YOUR plucky blue box bounds over gaps and hops onto platforms, the pulsating red wall never more than half a screen behind. If it catches you, it is game over. Then the wall falls into a hole, you bounce on it and it is gone. This is Super Mario reimagined by artificial intelligence.
Matthew Guzdial and Mark Riedl at the Georgia Institute of Technology in Atlanta have created a machine-learning system that has taught itself the basics of video game design by watching videos of people playing classic games, such as Super Mario, Kirby and Mega Man.
Advertisement
This allows the AI to learn about the interactions between characters and objects: what happens when this thing goes over there, what happens when this thing touches that thing. It then takes the patterns it has identified in individual games and remixes them into new ones. “Some are good, some are bad,” says Guzdial.
[video_player id=”xvAtC8pN” access_level=”everyone”]
The above game, nicknamed Death Wall, was one success. Another, called Killer Bounce, has the player jumping across blocks that disappear when you touch them, forcing you to keep bouncing. It was the result of the system taking the Super Mario rule that you dispatch enemies by bouncing on them and then making everything in the game an enemy ().
At the moment, Guzdial’s system produces very abstract, blocky games, but he would like to add more recognisable character designs and backdrops. These could also be machine-generated. Guzdial has previously used AI to generate new Pokémon, for example.
Eventually, he would like anyone to be able to use such systems. “People often say they want a game that’s a bit like one thing and a bit like another,” says Guzdial. “Wouldn’t it be great if you could show the AI Fortnite and Minecraft and it spat out something new you could play?”
This article appeared in print under the headline “Machines make the next Super Mario”
