Building Mario Levels with Machine Learning | AI and Games #39
Вставка
- Опубліковано 16 вер 2024
- It's been 10 years since the first ever Mario AI Competition, so I return to the world of Super Mario level generation research and catch up one some of the more interesting examples that have arisen in recent years.
--
You can read the article based on this video on the AI and Games website and on Medium:
/ mariomakers
www.aiandgames....
--
This video uses footage of Super Mario Bros. from World of Longplays:
• NES Longplay - Super M...
Video footage of 4K Metroid Prime 2 from UA-cam channel iamthegr34td3str0y3r:
• Metroid Prime 2: Echos...
This video is inspired by the following AI research papers and projects:
STEVE DAHLSKOG:
Steve Dahlskog and Julian Togelius (2012): Patterns and Procedural Content Generation. Proceedings of the FDG Workshop on Design Patterns in Games (DPG).
julian.togelius...
Steve Dahlskog and Julian Togelius (2013): Patterns as Objectives for Level Generation. Proceedings of the Workshop on Design Patterns in Games at FDG.
julian.togelius...
Steve Dahlskog and Julian Togelius (2014): A Multi-level Level Generator. Proceedings of the IEEE Conference on Computational Intelligence and Games (CIG).
julian.togelius...
Steve Dahlskog, Julian Togelius and Mark J. Nelson (2014): Linear levels through n-grams. Proceedings of Academic MindTrek.
julian.togelius...
ADAM SUMMERVILLE:
Summerville, A.J., Philip, S., & Mateas, M. (2015). MCMCTS PCG 4 SMB : Monte Carlo Tree Search to Guide Platformer Level Generation.
pdfs.semantics...
Summerville, A., & Mateas, M. (2016). Super Mario as a String: Platformer Level Generation Via LSTMs.
arxiv.org/pdf/...
MATTHEW GUZDIAL:
Guzdial, M.J., & Riedl, M.O. (2016). Toward Game Level Generation from Gameplay Videos. CoRR, abs/1602.07721.arxiv.org/ftp/...
VANESSA VOLZ:
Volz, V., Schrum, J., Liu, J., Lucas, S.M., Smith, A.D., & Risi, S. (2018). Evolving mario levels in the latent space of a deep convolutional generative adversarial network. GECCO.
arxiv.org/pdf/...
GitHub Repository:
github.com/The...
AHMED KHALIFA:
The Mario AI Framework (10th Anniversary Edition)
marioai.org/
--
AI and Games is series on research and applications of Artifical Intelligence in video games. It's is supported through and wouldn't be possible wthout the wonderful people who support it on Patreon.
/ ai_and_games
www.paypal.me/A...
www.ko-fi.com/A...
--
Get yourself an AI and Games t-shirt over on Teespring!
teespring.com/...
You can follow AI and Games (and me) on Facebook, Twitter and Instagram:
/ aiandgames
/ aiandgames
/ aiandgames
/ get_tuda_choppa
--
#SuperMarioMaker2 #DeepLearning #MachineLearning
It's been 10 years since the first Mario AI competition and four years since my video covering said event! While the competition ended in 2012 there has been a lot - and I mean A LOT - of Mario-related research since. Hopefully you'll agree that some of these new level generators are pretty exciting! It's been great to revisit that world and I want to give special thanks to Steve, Adam, Matthew and Vanessa, all of whom helped me make this video a reality.
Tommy, would you look into Divinity Original Sin 2? They have a responsive AI that counters the player characters and uses the environment to it's benefit.
Would be really awesome to learn how did they do it :D
do the metroid prime video
I would be really interested in a video showing how AI is being used to upscale older video games graphics. Keep up the great work!
Me too!
You might enjoy AI Nile Ducks > ua-cam.com/video/3iCOxF_UuB0/v-deo.html
Yes please, 4k Morrowind details!
yes metroid prime
If you look at Mark Brown's series "Boss Keys", you will notice that metroidvania levels and Zelda dungeons are actually surprisingly similar in concept, so work that went into generation of Zelda dungeons could perhaps be applied to metroidvanias. And as a Doom player, I personally recognize similarity between metroidvanias and Doom levels, and Doom community used OBLIGE level generator that creates relatively fun customizable Doom levels for many years now.
The perfect machine learning model will just recreate miyamoto's brain
These Super Mario elements are what I would have found in 1-1: brick block lines with question blocks sometimes, question blocks alone, goombas in groups of 1, 2 and 3, pipes, bottomless pits with a length of 2 blocks (not very threatening), risk rewards, secret 1-up block or star (1 in the entire level), satisfying moments where you can kick a shell into tons of goombas, 8-step staircase before flagpole, repeating sections, once without risk and then with risk, a question block "pyramid" with a powerup at the top and a secret bonus area which contains tons of coins on a 7*3 brick block platform and skips the level. 1-1 is a bit special because it needs to introduce new players to all the mechanics of Super Mario Bros, while not eating away an experienced player's time. And I also think Nintendo made a really good job at designing 1-1. (Btw, the section at the start, that you see all the time, is like a masterpiece of first level design)
Please do that video on AI texture upscaling. I've heard a bit about it before and it sounds fascinating
I wonder if it'd be possible to somehow make an AI that creates levels that are specifically fun for the player. Like, simply by feeding AI levels that people consider good, and leves that people consider bad, and trying to teach AI how to design good levels.
That would be really nice!!!
bringing practical science to video games always brings a smile to my face
Wow, I've just finished my Thesis in Procedural generation of levels using Genetic Algorithms and all this would have come handy, especially the Mario Framework since developing the game was the toughest part and it kept me from focusing on the generation part. Thank you for the video though it gave me a lot of ideas for future projects!
Oh man, sorry about that! Hope the thesis went well despite that (sounds interesting too). The new Mario framework only just came out two weeks ago, so I guess you couldn't use it until now. 😐
Best of luck with your future projects!
@@AIandGames I'm still waiting for the correction, I didn't have as much time as I have wished to but it came out pretty great. Thanks for the good wishes and keep up with the inspiring videos!
The fact that there are even people with PhD's working on Mario games and writing actually research papers on the subjects is insane.
This is just so fascinating
I recently got to talk to the developers of a piece of software called “Chematica” which uses programmed-in knowledge of chemistry to suggest synthetic routes to products. The problem they’d been working on was how to effectively present the options - which might number in the hundreds of thousands - to the user. The big deciding factor after cost was novelty - suggesting a wide range of low cost options and allowing the user to then explore the neighbourhoods of particular options once they see the lay of the land. (I think you could interrogate the graph of options directly.)
Your video got me thinking about whether AI approaches would be useful in helping people find interesting Mario Maker levels, or user generated content in general. Using these tools to evaluate some quantitative parameter like difficulty or clear time, but also to help present a diverse range of styles. (Goodness knows online content/product catalogues could do with that in general, but Mario levels seem more tractable.)
“Dahlskog” Sounds Swedish!
“... in Malmö University Sweden ...”, well that would explain why...
"The DISCRIMINATOR" Thats what I call an AI that was teached bullying Mario Levels :D
That's basically it. "YOUR MARIO LEVEL IS SHIT! NEXT ONE!"
Imagine if you got the AI to make a super mario bros level, and it just made new super mario bros levels in the original mario. Then Miyamoto turns up and points a gun at you. 'You were right about super mario 64. But we didn't just stop there with our AI. No, we've refined it to the point where the personalisation isn't obvious any more. Truth is, the game was personalised from the start.'
I'd expect ML to be more competent when trying to create Kaizo Levels. They are much more linear and the tile placement is super precise. Classical Mario levels are usually more open and have multiple solutions, whereas Kaizo levels expect the player to beat the level in one specific way. Do you know if anyone focused on that?
One thing that may help tremendously (though, I guess that's the first thing AI will replicate) is that there is virtually no verticality (except in fortresses in more recent Mario games).
This AI makes better levels than the entire mario maker community.
I would love a video on GAN texture scaling, but I wonder if general-purpose image-scaling GANs will come to TV sets and make that a bit moot before it’s widely used. Why remake Metroid Prime in HD with GAN textures when my telly could hallucinate a higher resolution for the GC original running off original hardware?
Truly amazing start.. So much potential. First video games, textures, and fake photos, next maybe even something like film?
I love how you pronounce "Mario".
Imagine a Mario Bros game that has different levels every time you play it
Building a level in isolation won't work. The machine learning algorithms would need to be sophisticated enough to understand the player responses required or suggested by a particular environmental obstacle, and it would aim to weave these obstacles together based on their accompanying responses into a cohesive gameplay narrative.
Just sayin'
EDIT: nevermind you cover this 14 minutes in, ignore me
Steve has been my teacher 0.o
😂 Believe it or not, game AI (both in academia and industry) is a relatively small world. I've known Steve for around 5 years now. He's a nice guy and I really like his research too!
You know what game would work well with neural networks integrated into it? Dwarf Fortress. I say that because it's so heavily algorithm based you could generate things with dwarf Fortress and make something which tells if it's convincing or not. You could have the network say "fondlefoot" is a silly stupid suggestion for the name of a town next to a river but "freshwater gulch" is fitting. Generate crazy designs for fortresses and see if the dwarves actually survive and see if it comes up with the exploits that players rely on to make the game run smoothly. Is it merciless, torturing everyone and forcing them to work to make things function? If you told it to make everyone happy, would that result in killing off the weak, making the strong happier? Will it just kill everyone and make sure that the cats are happy wandering around killing rats? If you told it to make a wood themed Fortress, would it build solely on land? If you got it to control the distribution of metals underground to create the most stable economy how would it go about doing it? I know that's not exactly how neural networks function but it seems like such a simulation heavy game that you could make one aspect controlled by a neural network and it would adapt itself to it.
Hey,thanks for your content, you inspired me to start learning to make game A.I. and programming!
Fantastic, best of luck!
I'm interested on those AI texture upscale episodes :D
9:21 Michael Mateas!
“Georgia Tech University” lol
I really really want to play Mario Maker now.
how is GANs used to retexture metroid prime?
18:23 notice how you say organs
19:46 Yes Please
but mario levels are simply an exercise in using the limits of Mario's skills. they are not a pattern onto themselves. these people seem to be approaching it incorrectly.
TEXTURES!
14:50 there is just no way that can happen, this is fake (the top of Mario's cap stays on screen even though Mario comes back up next to the number)
I love your accent, where is it from?
Let's try my best to emulate a proper comment for this video by taking the other comments into account:
Yes, please so GANs perfect just Mark TEXTURES start.
Just mocking. The way the study starts is so broken. The AI even breaks the final stairscase, that is a Super Mario staple.
Players not only get the quantity and spacing of objects, people are introduced to mechanics or rules early in a level or when a new element appears in scene. They won't enjoy dealing with patterns the level haven't warned. Technically humans has to teach computers how to teach humans how to teach computers...
I may be wrong but I sense a Fife accent?? 👀 also I luv the video I should rly be asleep right now lmao
You spent 25 minutes repeating the same phrases over and over again, most of the rambling made little sense..... try to focus on the key points and be concise.
Strange that there is so much attention to generating levels. Like why is it so important? But anyway if it's important what is the definition of how to make a Mario level, why not just contact the original developers?