run with a lot of downforce or little downforce. Even if the AI had their own, separate alien physics, the relationship between parts ought to remain the same (e.g. more downforce, less straight line speed). Again, I need to think about that some more.
     
  4. Lazza

    Lazza Registered

    Joined:
    Oct 5, 2010
    Messages:
    12,345
    Likes Received:
    6,572
    Quite possibly useful as far as downforce vs drag analysis for a track. Finer details like ride height are hampered by the simpler AI physics.
     
    Emery likes this.
  5. J0E

    J0E Registered

    Joined:
    Sep 26, 2019
    Messages:
    156
    Likes Received:
    61
    Yes, I can see where the "simpler physics" would be a showstopper for this approach. You wouldn't know what is modeled and what is not and whether the relationships between different car components are in place. This also makes me think that forcing the AI to use your favorite setup is more marketing than reality. You may think you're running against cars that are going to perform just like your setups dictate, but you're really not. Add to that session settings like AI strength, etc., and you're clueless about how they're really going to run. And applying any information gathered from this would probably be only slightly more valuable than doing a post on Reddit saying "how do I go faster?"
     
  6. Lazza

    Lazza Registered

    Joined:
    Oct 5, 2010
    Messages:
    12,345
    Likes Received:
    6,572
    The AI uses a lot of the same physics values, but runs at a slower rate and with some simplifications. So much of the setup will have some effect. They use completely different tyres as well. Of course it's the modder's job to make the performance match the player physics quite closely, and while there might be adjustments needed to overall spring rates or tyre grip / sliptables to make them a close match, things like wings (with associated downforce and drag) are exactly the same and won't vary that much with the different physics rate. Ride height and more sensitive aero (so bottoming out, and diffuser or front wing height) will be more affected for setup analysis purposes.
     
  7. J0E

    J0E Registered

    Joined:
    Sep 26, 2019
    Messages:
    156
    Likes Received:
    61
    So, Lazza, if you were to place the "use the AI to gather info on setups changes as described in the first post" on a scale of 1 (valuable) to 5 (educational) to 10 (waste of time...go race instead), where would your checkmark fall?
     
  8. Lazza

    Lazza Registered

    Joined:
    Oct 5, 2010
    Messages:
    12,345
    Likes Received:
    6,572
    If I were to guess on ballpark figures, for aero dependent cars I reckon you could judge wing levels on about a 3 (wouldn't go lower because of bottoming out issues). Gearing 5. Brake bias 6. Everything else 8-10. So a useful reference to get starting levels on a new track, let's say.

    How's that? :p
     
    Emery likes this.
  9. J0E

    J0E Registered

    Joined:
    Sep 26, 2019
    Messages:
    156
    Likes Received:
    61
    I appreciate the opinion.
     

Share This Page