MarI/O – Machine Learning for Video Games

Here's the author taking the world record (now superseded) at glitching mario world into the credits to "win" the game in 3:07:

https://www.youtube.com/watch?v=HxFh1CJOrTU

This is on a real console, not an emulator. They overflow the snes into the controller registers to achieve arbitrary code execution by hand. They do this by placing koopa shells at pixel-perfect locations to spell out ASM, overflow to the controller registers, and press a specific combination of buttons to jump to the sprite table and their literal shellcode.

At the start of this year there was a crazy few weeks where different speedrunners were competing to execute the "credits warp" in the quickest and most efficient way. It started at around 6 minutes, now it's down to around 2 I think, most of which is the game's intro. Fun stuff. SethBling was the first person to ever execute it on a real console.

Here's a video explaining it, from a different speedrunner: https://www.youtube.com/watch?v=vAHXK2wut_I

Thank you for sharing. The video explanation is really easy to follow.

For interested people, here is an interview with K. Stanley, the author of NEAT: http://aigamedev.com/open/interviews/galactic-arms-race/

I was wondering if this learning technique was applied to any other games. I came across this:

1. https://www.cs.utexas.edu/~mhauskn/papers/atari.pdf

2. http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.93....

Bizhawk is a SNES emulator for Game Casters.

Seth, the video author is a Super Mario World Speedrunner and Twitchcaster. http://www.twitch.tv/sethbling

Bizhawk has Lua Scripting built in : http://tasvideos.org/Bizhawk/Features.html

Seth implemented Neuro-Evolution in Lua and http://pastebin.com/ZZmSNaHX

now machine learning casting is a thing...

This is very cool, but it isn't really "playing" the game. Instead it's "memorizing" how to complete a specific level through trial and error.

If you took this neural network and tried it on another Mario level, it would fail miserably.

I have reasonable suspicion that it wouldn't be nearly as easy to train without "spinning" move which basically gives invulnerability against all enemies sans their bullets, iirc.

TLDR: a neural network playing Super Mario.

A very cool demonstration showing how the network evolves.

This is a really cool and intuitive introduction to this sort of algorithms, and the example and visualization make it very accessible. Great work, I'm sure it will inspire some kids to get into ML and neural networks.

I apologize, this is pedantic, but I think people often overstate the relationship between neural networks and the brain. I'm far from an expert, but saying neural networks are based on the brain is a stretch. They're inspired by a theory of how the brain works that is not wholly in line with what most scientists believe.

It's so tempting to say, this is based on the brain, when really it's more like, this was inspired by one psychologists model of the brain.

Awesomeness. How you explain machine learning principles/NEAT really shows a firm grasp. The fact it's done in lua...makes it even more eccentrically cool.

Veery nice visual introduction to the basics of neural networks