A.I. teaches itself to drive in Trackmania

23 thoughts on “A.I. teaches itself to drive in Trackmania”

1. Thanks for watching this video !
   This is the first time i'm using NEAT algorithm, so there is obviously still room for improvement. The main problem is that my AI doesn't have a map memory, and can't anticipate "what comes next" with its current inputs. I have some ideas to improve my AI, so don't forget to subscribe if you want to see the next steps of this project 😉

2. Perus elon musk ft. Tesla ja niin eespäin

3. uN aCcEnT vRaImEnT

4. very interesting 😀

5. The french accent is killing me

6. It would help if you chose a player that didn’t cheat.

7. Alors si t'es pas francais toi

8. I'd like to know how did you connected the AI algorithm to this video game, and what is the underlying code please

9. are you French ?

10. soundtrack sounds much better at 1.16x speed =P

11. props to this guy at 13:17 going full send the wrong way

12. Code pls!

13. Code pls!

14. People complaining about him having an accent while they can’t even speak French is incredible. People are so dumb. Don’t worry, you’re very good, just keep working!

15. AI: acts like stupid 5 year old child at first, learns and becomes informed 5 year old. Not so bright also.😊

16. R u french ?

17. 13:33 yunno those TikTok’s with the weird things like hands melting and stuff, just the same vibe

18. You should have taught the AI that it has a direction that it is facing, so it would naturally learn at what angles it can afford to hug a wall instead of always avoiding them. the fitness score being impacted by things like wall-hits negatively and maintaining high speed would also benefit the neural learning greatly…

    Finally, might just be me, but it looks like you had a way too wide range of mutation for your experiment based on your sample-size, as even during generation 99, you had what looks like 5 samples of cars that don't even leave the start. If the fitness score was truly effective, you should have been fully rid of problems like that just a few generations deep, much less 100 of them.

    I bring this up to contest the point made at the end of the video, as you said your supervised learning was superior to genetic algorithm learning, but it seems like a flawed stance to take when the algorithm used in this case had multiple points of extreme weaknesses. It's like saying that in a contest between your mom and a professional cook, your mom is better at preparing food because she made lasagna the way you liked it growing up when you didn't give the professional cook any pasta dough. You could get WAY better results if you gave the genetic algorithm it's due dilligence and adjusted a proper fitness score, allowed the time to accomodate the entire track from the start, and simply gave it the time it needed to shine. Instead of making an AI that learned to *drive trackmania*, you made an AI that upped it's score arbitrarily first based on a not very good fitness metric, then associated that with a completely different goal and thus taught the AI that driving far instead of racing was the goal, all while having such a strong mutation factor that it couldn't reliably get better at it. That's most likely the reason your result improvements flatlined so early, as any improvements below that was completely negated by chance.

    I urge you to try again if you really want a fair comparison, but if all you wanted was a clickable video, at least you made something. Good effort, but could have been way better, 6/10 🙂

19. muy buen vídeo, que ordenador tienes para poder simular eso?

20. Song name at 0:00?

21. super video ! (ton accent français est fou😂)

22. Why didnt you let them learn until they beat you……….

23. so the ia is stupid asf

Comments are closed.