Let AI learn "How to get a star in Super Mario 64"

Machine learning andneural networkMany engineers are challenged by trying to learn artificial intelligence (AI) using AI and so on so that the game can be cleared by AI alone. Google also uses game-specific "DQNWe have developed an AI called "AI"Super Mario WorldYaSuper Mario KartAn attempt has been made public on the Internet to let you play. With such an AI × game attempt, finally it is not a 2D game but a 3D game "Super Mario 64"Has been started to learn.

AI Learns To Play Mario 64, Gets A Star
https://kotaku.com/ai-learns-to-play-mario-64-gets-a-star-1823170045

Up to 24 people can play Super Mario 64 at the same time "Super Mario 64 OnlineAnd hacking Super Mario 64 and playing in Mario perspectivemovieWe are publishingKaze Emanuar, "A star in the back of the maze" or "Yellow rabbit running away"We are challenging to navigate Mario to the place of.

You can see how the AI ​​operated Mario is approaching until the star is in the following movie. It is like the AI ​​that is written on the bottom left of the game screen, one of which is "mAIrioIt is an open source one whose code is published on GitHub called "GitHub."

ETHAN WINS


Kotaku of the game media interviewed Emanuar by e-mail, explaining that the AI ​​being used is "very slow", these AIs apply extensively the solution under the specific circumstances It is said that there is a tendency to do. So it seems that it is not uncommon scene that AI repeats jumps and punches in the same place for a while. Nevertheless, these AIs have succeeded in getting the stars more than once, "While one AI succeeds in getting the star in about 15 minutes, this is more than a technical aspect It is because it was good, "Emanuar says. Also, with AI used by another user, we succeeded in getting one star in about 3 hours.

About adventuring the complex stage of Super Mario 64 to AI, Emanuar says "Since the entire 3D space is too complicated, we can not enter all the information in AI's brain, simplifying this input information, We need to work fine on detailed processing details such as small space of processing capacity. "

Emanuar seems to have had no experience using AI so far, it is trial and error while reading some books. Still, AI seems to have learned some interesting actions, "One AI has learned to jump and access the high place and another AI to learn to jump three times with these AIs The action learned is such an action that ordinary players never do, it seems to be a hidden technique of Super Mario 64. "

You can also see Emanuar 's Twitch channel accessible from below and see a part of the AI' s playing Super Mario 64.

m2mkaze - Twitch