• Welcome to the largest (sim) racing website in the world!
    Blurring the line between real and virtual motorsports.

Featured A Gateway to the Future: Neural AI in Simracing

Discussion in 'Other Racing Games' started by leon_90, Apr 25, 2019.

  1. leon_90

    leon_90
    Columnist for RaceDepartment Staff

    Messages:
    746
    Ratings:
    +1,108
    brain-technology.jpg
    How would it be, to face computer opponents that behave like human players?


    Simulations have made enormous progress during the last two decades. If you pick whatever title from the mid-nineties, whether it is a flight, combat or racing simulator, and compare it with its contemporary contender from today’s selection, you will immediately notice a big step up in almost all areas.

    Graphics are the most evident improvement, thanks to the newest engines, which are capable of rendering very detailed models, bigger textures and handle a larger number of objects on screen than they used to. This makes it possible to have rather complex scenes with more realistic scenarios. Of course, we all tend to say that physics are more important, and while this is very true, it is also true that having nice graphics surely is a big boost in immersion. We experience reality through our senses, and sight is one we very much rely on. What this means is that having better, eye-pleasing environments helps us believing that what we are trying actually is a believable representation of reality as we know it, making the title more enjoyable.

    Physics is another area that has seen a great advancement. While sims from older generations are still kept in very high regard by the community, and sometimes not only still played but also actively developed or enhanced by modders, it is undeniable that the most realistic sims from the current generation are much more detailed, physics wise. Tyres' dynamics, engines' characteristics, ambient properties, these are all areas that have been enriched with a lot more variables than in the past, making your virtual vehicle much closer to its actual counterpart, and thus giving you a better representation of its real nature, again with obvious benefits for the player.

    Audio is instead one area that while it has been improved, it has though not seen that much of a progress compared to graphics or physics. In the early noughties for example, sims used to include specific engines for audio rendering, and having a dedicated sound card would allow you to have a very different experience from those who did not. It is not an exaggeration or an overstatement. Playing through your audio chip integrated in the motherboard or through a dedicated Sound Blaster would result in feeling like playing two completely different titles; such was the difference in quality. Of course, you could say that integrated audio chips have become way better now, and that on the other side dedicated audio cards have stalled, with not that many being available for purchase, and new models being produced every 5 years or so. However, I cannot help but think that this is also because game developers themselves have stopped supporting those, at least in a sense of making full use of the advantages that a dedicated audio card can provide. It is a bit of a shame, since, as we were saying before, yes, sight is a very important sense for human beings, and the one we give more credit to in games and sims, but hearing is just as prominent for immersion, even though it is usually neglected. It helps having a better understanding of what your vehicle is doing, where it is placed in space in comparison to your opponents, and in what type of ambient you are moving in.

    rf2-roadmap-december-2017.jpg

    Nonetheless, the AI compartment really is where a stagnation has occurred in the last decades. In fact, Artificial Intelligence has not really evolved since the early sims, at least not in the fundamental ways it operates. It can handle more factors now, it can respond to a lot more parameters than before, but it is still hard bounded to an old philosophy of work. AI has to interact with the various iterations it has in the scene (your offline opponents) and, most importantly, you, the player. To do so, putting it in an oversimplified manner, it works through a series of coded lines, which tell it how to respond to certain situations that may, or may not, arise. Where is the limit in this kind of implementation? You cannot code a satisfying response to all of the possibilities that can occur while playing in a simulated environment, especially given the presence of human intelligence, which is unpredictable. Moreover, sims are getting more and more complex, which means even more parameters to respond to for the computer. You can see the limitations in current AI whenever you do something they are not coded to respond to, that they are unprepared to face. Then, glitches and weird behaviours happen.

    The solution will be a new sort of AI, which is already at work in different areas of computer technology but just now breaking through in the mass market, and so in gaming and simulations. I talk about Neural AI.

    project-cars-bmw-1m-stanceworks.jpg

    When we talk about Neural AI, we often hear the term “Machine Learning”. What does it mean? Now, of course I do not pretend to be an expert in computer programming, since I am really not, and I won’t be arrogant enough to expect giving you a perfect, in depth-answer to that question. I think that there are better-suited accounts[1], written by actual professionals that can give you a more accurate understanding of the way Neural AI works. What I am here to tell you, and discuss with you, is the importance of this new approach, how it can revolutionize gaming by overthrowing the old AI with its limitations and bounds.

    Neural AI is no longer chained to a series of fixed parameters, only to which it is able to respond. It is no longer coded to counter act to specific actions, but to general conditions. It means that developers are now doing a very different job than they were in the past. They do not tell the AI exactly what to do anymore. They tell it what they expect it should do in a determined situation. They are, basically, trainers, so to speak, and the AI is now like an athlete trying to reach its objectives successfully. Doing a very simple example, given a certain track, devs ask the AI to set a specific lap time by respecting certain tasks like staying between the white lines, handling traffic, managing the vehicle characteristics. The AI then starts training, and “learning”, by itself. It goes around and around until it understands how to reach the time required in a satisfying (relating to the general parameters set from the developers) manner. It can be trained to achieve whatever objective it is desired to get. Replying to the original question, we say that it is learning because it is indeed. The old AI in fact, as we said, could only respond to situations it already knew because they were coded in it. It had no memory. Neural AI can respond to any situation, because none is coded; only its expected behaviour is. Therefore, when confronted with a problem, it tries to overcome it successfully and if it does not, it ‘thinks’ of new ways to solve it. It applies a strategy. We can finally have an opponent that is actually able to withstand any possible scenario and the unpredictability of human behaviour, because it learns from past experiences, remembers what went wrong and what went well, just as we do, store it for the future and then tries to do and be better next time around, every time.

    Isn’t this something we have been asking for, for a very long time? Computer opponents will be much more “human” than they have ever been, much for our enjoyment, and simulations will make another big step forward in realism. This truly is the future; this will really be the golden age of simracing.

    assetto-corsa-competizione-trailer.jpg

    Like what we do at RaceDepartment? Follow us on Social Media!
    Instagram
    Youtube
    Twitch
    Notes:
    [1] You can start by reading here: https://en.wikipedia.org/wiki/Machine_learning
     
    • Like Like x 25
    • Beer Beer x 3
    • Love Love x 1
    • Agree Agree x 1
  2. Marc Collins

    Marc Collins

    Messages:
    1,156
    Ratings:
    +334
    It will be wonderful! Too bad our PCs today can't even handle the best AI algorithms and programming we already have if you want to race a large field of computer opponents.
     
    • Disagree Disagree x 4
    • Agree Agree x 3
  3. TylerDurden4321

    TylerDurden4321

    Messages:
    164
    Ratings:
    +97
    DrivATARS™

    Those are great, aren't they... huh? ...No? ...ANYone?? :unsure:
     
    • Haha Haha x 12
    • Like Like x 2
    • Love Love x 1
    • Beer Beer x 1
  4. NoThrottleControl

    NoThrottleControl

    Messages:
    74
    Ratings:
    +47
    Live for Speed had learning AI 15+ years ago. And it used the same physics as the player as far as I remember.
     
    • Like Like x 4
    • Agree Agree x 1
  5. mms

    mms

    Messages:
    143
    Ratings:
    +106
    I don't think it's the AI algorithms that today PC's are struggling with, AFAIK it's simulating the physics for the AI cars that does.
     
    • Agree Agree x 13
  6. Kevin Bradley

    Kevin Bradley

    Messages:
    145
    Ratings:
    +115
    One day these 'trained' bots are gonna rise up against us and kill everybody in sight.

    Exciting possibilities in offline sim racing until then though :D
     
    • Haha Haha x 12
    • Like Like x 1
    • Beer Beer x 1
  7. Anomalistics

    Anomalistics

    Messages:
    75
    Ratings:
    +26
    Very impressive write up. MotoGP 19 will consist of neural AI, so we will have to see what happens with this. You're quite right in your post though, the AI has been stagnant for years. We only see the very basic implementation of AI and logic; and even this is done poorly. An example of this is ACC.

    As it stands currently, I would say RF2 is leading by an example in terms of the awareness and positioning. It still has fundamental flaws with slip streaming and overtaking though. I am hopeful that with the release of iRacing Ai in the coming months, we will finally be able to experience something ground breaking.
     
    • Agree Agree x 3
    • Like Like x 1
  8. BrunoB

    BrunoB

    Messages:
    462
    Ratings:
    +114
    Principally neural AI could be way better than the complete un-intelligent AI all racing games are using today.
    Hehe but I will se a good FUNCTIONING neural AI system on a home computer before I believe it.
    Until that is fullfilled I just consider sim devs info about it as rather hot air.:cool:
    Because until now they cant even produce AI that use same physics and TM as the player car/MC.:poop:
     
    • Agree Agree x 1
  9. Ben O'Bro

    Ben O'Bro
    Premium

    Messages:
    975
    Ratings:
    +1,753
    so we can expect a mess on the first corner ? Nice.
     
    • Haha x 11
    • Beer x 8
    • Like x 3
    • Agree x 2
    • Love x 1
  10. Bazooka_Joe

    Bazooka_Joe
    Best of the worst, sometimes worst of the worst.

    Messages:
    266
    Ratings:
    +156
    So we'll end up with frustrated, angry rammers that are often met in online lobbies in so far clean offline modes too? No thanks!

    edit: @Ben O'Bro - one mind :D
     
    • Beer Beer x 3
    • Disagree Disagree x 3
    • Haha Haha x 1
    • Agree Agree x 1
  11. RobertR1

    RobertR1
    Premium

    Messages:
    743
    Ratings:
    +347
    I rather the AI be used for vehicle tuning first. I want a virtual "team" behind me as a driver.

    One one hand the sim is trying to replicate real life. On the other hand, I'm the driver, mechanic, engineer and strategist for a GT3 team which is absolutely not close to reality at all. Those functions should be AI driven and combined with audio feedback as a driver would give. "I have understeer in T1 and oversteer in T14" etc.
     
    • Agree Agree x 15
    • Like Like x 9
    • Beer Beer x 1
  12. Marc Collins

    Marc Collins

    Messages:
    1,156
    Ratings:
    +334
    I think it is both, but probably the physics is even more of a load. Reiza expanded the AI awareness (ranges and frequency) in AMS as part of the development process...still, it makes a difference just to do that let alone using more sophisticated algorithms.
     
    Last edited: Apr 26, 2019
  13. guidofoc

    guidofoc
    Premium

    Messages:
    619
    Ratings:
    +381
    The article is interesting but it covers a lot of other stuff and only halfway through there is something about the topic. Also, can you give us some example of neural AI? Otherwise this thread will be the usual 'my AI is better than yours'.
     
    • Agree Agree x 4
  14. Anomalistics

    Anomalistics

    Messages:
    75
    Ratings:
    +26
     
    • Like Like x 8
  15. 147852369

    147852369

    Messages:
    26
    Ratings:
    +14
    So true. Developpers are adding more and more settings to the cars but you're left alone with no help. Most sims today don't even have some integrated telemetry, how is this possible?
     
    • Agree Agree x 8
  16. Niels_at_home

    Niels_at_home
    Reiza Studios

    Messages:
    256
    Ratings:
    +398
    AI is just really complicated stuff. I think its mainly down to finding the AInstein (ooff) who is the cleverest at programming something that is low enough on CPU usage but human enough in behavior.. AI is all hip and trendy and I'm sure lots of cool things are happening, but at the end I feel the main part still is the cleverness of the programmer here..
     
    • Like Like x 4
    • Agree Agree x 4
    • Beer Beer x 3
  17. mantasisg

    mantasisg

    Messages:
    1,819
    Ratings:
    +1,992
    Turbo super realistic AI just like ultra realistic physics, should make players complain. Developers are hurt by complaining masses. So sadly such ultra realism will never happen. At least not in BIG BIG title.

    To me it sounds terrible, unattractive and too simplified for the player end, and insanely complicated for developer end.... why not just disable setups completely ? For you to know, just for example, "Assetto Corsa" means "Racing Setup". Setups are very interesting and important part of simulation. Without them there is like half of the whole thing gone away.

    P.S. if such feature would indeed be presented, I wish AI would suggest you how to be better driver rather than magically making you a setup to make car to drive on rails :D
     
    • Agree Agree x 1
    • Disagree Disagree x 1
  18. Will Mazeo

    Will Mazeo
    Premium

    Messages:
    2,008
    Ratings:
    +764
    rF2 has learning AI too
    But I think if an AI use same physic as the player things are not very right... something is super simplified then (at least with current tech), but that's just my opinion. On the other side rF2 AI use rF1 player physcis

    To the whole "neural" thing I can just say... I'll believe when I see it...
     
    Last edited: Apr 25, 2019
    • Agree Agree x 1
  19. Salvatore Sirignano

    Salvatore Sirignano
    Premium

    Messages:
    147
    Ratings:
    +73
    Geoff Crammond did it best with Grand Prix 4, some 17 years ago. :inlove:
     
    • Agree Agree x 4
    • Like Like x 2
  20. RobertR1

    RobertR1
    Premium

    Messages:
    743
    Ratings:
    +347
    You'd still have the option to override and still would accept the changes made. You wouldn't be locked out of it. You'd have the AI act as brainpower and as a point of reference. Putting a car on rails is still limited to a drivers ability to drive fast. You're not going to turn a tail end driver to a front runner. Let's shift from setups for a min and move to race strategy.

    In a long race, you're expected to come up with strategy changes on the fly, keeping an eye on everyone's pace and then hoping you don't come out in traffic to make you strategy work. That's a perfect use of AI built in in giving you strategy options to choose on the fly based on what's happening on the track. Maybe come in early and change tires to come out in clean air? Maybe stay out based on your degradation rates, increase your mileage and get across the finish line having gained positions?
     
    • Agree Agree x 4
    • Beer Beer x 1
  1. This site uses cookies to help personalise content, tailor your experience and to keep you logged in if you register.
    By continuing to use this site, you are consenting to our use of cookies.