Stefano Casillo moves on to new adventures from Kunos Simulazioni

Yeah, thanks for the info and although temporarily back...
Still a sad, sad day.

Waiting for his 1:1 world ship simulator, would buy it it in a minute with no questions asked.

 

I don't know the Neural AI from MotoGP 2020, but imo the key technology in this area is imitation learning. We don't want to have 30 AI opponents who drive every track perfectly, all within a millisecond. What we want are AI opponents that imitate human driving behavior. Training such an AI can be computational demanding, but the good news is that this can be done completely during development. The evaluation of such a model during the game is way faster, even compared to classical algorithms (of course, depending on the complexity of the models). The real effort is to get the data for training neural models. And here we have a great chance with the sim racing community to support the devs with data.

 

What he has done to the AI on his temporary comeback so far looks amazing. So amazing to the point that it puts me off playing against the current iteration of the AI since they are so boring.

 

Indeed, very good improvements and, best off al, it seems with simpler code (Stefano words, I'm totally ignorant about any sort of coding).

The "I" in ACC "AI" does have a real meaning
Even Stefano seems amazed sometimes with the outcome.

 

Agreed. The AI looks much better in the most recent video.

I find it fascinating the different approaches that all the titles take to AI performance. I've seen titles like AMS with a very simple implementation (their words) that works exceptionally well and i've seen folks like iRacing do a really deep series of simulations that affect everything from driver aggression to pit strategy aggression. Both styles seem to be successful from the customer end point perspective.

 

tried iracing AI race, seems atrocious to me.

 

@kunos

Stefano, perdona la domanda in italiano ma altrimenti faccio casino...
Ho visto in ritardo il penultimo stream su Acc... Ma in Acc, la IA non aveva già valori preimpostati per ogni pilota?
Sono strasicuro di aver letto nei primissimi mesi qui nel forum che ogni pilota virtuale avesse già un suo valore preimpostato. E questo si spiegherebbe anche vedendo i risultati delle gare della cpu, i migliori (piloti, non macchine) stanno davanti, quindi Bortolotti, Marciello, Stolz, ecc... i più scarsi dietro.

Perchè inserire in parametro per variare il range di IA? Non esiste già delta di prestazioni in base ai piloti virtuali?
Se già adesso imposto IA al 100%, o 90%, vedo risultati e tempi sul giro sensibilmente diversi tra i più veloci ed i più lenti, anzi, a dirla tutta forse il delta tra il più veloce ed il più lento in certe occasioni è già abbastanza elevato.

Cortesemente, riesci a spiegare un pò meglio sta cosa?

 

I think Stefano is working on adding random range of "performance of a day" to mix app order a bit

 

https://www.twitch.tv/kunosstefano/clip/SpinelessPiercingMouseYouWHY

Even now, in a total wip state, the result is awesome.
I've watched all the today's stream and the jump from "our" AI to the experimental AI is huge, really huge.

Congrats Stefano.

 

Thanks Panky.
If i've understand correctly, the gap between the fastest and the slowest can olny increase with the new options, am i right?

If it is, from your (all Kunos team involved into it) experiences, isn't the actual gap already "big enought" to be believable?

Anyway, it's always good to have options to personalize our own expericence, so, well done (well... almost done right now... )

 

Is this new AI a (maybe beta) download one can get somewhere?

 

As soon as it is released...

 

With time difference. I can't get to watch it live so no point of joining twitch to comment later.


Just want to say the improvements look amazing to what we have now with ACC.


This will transform ACC single player, as MP has gotten very competitive and requires lot of practice now.


Thanks for doing this and I really enjoy watching the stream replays.

 

Late to the party @kunos
I've watched the tuesday's afternoon stream only now, and i've see the "problem" with AI's zero grip which is anyway able to run fast.
Can it be the reason why AI, in raining conditions, is way faster than human,?
(or at least, there are some areas where human can't never match AI's performances.)

Also in the Aris's stream (when he tried racing @Zolder in rain conditions) this was absolutely clear to everyone.
The momente when AI's cars put let down power for acceleration, they simply fly, when the human's car simply can't do that.
Watch here:


From 2:40:36.
In every lap, the corner exit from T9 (it means, accelerate from a very low speed, where traction is needed) there's a huge difference in acceleration between Aris and all the cpu's cars.
They simply have a perfect traction, way way better that "traction control system" in our car.

This always happens in all this kind of conditions, like exit from T1 @Monza (under rain obviously).
Traction needed -> CPU just fly.

 

Can we expect to see a couple cars crash out during a 20 minute race and perhaps some aggressive pit strategies from the AI if rain drops start to fall?

 

I am trying to watch some parts of the two streams from today, but i am not sure I have understood the areas where Stefano improved the AI. The only part i seems to have understood is the AI cars are now slightly better at avoiding a stopped car on track.

 

there was the "move to the middle" at the start and some other strange behaviours that could be eliminated today.
there was a guy named "aristotelis_drives" or similar and he said something and stefano did something in the code and - voila - miracles can happen

 

What i've seen is they are much more likely to pass. He seems to be testing at 80% aggression which is good because it means you can crank it up for more aggressive driving. If we can get the AI to not only pass humans but each other and be willing to hold their line side by side in longer corners like at Monza....that would be amazing.