Training AI to Play Fortnite (Reinforcement Learning + Computer Vision)
HTML-код
- Опубликовано: 26 авг 2024
- Source code, pretrained model checkpoint, and documentation for the Fortnite reinforcement learning agent: github.com/wkw...
I open-sourced ScrimBrain because reinforcement learning isn't practical enough yet to help the average game developer do automated playtesting, or to help esports pros practice. My business is making games and videos, not doing AI research (I don’t make any money doing that), but I think RL is the best approach for these use cases.
If you want to support the channel, play my games! CoachDody and I spent months making Steal My Wall!, a fun and intense competitive Fortnite map that helps you master fundamental boxfighting skills like stealing pieces, taking good peeks, and using exploits to get into your opponent’s box: www.fortnite.c...
In case you’re new to Fortnite, all maps are free (devs get payouts based on player metrics) so it doesn’t cost anything to queue into a game. Our map includes solo practice areas with tutorials to explain the skills that other combat maps assume you already have. I’m not a competitive Fortnite player and I’m also not 15, so we designed the solo practice for players like me who want to learn the basics efficiently.
I feel like what we call "AI" in recent years is misleading and over hyped because it's too attached to an expectation of being close to achieving sentient AI and it easily SEEMS that way on surface. I think this technological breakthrough of LLM is inherently a dead end in terms of future progression when looking at it as 'method' of AI. It's more like Mimicked Intelligence. We should be calling it MI.
I agree! The only difference between the nigh useless GPT-3 and the amazing ChatGPT is that the latter has also been trained specifically mimic a human in its outputs.
Alparently people might have found the true source of our consciousness being microtubules in our brains performing quantum wave collapses, in other words no ai exist without quantum computing
Thanks for sharing your learning Will, I'm looking forward to checking out the repo.
nice to see youre back!
Love to see it Will! Hope you continue rolling out videos!
As someone heavily interested in game based AI, this is fucking awesome!
Cool as hell work, cant wait to see your future projects!
RL needs a whole company of engineers to make it work. For some reason it fails a lot at complex environments . I've tried all possible ways to teach it how to play a fighting game (NSUNS4) but i have failed every time .
Nice to see your still e-alive. Nice to see the thought process behind a persomal process. I feel like chat GPT killed alot pf research not related to having a chatbot as the final product behind AI research.
Exactly. It's a dead end in terms of actually progressing AI.
blessing for the algo
THIS IS THE GREATEST AI FORTNITE TRAINING OF ALL TIME!!
❤
Hmm I'm not a Fortnite or Ai expert but maybe you can train the AI in the following:
Spawn a square arena
Spawn bots from a particular point.
Place the AI.
Reward - Small
Ai aims the bots
Ai Shoots the bots
Ai damages bots
Ai Build walls
Ai certain distance from bots
Reward - Large
Ai kills bots
Ai clears the stage
Punish- Small
Bots aims Ai
Bots shoots Ai
Bots too near Ai
Punish - Large
Ai is damaged
Ai is killed
Just an idea
"no useful game ai tool developments in 4 years.. fine I'll do it myself..." Absolute player 🙏🙏
The only notification I get from youtube is when you upload.
Nice video.
I looked a bit at your code, especially the step function of your custom env. I couldn't find any delay between when the agent performs an action and taking a screenshot for the next observation. So the only time that passes inbetween is the time it takes for the function to take a new screenshot (or your sampling rate is synced with the game at 60Hz?) Did you try maybe to reduce the sampling rate of the agent? Maybe that helps to reduce data over the time dimension and lead to faster learning?
Yeah, the sampling rate is limited by the speed of the screenshot, the reward function, and model inference. In my first training runs with the chicken (not documented or in the vid) I hadn't optimized the performance of the screencapture/reward function yet and was getting
@@willkwan yeah, i think with low fps you need some kind of synchronization and ensure the rewards are given for the right action and not async. it may still work if the world is static and the only thing you do is navigating/traversing the map. but as soon as you have moving objects that you need to track/target, i think it quickly becomes a nightmare. so either synchronization or high fps. but high fps i think leads to a lot of redundant data and then I think a larger frame stack size is needed to make the observation space more diverse. but i don't know, just a guess.
amazing
🔥🔥🔥🔥
Can't believe you spent so much time on this. Pure Shooting AI can't get done. It's like hitting the lottery but you are also a baby. Only a pseudo AI can be made with a bot carrying the AI. Rewards have to be immediate and moving around yields no reward. Neither does shooting.
If you want to make a cool AI, Tony Hawk pro skater.
Training in real time might take to long to get enough data
start 2:38
this video fs will go viral lol
Play The Map Called Fun World 4
Very cool, but why didn't you use DrEureka?
Didn't know about it, but with how long the training takes and the lack of parallel environments, I guess it would take a long time to try automated reward function tuning.
bro game sucks ass, the building is zoomer cancer. Pick a good game next time.