Ready Player Two?

December 8, 2018 7:47 PM

Back in 2009, the first Mario AI competition was held, using Markus Persson's Infinite Mario. Robin Baumgarten won, but the contest fizzled out after a couple of years. Then in 2015, youtuber SethBling (best known for his Minecraft and Mario hacks) coded a machine learning Lua program to play Super Mario games, cleverly named MarI/O. What was revolutionary is that this neural net was evolutionary. (Good explainer if you don't want to watch video.) But what was really meant as a demo took on an extra life of its own with Luigi/o, a livestream that's been running for months, improving the code, and finally completing every level of Super Mario Brothers.

It's now running through all the levels in Hard Mode, after which they plan to tackle the Lost Levels (Japan's SMB2). Meanwhile Robin Baumgarten has taken on even weirder projects.

posted by rikschell (3 comments total)

Possibly related: an AI project at Uber can now completely demolish “Montezuma’s Revenge” on the 2600 thanks to coupling the usual neural net sort of thing with a crude kind of memory. https://venturebeat.com/2018/11/26/uber-ai-reliably-completes-all-stages-in-montezumas-revenge/
posted by egypturnash at 9:23 PM on December 8, 2018

This is how they take our jobs.

Also, super cool. Great post.
posted by Fizz at 5:02 AM on December 9, 2018

Also in related AI learning games news: AlphaZero taught itself to play three different games [Ars Technica]

“Google's DeepMind—the group that brought you the champion game-playing AIs AlphaGo and AlphaGoZero—is back with a new, improved, and more-generalized version. Dubbed AlphaZero, this program taught itself to play three different board games (chess, Go, and shogi, a Japanese form of chess) in just three days, with no human intervention.”

!
posted by Fizz at 5:08 AM on December 9, 2018