This Video Game Tournament Will Test the Wits of Computers, Not Humans
Computers will need to bring major problem solving skills and smarts to a new video gaming tournament that is intended to contribute to cutting-edge artificial intelligence research.
The Obstacle Tower game and challenge was revealed on Monday by AI researchers from Unity Technologies, an emerging player in technology for the video game industry. The goal is to encourage researchers to use the game to test the capabilities of AI technologies like deep learning, which computer scientists use to train computers to understand images like dogs in photos.
The hope is that the AI techniques developed by researchers who participate in the Obstacle Tower Challenge could eventually be used to improve self-driving cars or cutting-edge robots, said Unity’s vice president of AI and machine learning Danny Lange.
To entice researchers to participate, Unity partnered with Google to offer cash prizes, travel vouchers, and credits for the search giant’s cloud computing service that cumulatively total over $100,000. The challenge runs from Feb. 11 to May 24 and is divided between two rounds; organizers will announce the winners on June 14.
Obstacle Tower resembles a traditional first-person video game in which the heroine—a cartoonish-looking woman with purple hair—must navigate a maze of corridors to climb up the tower.
On each tower floor, the heroine must find a key that will unlock the correct door so she can exit the maze and proceed to the next level. There are several obstacles to navigate along the way, including giant pits and platforms that she must jump between to move through the maze.
In recent years, video games have become a popular testing ground for some of the world’s leading AI researchers, like at Google’s DeepMind unit, to showcase advanced software that beats human players. For instance, DeepMind’s AI software, AlphaStar, recently defeated humans in a tournament involving the popular computer strategy game StarCraft II.
Lange, who has a been a machine-learning executive at ride-hailing giant Uber and Amazon Web Services, said Obstacle Tower was created with the help of AI researchers at New York University to be more difficult for computers than current video games. The game’s makers want it to be a benchmark for other researchers to test how well their software performs at various AI-related tasks.
Because the game’s levels are randomly generated, computers playing it must consistently adapt to new environments—a more challenging task then simply mastering the same level repeatedly because “the system has to learn to generalize, and that is what AI is about,” Lange said.
Using an AI technique known as reinforcement learning, the computers in the competition must learn to play the game through trial and error—that keys scattered around the maze will unlock doors and that falling into deep pits is dangerous. In later levels, the computers will be faced with new challenges, like having to move bricks in certain places to solve puzzles.
The computer, Lange said, “has to experiment and figure out what do I need to do.”
Data Sheet, Fortune’s technology newsletter.
At certain times during the game, sunlight may splash onto the tower’s corridors, causing visual distortions, while certain doors may have ornate decorations that make them appear different than previous doors in the game, Lange said. This is challenging for computers trained to recognize images because sometimes even the slightest variable—like a man wearing glasses—can cause a computer to stumble when identifying someone in a photo.
Lange said that researchers will have to create neural networks, the software used for machine-learning training, that work together in order to navigate the heroine through the maze. For instance, a neural network designed to recognize objects like doors would need to communicate with the neural network responsible for the heroine’s decisions like whether to open a door. Lange likens the cooperation to the different parts of the human brain that must work in unison so that people can see and respond to their surroundings.
Although the AI techniques Lange hopes researchers will create by participating in the game could eventually be used to improve autonomous cars, some could be incorporated relatively quickly into popular consumer-facing apps. Lange explained that techniques derived from the game could be used to improve how retail apps recommend products for people to buy. He’s already familiar with such apps, having previously helped build Amazon’s systems that suggests products for people to purchase.
Like how computers playing the game must explore their digital environments to figure out their next steps, Lange imagines similar AI systems “exploring” new ways to recommend products for people to buy. Some of the most advanced AI systems that researchers have created to master video games are designed to be “curious” by exploring new places and taking risks.
Lange imagines an AI-powered retail app that takes the risk of showing shoppers something that they may not be interested in, but by chance, happens to cause a “chain of reactions” that ultimately lead to more sales. For instance, the app could recommend a book about barbecuing, which could lead to a shopper to eventually buy gas grills or other barbecue-related merchandise.
For now, Lange hopes that the game and tournament will lead to more AI researchers using Unity’s video game tools, which already power popular video games like Pokémon Go and the fantasy card game Hearthstone.
“So gaming is just a fantastic equivalent for nature, and if nature evolved intelligence why wouldn’t games do the same?” Lange asked.