MarI/O - Machine Learning for Video Games -

MarI/O – Machine Learning for Video Games

Views: 10503994
Like: 171036
MarI/O is a program made of neural networks and genetic algorithms that kicks butt at Super Mario World.
Source Code:
“NEAT” Paper:
Some relevant Wikipedia links:

BizHawk Emulator:

SethBling Twitter:
SethBling Twitch:
SethBling Facebook:
SethBling Website:
SethBling Shirts:
Suggest Ideas:

Music at the end is Cipher by Kevin MacLeod


  1. So when the first single cell organism died after making a fatal mistake how did it reproduce itself with memory to correct that mistake?🤔

  2. how does a computer learn when encountering a different map? does it already know what it will do after ten seconds or it solves it as it goes?

  3. isnt this guy some kind of crazy minecraft redstone engineer?

  4. didn't expect that he was this kind of programmer

  5. When your stuck on a level on mario so you ask your big brother to help, he is bad at the game, but he can code…

  6. Excelente muestra de Machine Learning. La técnica de este programador fue la de 'Prueba y error', el algoritmo primero ejecuta pruebas y tomando en cuenta los puntos críticos, guarda el estado analiza las posibles acciones ante el estado crítico y vuelve a ejecutar la prueba con una acción en el punto crítico. En este caso la inteligencia aprende de los errores. Otro punto a considerar es que el algoritmo únicamente entiende la acción de saltar en el punto crítico, eso significa que no fue programado para acciones como frenar, detenerse, etc.

  7. If Nintendo partnered with NASA and slapped Mario on some Mars simulations we’d be there by now

  8. I've just realized the Gens (the first ones) are just like a random demo play.
    This topic seems interesting and sounds cool so i guess i'll look up for other videos like this hehe. Gotta learn a little about AI. Great video by the way it was fun to watch!

  9. The Model was Great!

    But what if the Sprite changed their movements every time you restart the game?
    What if the Sprite is an AI itself like most of the modern games do?

  10. A single supercomputer in 2017 have the computational power to simulate a singular human brain. While terribly inefficient (megawatts of power required verse 300 watts), it is possible. What's next?

  11. Really cool overview. Would love to learn more about how it works and was implemented.

  12. I wonder how other strategies in this whole process would affect the behaviour of the AI. Most of the time you work with population based heuristics, you just follow a classical GA (Genetic Algortihm) approach. These consist of the three aspects of evolution: Selection, Mutation, Recombination
    Basically, different approaches, e.g. just allowing mutation the whole time and forget selection and recombination, explore different areas of the solution space(this would be the settings of a certain AI instance in this game). Therefore, they could follow completely different strategies and even be better, than the one we see in this video.

  13. I'm a data engineer, never heard of you until this. I've just subscribed.

  14. Every single human will be able to beat this bot

  15. I thought he'd put the neural network as a different window with normal textures but that too is Mario styled! omg you didn't have to go so far 😂

  16. I watched a live twitch of machine learning this

  17. how much time does it took to train the model?

  18. Why are you using generations and species what is the total number of species before Mario beat the level?

  19. Hey seth! I'm currently studying machine learning and just want to thank you for being my inspiration of sorts when I first saw this video. Really interesting video and well done for creating it!

  20. the problem is that you arent studying visual input to a neural network, hook up a camera of some sort and allow that visual input to run through some kind of neural network and you might end up with different results.

  21. After training, is it only capable of playing the level you trained it on, or can it play the next level (without having seen it before), too?

  22. I want to see this A.I. on Monster Hunter Freedom Unite….that would be insane.

  23. This would only work if the stage remains consistent. If enemies change directions or stage direction changes to the left, would the algorithm be evolving in circles? lol

  24. Great video. I am attempting to convert to javascript. I am running in to issues with indexing as lua using base1 javascript base 0…. Anyhow has anyone else attempted or completed the conversion?

  25. like 6 years later: video still interesting

  26. So how do we test with this program on another level?

Leave a Reply

Your email address will not be published.