AI Issue

Discussion in 'Wish Lists' started by green serpent, Jul 13, 2016.

  1. Euskotracks

    Euskotracks Registered

    Joined:
    Nov 29, 2013
    Messages:
    1,092
    Likes Received:
    191
    Thanks for your unpayable help but saying AIW editor is a good tool is like saying MsDos is a good operating system. Just because the less are capable of properly using it does not prove anything.

    A good tool starts by having a user guide or a properly made video tutorial. More people have learnt to convert tracks to rf2 with my tutorial than to properly use AIW editor .




    Enviado desde mi ONE A2001 mediante Tapatalk
     
  2. hexagramme

    hexagramme Registered

    Joined:
    May 25, 2013
    Messages:
    4,245
    Likes Received:
    194
    The AIW Editor can seem very complex, but once you start understanding the ins and outs, it's really quite simple, like MarcG just pointed out.
    This is not written to belittle people who are having trouble using the editor, but merely as a small "you can do it!" kinda motivational thingy.

    One thing is for sure though: the editor certainly doesn't suck. :)

    Another good point by Marc, if it did suck, every single AIW would be terrible, and that's far from the case.
    Sure, some tracks have rush job AIWs which certainly doesn't do the trick.
    But if time and effort is put into an AIW, it'll provide some damn fine AI behavior.

    Creating the AIW is easy as pie, and there are a few very helpful step-by-step guides around to help beginners get started.
    As for all the little nuances, you kinda learn as you go, picking up tips and tricks partly from trial/error, and from other AIW creators.
    What makes an AIW stand out is attention to all the small details, and this is something anyone can learn.

    There is literally no upper limit for how much "love" you can pour into an AIW.
    You can go all out "obsessive perfectionist" on it, if you wish.
    You can add car specific fast paths tailored for every single car you can think of, if you have the time and desire to do so.
    You can choose to "cut corners" and save some time and still get a great AIW out of it.

    The only real limiting factor I've come across so far is time. How much time I'm willing/able to put into making it.
    Of course driving/recording multiple millimeter perfect laps in each and every rF2 car would be tremendously time consuming, but possible if one wished to do so.

    OK, that went on for longer than I planned.
    TL;DR AIW creation can be daunting at first, but all in all pretty easy once you get the hang of it, and the editor provides pretty much all the functionality needed to make a great AIW.

    It does, right here. :)
     
  3. pmarin2

    pmarin2 Registered

    Joined:
    May 26, 2015
    Messages:
    17
    Likes Received:
    0
    Thanks for the tip, maybe I'm doing something wrong, but in rFactor 1 the AI gave some information on every completed lap, stating if there was a time difference on his runs. With accelerated time, I used to let it drop to 0.0s and then I knew I had a good result. When I do the ai learn, the AI starts lapping, and although the raceline is shown, no information is given. Is it because I need to enter another command?

    Thanks!

    ----

    Update: I left the AI running until its lap time didn't get any better (40 or so laps using time acceleration - CTRL-X by default), and although while practicing they shaved 2 seconds from their first lap, I have to agree with Euskotracks that cars go crazy during a race. I tried it in Road America with Apex's GT3 mod, and all the cars that I run through the ai learn process just kept spinning one after another. Here is a screenshot of them in action

    View attachment 20428

    It seems that editing or making a new AIW is the way to go...
     
    Last edited by a moderator: Jul 15, 2016
  4. green serpent

    green serpent Registered

    Joined:
    Feb 2, 2016
    Messages:
    704
    Likes Received:
    713
    I went through the process until the lap times didn't improve and then saved. But back in the race it was the same problem. The issue really isn't with the racing line itself, which seems fine, it's the fact that they are going into the corner too hot while braking and turning left, and then suddenly turning right. The car is the NSX-R, which I believe has more rear brake bias and is very easy to rotate while trail braking.
     
    Last edited by a moderator: Jul 15, 2016
  5. Juergen-BY

    Juergen-BY Registered

    Joined:
    Jun 16, 2012
    Messages:
    3,089
    Likes Received:
    440
    I went through the process, too, with LRP and the Megane`s. But tbh, i can`t see any improvement.

    Seems to be the only solution.

    View attachment 20429
     
  6. Juergen-BY

    Juergen-BY Registered

    Joined:
    Jun 16, 2012
    Messages:
    3,089
    Likes Received:
    440
    Should the proper usage of the tl3 file somewhere viewable in the log file?
     
  7. Christopher Elliott

    Christopher Elliott Registered

    Joined:
    Jul 31, 2014
    Messages:
    4,551
    Likes Received:
    7,538
    I think both sides of the spectrum could come together a little bit more and try and stay objective and not take things personal. We're fine with criticism, I think that's how things evolve and we want user feedback, but I also understand that some people don't all express their thoughts in the same way.
     
  8. Christopher Elliott

    Christopher Elliott Registered

    Joined:
    Jul 31, 2014
    Messages:
    4,551
    Likes Received:
    7,538
    Thanks for the feedback, we'll look into that.
     
  9. Euskotracks

    Euskotracks Registered

    Joined:
    Nov 29, 2013
    Messages:
    1,092
    Likes Received:
    191
    Some more feedback about the topic.

    I did some tests yesterday using learning methods 1 and 3 (latest).
    I used different versions of Pocoyo Park being the AIW the only difference.
    The learning process was done with AI strength at 100%. It is unclear to me if this makes a difference in result as it does in the laptimes. I have asked about how should be done with no answer (top secret?)

    Without learning (original situation) there was a 3s difference between the laptimes of both versions: version A was running at 1:28s and version B at 1:31s. Hence there is a significant room for improvement for version B.

    However, none of both learning methods did improve version B by more than some tenths. Ini and tl3 files were created with values in it for the different waypoints.

    However, it is really interesting to see that increasing AI strength to 120% lowered times by 5 seconds for both cases.

    In my opinion a bad AIW is a bad AIW and the really effective way of having good AI is by defining a good AIW.

    The efforts should be done trying to help creating good AIWs or improving AI behaviour so that it does not get so sensitive to bad AIW.

    If I were ISI I would ask myself why AI cannot drive fast when following a line taken from the replay of a hotlap.

    Enviado desde mi ONE A2001 mediante Tapatalk
     
  10. Christopher Elliott

    Christopher Elliott Registered

    Joined:
    Jul 31, 2014
    Messages:
    4,551
    Likes Received:
    7,538
    That's some interesting testing right there, thanks.

    It's true that the AI learning is probably more geared towards helping addon cars with AI issues on tracks with an already solid AIW. The benefits might be a lot more minimal on a track with a problematic AIW.
     
  11. Euskotracks

    Euskotracks Registered

    Joined:
    Nov 29, 2013
    Messages:
    1,092
    Likes Received:
    191
    Hi Christopher

    I have been doing some more tests with quite interesting results. I tried two methods which combined I would expect to provide good results. However, in order to do so I would need some features to be introduced. Let me explain...

    I have used the feature which allows using replays for generating the fast path. BTW, I would suggest that the variable in player.json that prevents resuming for replay would be set to 1 in order to make it easier for modders to use this feature. It doesn't make much sense to use the resume from replay in dev mode. I think.

    I have used the replay file that I had where my own laptimes where the best with the megane which is the car I am using for all the tests. I had recorded a 1.24 laptime which is about 2s seconds from the fastest lap set by the aliens of our community (1.22s). The resulting AIW turned out to provide 1.28s for AI100and 1.23s for AI120. Similar times as with the best AIW I had found up to today (the one of my last post mentioned version A)

    I would have tried to use the laptime of 1.22s from another driver in the replay file but it is not possible to resume to AIW editor from a different driver to one who generated the replay. I am not sure about the fidelity of the replay of opponents but it would be nice to try. Not only in order to use other humans drivers laps , but also AI opponent laps seto to 120 after learning the track.

    It should be noted that, at least in some cases, it is required that the replay does not start or end exactly with the lap that wants to be used. For whatever reason (using this feature at least) the recording starts in the first waypoint and not in the start/finish line.

    This doesn't seem a big issue but prevents (in the mentioned cases) using the replay of a single hotlap. If you combine this with the FACT that only laps of the driver that generated the vcr file can be used to record a new fast path, the bypass to this, which would be to create a new replay file using "Replay Office" tool doesn't work.

    I have also given a try to track learning method number 2 (wis file) which for whatever reason I cannot get it to work all times. I am unsure about the required conditions to make it work properly. When I finally have made it work (excited about learning new track) it has improved laptimes to low 1.23s for AI120 (over 500 laps in total, after 240 laps or so, no improvement was obtained).
    For whatever reason the wis file did not successfully created (2kb and no relevant info kn it). It would be nice if this method is clarified...

    Furthermore, it should be noted that when you use any of the three learning methods in dev mode, the correspondent file is not generated within the settings folder as it is in single player. At least I have not been able to. I cannot tell for sure either if copying the structure from single player to dev mode would take any effect since I haven't tried.

    In any case, if the replay from the best lap of that learning AI could be used by resuming to AIW editor from the opponent's car, the above wouldn't be a problem since it would make unnecessary to implement the learning into devmode. I would expect also that offline races can provide an accurate path for opponents which might be a problem when using a replay from an online race.

    As a summary in order to fully exploit rF2 learning features, it seems that the following changes would be very welcome.

    - possibility to restore to AIW editor from an opponent's car when using a replay.

    - ensure that recording laptime starts at the finish line and not in the first waypoint.

    - clarify the conditions to properly make each learning method work:

    Practice: private or not?
    quali and race should be off?
    fuel and tire consumption off, static realroad... number of AI=1
    player.json settings
    AI strengh, AI aggression, AI limiter
    Recommended number of laps
    How to save learning
    How to apply learning.

    Enviado desde mi ONE A2001 mediante Tapatalk
     
  12. Emery

    Emery Registered

    Joined:
    Oct 24, 2010
    Messages:
    3,035
    Likes Received:
    1,654
    Type 2 AI learning needs to advance the session before the data will save. Thus you need practice and at least one other session active.
     
  13. Euskotracks

    Euskotracks Registered

    Joined:
    Nov 29, 2013
    Messages:
    1,092
    Likes Received:
    191
    I have been able to make opponent's car work for saving the AIW. I tried again and pressing + after resuming worked this time. I had to reset best laptime since the first lap established a false record of less than 1 second. Afterwards it started to save the path. In my previous tests and in some other later ones, it wouldn't record laps after resetting best laptime. I am not sure why yet.

    I also was able to save .wis file when using learning method 2. Leaving the track during the session triggered the file saving. It had improved laptimes to 1.23.2 with AI120. The necessity to pick an extra session (quali or race) is confusing since it is not required for the other learning methods. However, when setting autocalibrate AI to 3 to use the wis file, the result didn't seem to work properly since laptimes were the initial ones prior to the learning. ¿? For sure I tried with the car number for which the wis file was created.

    What I later did is, as said, use the saved replay of the learning process and use it in AIW editor within dev mode. As said when using opponent's car it might get tricky to make it work properly. I would need to further investigate about this...

    The result of the new AIW was quite good. At least as good as the best one so far. The laptimes and racing lines are very comparable to real ones when using AI120. I still need to correct the AIW so that the end and start of the lap are better aligned because the AI lifts throttle and loses one or two tenths.

    I think it would be quite hard to beat AI at 120% due to their lap consistency. In the real 20 lap race, humans got their best laptimes at the beginning of the race with tire degradation of over 1 second impact in laptimes.
    In the AI race the best laptimes were being set at the end of the race.

    I guess it would be nice to have a multiplier for AI tire degradation in order to minimise discrepancies between tbc and tgm values.

    Enviado desde mi ONE A2001 mediante Tapatalk
     
    Last edited by a moderator: Jul 20, 2016
  14. Christopher Elliott

    Christopher Elliott Registered

    Joined:
    Jul 31, 2014
    Messages:
    4,551
    Likes Received:
    7,538
    Just so you know, I'm reading along and your feedback is being noted. Thanks :)
     
    Risto Kappet likes this.
  15. Emery

    Emery Registered

    Joined:
    Oct 24, 2010
    Messages:
    3,035
    Likes Received:
    1,654
    I've noticed this effect, too. Which begs the question of what is wrong with AI tire degradation as most mods implement it?
     
  16. Euskotracks

    Euskotracks Registered

    Joined:
    Nov 29, 2013
    Messages:
    1,092
    Likes Received:
    191
    Thanks ISI for reading as well.
    Please note that some of the statements regarding of not being able to use opponents replays were not accurate as I clarified later in the next post to the one you quoted.

    I would also like to add that learning method number 3 worked very bad after a while. In the beginning the laptimes were improving slowly, but it got to a point were the method started diverging and the laptimes were really terrible.
    Learning method number 2 showed much better convergence and the achieved laptimes were also faster.

    In all the cases, I am talking about the laptimes while in the learning process. When I later have wanted to use the learnt path, I seem to be unable to properly do it for any of the learning methods. At least I cannot tell the difference. I will double check this as well but any help would be appreciated.

    Both learning methods were performed with AI at 120. I will have to check myself it this makes a difference or not wrt for example using AI100. There could be two types of difference.

    1. If the learn track can be applied later, performances using the same AI strength could be compared. Influence in the learning.

    2. If using the best lap of the replay of both learning processes to redefine AIW, the components with the obtained AIWs can be compared using the same AI strength for the comparison. Influence of speed (laptime) in the AIW fast path definition.

    Enviado desde mi ONE A2001 mediante Tapatalk
     
    Last edited by a moderator: Jul 20, 2016
  17. stonec

    stonec Registered

    Joined:
    Jun 19, 2012
    Messages:
    3,399
    Likes Received:
    1,488
    With AI using the simplified model, I guess it would be impossible to make their tires degrade exactly same way as human player's tires. I assume that most mods use the same wear curve for both TGM and TBC files, but human tire model has abrasion, thermal deg, flatspotting, changing contact patch... all kinds of things that affect wear, which the AI's don't.
     
  18. green serpent

    green serpent Registered

    Joined:
    Feb 2, 2016
    Messages:
    704
    Likes Received:
    713
    Most of this technical stuff is way over my head haha. In all honesty I like it how the AI stuff it up every now and then, which makes them more human like. What I don't like is when it's the same corners in a similar way. At 'Matsusaka' West circuit the AI run wide on one, and get sideways on the second to last corner consistently(NSX-R). I was doing a race with a friend watching on and he was like 'you know what corners they'll stuff up, so just wait till then to pass'. Also immersion wise it doesn't look good when multiple AI cars do a carbon copy slide at the same corner over and over again. It would be so much better if the AI stuffed up at random corners. This is coming from a layman so I understand how hard it must be to get all this stuff right.
     
  19. pitkin

    pitkin Registered

    Joined:
    Dec 30, 2010
    Messages:
    92
    Likes Received:
    12
    Can you fine tune the AIW on a per corner basis? It seems daft to record a brilliant lap but be a little out on just one corner and end up continually driving to get that perfect lap. Something like the AIW Cam editor that was available for RF1 would be so useful for creating the perfect AI line. Sometimes just a single corner needs tweaking from car to car.
    I find (being a not very good driver) having to drive the perfect lap myself very time consuming. I found recording a fairly good line in RF1 then loading the AIW into the AIW CAM editor and viewing from above to tweak the line more friendly.

    The dev mode AIW creator relies heavily on being able to drive well to get a good line.
     
  20. ECAR_Tracks

    ECAR_Tracks Registered

    Joined:
    May 1, 2016
    Messages:
    456
    Likes Received:
    465
    Just to keep up, adding the TL3 files in any track mas file has no effect, the game doesn't load the learning file. I have to put it manually in the playerfolder\settings\car_x and them the LOG file certifies the game is using the TL3 file, otherwise there's no mention (trace=1000)
     

Share This Page