Base map> Matsusaka GP Clean Weather, Default Setup, AI 100% My gap from AI table leader: Ford Focus BTCC +0.777 (OK) Radical GT3 Power Pack -0.452 (OK) Corvette Callaway GT3 Power Pack +0.094 (OK) ISI Chevrolet Corvette C6R -0.034 (OK) Radical SR3 Left Drive +4.922 (Not OK) Seat Leon Cup +6.113 (Not OK) So why can I keep up with several AI cars and others I eat serious smoke while being 5 or 6 seconds slower? They're not so different in terms of handling than others. Someone can confirm this sort of experience?
Of course some mods are overly fast (or slow). You should have mentioned the 3PA Gasoline Alley that was never calibrated to Indianapolis...
I didn't know that. I was considering the AI lap times is only dependent of AIW circuit file. Why some cars have those unreal fast AIs is beyond of my understanding.
There is more than just the aiw that goes into ai performance. The ai do not use the same tire as the user. There are also ai adjustments in the physics files like the hdv. Not to mention the talent files. So you can make identical cars perform completely different. It is entirely up to the modder.
Often mods created for league use have almost no value for offline racing because tweaking and optimizing IA is surely a lot of work, and there is no purpose if the contest is created for online racing.
The too fast AI for this has been known almost the instant S397 released this car, unfortunately no "hotfix" ever came and now any further updates won't come until sometime in 2018.
Certainly. It is annoying having to do trial and error sessions for pretty much every track/car combo. If there's a feature in which i think R3E is ahead of every other sim is their Adaptive AI. The fact that it's a closed platform makes it easier to implement ofc, I don't even know if it would be possible in RF2 since every car and track get assigned an ID to keep a log of lap times, and probably you can't control that in an open platform like RF2, but as a single player if there's a feature i'd like every sim to have, is that one.
It would be interesting to know if using any of the three learning methods, these big differences are maintained or maybe they are more even. I would understand that the method that provides more consistent results would show which is the way to go in terms of AI tuning strategy. I am not sure what approach each of the learning method uses in order to modify the standard (without any learning) behaviour. ISI/S397 must know it for sure. Do they alter path or do they affect AI settings for the vehicle? Or both? It is clear that each car has its own AI and talent settings but IMO they shouldn't be strictly necessary. IMO the learning method should try to adjust first the vehicle parameters so that it follows certain path in a quick and consistent way (using a 5/10 lap average should be fine for this). After adjusting these parameters slight modifications of the path could be attempted in order to further improve it. I wonder how car modders do this job or if they do it at all. I haven't seen documentation regarding this AI behaviour.
For the car, the first thing to do is make sure the AI is cornering at the same speed as the human does at the same downforce. With a non-aero car, like the historic F3s, the AI is sooo much slower than a human in turns that to get similar cornering capability, you have to turn up the AIGripMult to 1.08 (e.g. 8% more than ISI gave the car) in the TBC file. That's the quick & dirty way, but you don't want to go higher than 1.10 or else you have AI that can't be bumped out of the way (instead, they're glued down and you're the one who loses). If you can't get the corner speeds to work without going over 1.10, then you really need to work on making a base setup for the AI and looking into AI suspension parameters in the HDV (AIDownforceZArm, AIDamperFast, AISpringRate, etc, but not AITorqueStab since that is obsolete). Another quick & dirty thing to play with is in the HDV: AIPerfUsage. This alters braking and cornering for the AI. Default setups for the AI need to be as neutral as possible on a skidpad, probably slightly favoring oversteer. If they understeer too much, they won't be able to put the power down on exit without plowing off the corner on many tracks, particulary in slow speed turns. [edit: a good high speed test turn is the final turn at Portugal.. if the AI lifts, you need to do more tuning] Default setups for the AI must be insensitive to weight transfer because they are binary with the brakes & throttle. An AI car that pitches uncontrollably will spinout too often. [I recall a line that affects how hard the AI hammers the throttle but can't find it today]
Type 1 AI learning affects only the path, trying for the shortest overall length. Type 2 AI learning affects usage of throttle & brake via the Wisdom file (.WIS). Don't bother with the chat line "/AI Learn"... all my testing ended up with slower AI.
As long as the various tracks AIW are within 3% for a particular car, then they're useable because we're human enough that few people perform consistently from track-to-track without hours of practice. Automobilista is pretty good in this regard because they've worked out a speedy way to manage consistency across their growing library of tracks; no other studio is near as good at it. Raceroom does a decent job, but the growth rate of their track library is significantly slower than Automobilista. The huge outlier for rF2 was Sao Paulo, incredibly slow AI, but S397 fixed that. I've heard mixed reviews of Raceroom's adaptive AI. My own experience was negative, but it's entirely possible I had no patience to wait for it to work. When they roll out multiview triple screen at the end of this month, I'll be spending time in Raceroom and will try out adaptive AI again.
Yeah it's true, and that has actually been my experience. I was doing a championship with the Simtek mod and in 8 out of 10 tracks I was within 4% of AI difficulty but still, a 4% can be the difference (in a 24 field race) between qualifying first or last, and in a 40min race that can translate into hotlapping in 1st for 40mins or trying to play catch up for 40mins, so I always do trial error sessions to find that sweet spot where i'm running at the same pace as the AI and qualify between 5-15...then i know i will have a fun and challenging race. Overall i still consider RF2 AI more intelligent than Raceroom's but I was just talking about the concept of them adapting to you, as in R3E i can forget about this trial/error sessions. Next time you try, I'd recommend deleting the aiadaptation file first and starting fresh (the game will create a new file). Run 2-3 quick races in the same class/track combo and it should get up to speed. It will use that also for other classes/tracks, so it should already be close to your level should you choose another car/track. The key is being able to do consistent lap times, as it will adapt faster.
I would like a function like this being an offline player: an AI that adapts in real time to the pilot. Of course in case of big mistake it should not that the AI rolls at 50km / h while waiting for the player ... I do not know exactly what form it would be but it is a function that I would like.
I guess most of you know this both threads and. I can't say if they are still valid in all aspects. So if, then please use it on your own risk. For me in any case they was (and still are) very helpfull for a better understanding how things are working together. https://forum.studio-397.com/index.php?threads/the-definitive-ai-settings-thread.38584/ https://forum.studio-397.com/index.php?threads/ai-learning-step-by-step-guide.47657/ many thanks to MarcG
You say that, but once you realise that rubber-banding is happening, you realise that your speed/consistency is no longer important. EA (The publisher) games often have pretty obvious rubber-banding and it is horrible. Test Drive Unlimited had a quite subtle version, and the first time I did the race around the island (Takes about an hour) I really enjoyed it. The second time, going noticeably faster, didn't produce any bigger gaps between me and the AI. After that it was just me testing how much the game was modifying the AI speed based on my actions. "I do not know exactly what form it would be but it is a function that I would like." And that sentence redeems you.
You probably right I don't know. I didn't play video games since ~15 years... (apart from that : RBR, Rf1, rf2, and.... 7 hours of assetto corsa) So I do not know the disadvantages of this function.