AI gets eaten by clam
Вставка
- Опубліковано 22 чер 2024
- A self taught AI learns to play New Super Mario Bros. Wii on an underwater level.
The AI learns from a stack of 4 frames from the game, downsized and greyscaled. These are then put into a convolutional neural network. If you would like more explanation about this please comment below!
00:00 Explanation
1:38 The AI
2:28 Training Starts
4:21 8 Hours
6:27 The moment...
7:15 Final Agent - Ігри
This channel has a lot of potential to become popular in the future. I'm glad I randomly stumbled across your video. Keep up the good work!
Thanks, will try my best!
keep posting on this channel please, it honestly has a lot of potential I believe especially for future with AI trends
question: how does the AI behave in another level after being succesfully trained in one level? I mean, is the training in this level relevant or useful for the AI in order to finish another level from the same game, or perhaps, a level from a similar game?
Probably couldn’t do another level after finishing this one, but would pick up another level much faster as it has an understanding of things like the swimming and enemies. Other non-swimming levels would probably be tough as the movement is very different. I may explore this in a future video!
The AI didn’t enter the pipe and finish the level smh
Jk lol, great work! Maybe a cool challenge could be to have it 100% the entire game? Probably not feasible, but just thought I’d mention it.
Thanks! That would be very cool, and if I had enough computers I would definitely try! I’m looking to try it out on a least a few different levels though to see how it does!
@@aitango cant you teach your ai to do one level at a time and slowly make it trough the entire game (i dont know anything about that so sorry) it was a cool vid tho u are a very smart guy 👍
The best way would be to teach it a generalised method, so specific levels. Then, from there, see what it picks up on other levels. For example, most of the hazards in the first world are pits, Goombas and Koopas. If you cover any specific gimmicks, then any levels with no or less gimmicks should actually be easier.
When even the AI hates the water levels. It kept jumping above the water level 😂
I always hated the water levels, I’m just glad an ai can do it for me now
Do you have somewhere we can see your code like a github? I would love to contribute to a project like this
I will be releasing the source code at some point, but I'm currently revamping the code because at the moment its a complete mess. Once the revamp is done and I've gotten it to a remotely presentable point, I will release it!
Just interacting to make youtube promote the channel :)
Very much appreciated!
It seems to me that playing the same level from the start over and over again is overfitting it to that specific level rather than actually teaching it to play any level.
Maybe you should use a similar strategy to the one you used in your mario kart games, by making the AI play random bits of different levels at a time so it can generalize better
Also, maybe you could get good results by using a RNN architecture. This might give it a sense of continuity to its actions.
Then again, I have very little experience with deep learning so I might just be talking a bunch of rubbish
Well many components are repeated multiple times in the level, so it's properly learning those, the exact path is probably partially hardwired into it, but if you took this AI and put it in a new underwater level, it would likely still know to swim at the surface for max speed, not touch the fish, and move to the right.
The AI doesn't have a memory so its not just repeating the same steps in the same order every time.
It's also worth pointing out that human speedrunners also overfit strategies for specific levels that wouldn't work generally too
It is likely overfitting at least a little, I think Sam explained it well. At least using multiple start locations would probably help though to stop really bad overfitting at the start of the level. I've been considering using an RNN architecture, and is definitely on my to-do list. I will have to see how many samples they typically need though, as if they need many more it may be problematic due to the speed of collecting samples in dolphin.
i really wanna see an AI tackle the tower level now by having to go upwards, even if it physically can’t beat the boss
That would be interesting, I have considered doing some upwards based levels, or even the boss fight!
@@aitango video title: “Larry loses to an AI that doesn’t know how to crouch”
Just curious but where should I look to get started with making something similar to this?
Depends on what you currently know! Firstly you would need to learn a good amount of programming, then look into reinforcement learning. There are lots of great resources online and just on UA-cam for both of these. For Wii games specifically, I may release the code at some point in the future
liked the vid but couldnt finish because of the stock music
I like this
I am glad!
Do you know you can play this game on the pc it’s much better since you doing AI coding
We call that 'illegal'
@@imlegos2153 o danm
I'm using dolphin which is an emulator on PC! Guess I am a fugitive now