Google DeepMind demonstrated a few years back that artificial intelligence (A.I.) could learn to play retro video games better than the majority of human players, without requiring any instruction as to how they should accomplish the feat. Now, researchers from Georgia Tech have taken the next logical leap by demonstrating how A.I. can be used to create brand-new video games after being shown hours of classic 8-bit gaming action for “inspiration.”
The results? New titles like “Killer Bounce” and “Death Walls,” which look like they could have stepped directly out of some grungy 1980s video arcade, designed by machines way more sophisticated than any 1980s computer scientist could have imagined.
“Our system operates in several stages,” Mark Riedl, associate professor of Interactive Computing at Georgia Tech, told Digital Trends. “First, we take video of several games being played. In this case, the games are Super Mario Bros., Kirby, and Mega Man. Our system learns models of the level design and game mechanics [and] rules for each game. The machine learning algorithms we use are probabilistic graphical models for learning level design, and a form of causal inference for learning game mechanics.”
Had these models been used to generate new games, Riedl said the resulting games would have looked just like the ones that inspired them. (This has been the basis for previous work by the team.) Instead, their algorithm carries out something called “conceptual expansion,” which infers the existence of models for games that do not exist, but potentially could, based on what is learned from the input game video. The A.I. then generates games that fall into the overall hypothetical game models — bearing enough resemblance to other titles’ game mechanics to be familiar, but not exact copies.
For instance, in the game Death Bounce, the system has learned that some objects disappear when hit from above, and applies this concept to the ground instead of enemies. In the fame Killer Walls, the A.I. creates an enemy wall based on a combination of its understanding of enemies and wall obstacles.
Ultimately, what interests Riedl and Ph.D. student Matthew Guzdial is the question of whether machines can play a role in carrying out creative acts, such as designing video games. “Games are really complex and really hard to make, even for experts, so the ability of an algorithm to create interesting, working games is a notable achievement,” Riedl said.
A paper describing the work is available to read here.
