SeerV7.1 vs. Kito + Vesh
HTML-код
- Опубликовано: 28 сен 2024
- ► Playlist: • Seer: Reinforcement Le...
.................................................................................
Seer: Reinforcement Learning in Rocket League
Socials:
Twitch: / ultrawidegc
Twitter: / ultrawidegc
Instagram: / ultrawidegc
Reddit: / ultrawidegc
Tiktok: / ultrawidegc
I created a discord for you guys to talk with me! / discord
.................................................................................
#UltrawideGC
It's so nice to see some progress with the bumps coming out from Seer.
I have watched a few of these Seer series and this feels like the first one where they have been using their flip resets. Love it!!!
45200 is insane
Seer does use reset sometimes or kind of double taps it in. With just a little more training, if could be unstoppable!
Which activation fhnction does Seer use? (For the hidden layers, mainly) ReLU? Swish? I wonder what is optimal for Rocket League
GO SEER!!!
Can you train it with training packs to get specific mechanics?
CAN WE SEE A 3V3 MATCH?
Hey ultra wide, you've mentioned in a previous video issues with iterating on designs because of lack of compute. Is there a world where you can setup some sort of parallel training cluster from the community? I'd happily donate my electricity to this cause!
Thanks for all your groundbreaking work. Looking forward to the day when seer leads the way in strategy and mechanics.
If there is an issue with leaking model weights, I'm sure your can come up with a way to encrypt everything!
Which would you say is faster for Seer to learn? Just keep playing itself in games or to show clips other humans playing rocket league and it learning to play like them?
I think that’s how Tesla has done their neural network stacks for Full Self Driving. They just fed millions of clips to it and it learned that way.
I guess that would mean a lot more work for you haha 😂
You could backpropagate it to match RLCS Pros, but theu arent perfect. Self play is the road to superhuman. Also, it is easier to do, as you can simulate as many games as you want, unlike driving where you need a real-world environment.
@@BadChess56 yeah RLCS pros aren’t perfect, but I think Seer could quickly learn the mechanics of pros & from there perfect them with basically zero hesitation, instant calculation, and perfect boost usage.
I mean this version of seer has basically perfect rotations, insane challenges, etc. once it has learned complex mechanics it’s OVER