A new idea for sims - ai constantly learning from humans

I'm guessing Kunos is trying to get a good AI with a decent balance of awareness, collision avoidance and aggression that works for all cars and all tracks. Once that has been achieved hopefully they can add individual characters to each driver which will make the whole single player experience much much better.

If they spend too much time on driver individuality at the moment and then have to make major changes to the main AI calculations it'll probably have been time wasted. But it's an important part of AI that many sims don't get right, I play Grand Prix Legends and racing Jim Clark is a different experience to racing Jack Brabham or Jackie Stewart or Dan Gurney or Graham Hill.... adds a lot of depth to the game.

 

I would personally love to see an Ai car bump another one off the track and then see that off track Ai come steaming into a hair pin and exact revenge by T boning the other Ai and then see a spam of vote-kicks against both AI's.

 

Self evidently not as AC ably demonstrates.

As for the OP's request, really it'd be good if they put AI control into the hands of modders and app writers if it isn't already - let the community fix it.

 

I just don't want it to go the way of rf1 where you need a custom aiw for every car & every track to make them fast - so unless you know how to make them yourself, you're stuck with only having competitive AIs on a small set of tracks.

 

That would be better, IMO, than being stuck with AI that don't work at all.

 

I gave this a bit more thought. Below you see a diagram of what I think an AC AI might look like:


Since the AC AI uses the AC physics, the output of the AI module would need to be Throttle, Brake and Wheel Turn. Inputs are position (on track), direction/pitch, velocity, and potentially things like grip levels/slip angles, which might limit what you can and cannot do. (i.e. a correction on a straight will be easier than in a corner, already at grip limits). Of course, you also need to input at least basic info on other cars within a certain distance (let's say 20-30 meters or so, or whatever seems appropriate). The AI would need to come up with these outputs, which are then filtered through the Physics Engine to update the car position, etc. which then itself becomes input again for the next cycle.

I believe that AC devs mentioned the Physics run at 50 Hz or so, but the point is that this is a continuous cycle of updates, whether that's 50 times a second or another number is neither here nor there, but shows we only have a very short time to calculate the AI for one car, let alone for 20 cars on track. Things will have to be simple and quick to be able to get it all done in ~20 ms (or less, preferably).

In blue you see a straw man of what the AI module should do, a series of rules, if you like that it needs to apply. As you can see, some are a bit contradictory, so you'd have to play with priorities. Yes, going around the track fast is critical, but should be overridden if doing so would cause an accident or ram an opponent off track.

You also need something of a random element, to allow the AI to make mistakes, and appear more human.

------

Now, training this model (and note, I am not saying what model we're talking about here, whether a neural network or a basic rules engine or something else) might be able to leverage human lap data, introducing an element of "supervised learning". This may be useful (even with careful selection of which human laps you'd use), but it may be more beneficial to use unsupervised learning, so it can learn by itself based on these instructions and optimize outcomes. (specifically, lap times and race position)

Learning would require to lower the physics frequency, temporarily, so that all the calculations can take place. However, once the AI is trained, the predictions might be able to fit within the normal physics frequency.

Now, I would imagine that a lot of the current AI (except the human lap data) already works this way up to a point, but it is never entirely self-learnt, given the ai and fast ai "templates" that it uses. We don't know, but I wonder if they experimented with learning as described here (without the assistance of those two files) and realized quickly that things could just not get computed quickly enough to match the frequency of the physics engine.