Mario may be super, but even he must get bored hurdling the same Goombas and falling off the same cliffs over and over. Fortunately, a new artificial intelligence (AI) algorithm can endlessly produce new levels, and even tailor them to a player’s skill level.
Computer scientists have spent decades honing “procedural content generation,” the use of algorithms to automatically design new characters, landscapes, and weapons for video games, saving humans hours of labor. (The 2016 game No Man’s Sky can generate up to 18 quintillion unique planets as players explore the galaxy, a daunting task for any human designer.) But programmers still need to handcraft the rules that tell the computer how to create such content. In recent years they’ve applied machine learning, an AI technique by which computers learn from examples, so AI can simply produce more content in the style of existing content without needing explicit instructions. Game levels are particularly tricky to generate, however, because small changes can make them unplayable—a stray wall can seal off a critical passage, for example.
The new method learns to imitate human-created levels and then allows for customization. It has two phases. In the first, a “generative adversarial network” (GAN) learns through trial and error to transform strings of numbers into levels indistinguishable from human-created levels. A second phase then helps find strings of numbers that lead to levels that are not just realistic, but that fit certain requirements—such as having a lot of enemies or jumps. The authors gained precise control over how difficult the levels were, they report in a paper to be presented in July at the Genetic and Evolutionary Computation Conference in Kyoto, Japan. They believe their approach would work for other games, too.
Another new method uses GANs to produce new maps for Doom, the classic first-person shooter game. The algorithm creates Doom maps that match human-created ones visually as well as on certain higher-level features such as the balance of large and small rooms, the authors report in a paper posted to arXiv last month. (Both papers were mentioned in Import AI, a newsletter.) Procedural content generation doesn’t just save designers time and save Mario from tedium. It could also help video games adjust to the skill levels of players on the fly. Super.