@Lazza , and everyone else of course: Are the so called empirical tyre models like the one from RF1 the reason why other games are capable of running player physics? And about the difficulty of making AI drive like a person. Indeed in Assetto Corsa the AI maxed out is 6% slower than real life with their F138/Silverstone combo. Or maybe the car setup they use is very slow like the default one.
@caravan_driver No, if we're separating the models based on the type of data given to it (ie rF1 and similar take 'only' some sliptables which then run live, whereas rF2 needs tyres designed and then built for hours), either model can be used for AI. The player physics in rF2 are demanding and run in realtime, and for that reason only the AI can't realistically use them - unless they wanted to design a game with only very small fields of AI. Even rF1 didn't use full player physics for AI. They were limited in frequency (40Hz for AI vs up to 90Hz for the player; these are the external rates [input/output], the player physics ran at 360 internally, not sure about AI) and they used a simplified model, and that was probably due to CPU use at the time. If you're trying to maximise player physics for current gen PCs, and still allow large AI fields, you're going to end up using different physics for the two. rF1 on modern PCs runs at breakneck speed and could probably handle AI fields using player physics, but having all cars using the same physics at the time (2005) would have meant slowing down or simplifying the player model. It sounds like I'm putting other games down, because really I'm saying instead of rF2 lacking something in not having player physics for AI, it's the other games that lack in player physics compared to rF2. And that's true to some extent, but what matters is the outcome. Anyone can write code that taxes a CPU, the point is what you actually do with it. As for AI matching humans, I think the difficulty is matching up a realtime model (track layout with kerbs etc, track grip, rubbered line, damp patches, drying line, tyre wear) with getting the AI code to predict what to do next. That's all people do (predict what to do next) but AI tends to be more reactionary. With simpler games of the past that didn't have all that variation, getting AI to match players wasn't all that difficult. Take just a simple kerb: if you decide kerbs should be used then your AI might make use of it, but if this particular kerb is bumpier than most, or it'll unsettle the car and make the next corner more difficult, or it stopped raining 10 mins ago and it's still wet although the track is dry, maybe you shouldn't take it. Or use it only half as much as normal. But maybe this other car needs an entirely different approach. rF2's strength (variation, open content) is its weakness in that regard. The more complex the AI physics is, the more variables you throw into that mix, and the harder it is to make the AI think.
It all depends on the learning technique and decision making criteria to be applied. Humans learn by trial and error plus the perception of improvement or regression. If you have tools like delta best that provide a more accurate and quick feedback you learn faster. It doesn't matter if they are playing a videogame from the 80s or driving a real car with the most complex physics like it is real life. IMO the way to improve AI behaviour would be to divide the track into several sections trying to improve each one of them. The division between sectors should be carefully taken so that errors in one sector wouldn't penalise others. This type of analysis can be easily done in motec as well.
It is anoying that when you put the difficulty level at let’s say 110%, you teammate in offline endurance racing (your own ai) can’t keep up with the pack and it is way faster when setting the diffuculty level at 95%. So why is it that the ai is set to 110% when you set the difficulty level at that, except you’re own ai teammate, who is set at a constant 100%?. You’re own ai teammate is also using the ai physics and tyres.
