A computer at the Georgia Institute of Technology is generating reams of text that would look like gibberish to most people. To Matthew Guzdial, a PhD student at GIT’s School of Interactive Computing, the text looks like video games.
”Matthew’s like in The Matrix, he’s reading these printouts and saying ‘Oh yeah, that’s a good game,’” Mark Riedl, an associate professor of AI and machine learning at GIT, told me over the phone.
Guzdial and Riedl are co-authors of “Automated Game Design via Conceptual Expansion,” a new study designed to teach an algorithm to design its own video games. The study is the latest iteration of a long running project where Riedl, Guzdial and many others are attempting to teach AI how to be creative, often using video games.
Riedl and Guzdial fed a machine learning algorithm video of humans playing the first levels of Super Mario Bros., Kirby’s Adventure, and Mega Man for the Nintendo Entertainment System. Once the program had absorbed all of the footage, it was able to probabilistically map the relationships between in-game objects and how the relationships change to generate “game graphs” that laid out how the games worked, and which the scientists can visualize in diagram form. The pair then asked it to create its new game graphs based on what it had learned—essentially, creating new games.
“It’s the mimicry approach to creativity, which isn’t a bad place to start because humans also learn to be creative by mimicking at first,” Riedl said.
To start, the team tested the AI by asking it to recreate certain games. For example, after hours of training on the opening levels of Mega Man, the program was asked to complete the game based on its learned assumptions. “If it comes up with something that looks like the original Mega Man, we call that success,” Riedl said, “but it’s all subjective."
Mega Man gave the AI problems. The original game includes the Magnetic Beam, a gun players can use to create platforms that allow them to traverse the level. “The system wasn’t able to infer the mechanics from the levels as well [as other games],” Guzdial said. Super Mario Bros. was simpler for the AI to understand because there are no power ups that let you jump higher or longer.
But successfully creative humans tend to make a leap from mimicry, and the AI isn’t just replicating what it’s seen, either—it’s also mixing mechanics and level design to make new games. “The other day it made this really cool game where every time you jumped on a platform you also destroyed that platform,” Riedl said. “You had to continually moving right, you couldn’t pause ever. It was really cool.”
Right now, Guzdial is one of the few people who actually knows how those games play because he can read the code and interpret what the AI designed. “It’s literally printing out levels in text representations, like big ASCII art style levels,” he said.
“It generated lots of games and most of them didn’t fit the bill,” Guzdial said over the phone. One of the early problems for the project was determining what makes a good game, and the pair plan on letting the public decide. The team is building an emulator and plans to release some of the AI-designed video games online in October or November.
“Creative” algorithms are still in their early days and Guzdial and Riedl aren’t sure what AI’s place will be in the game design process.
“We don’t know the right way humans and AI should work together on the design level,” Riedl said. “The AI could be an assistant, it could be an outsourcer that fills in the blanks, it could be a peer, and—at the extreme—it could take the lead and the human could be the assistant.”
Get six of our favorite Motherboard stories every day by signing up for our newsletter.
Correction: An earlier version of this article incorrectly referred to Georgia Tech's School of Interactive Computing as the School of Interactive Gaming. Motherboard regrets the error.