Artificial intelligence models have successfully mastered modern video games such as Dota 2 that boast complex three-dimensional environments and controls. Now, Microsoft Corp. wants to up the challenge in a bid to advance AI research.

The company on Thursday open-sourced TextWorld, a framework that can generate games for training neural networks. More specifically, it creates text adventures, which don’t have 3-D graphics or complex controls but can be immensely challenging for an AI.

Text adventures, which trace their roots all the way back to the 1960s, are essentially a type of interactive story. Instead of presenting a visual environment for the player to navigate, the game describes a situation or puzzle in writing. The player can advance the story by typing in input such as “move left” or “open the box” until they reach the ending.

Microsoft created TextWorld because these types of games present a multidimensional challenge for AI models. To get through a text adventure, a neural network has to parse the information on the screen, figure out how to proceed and then formulate its answer in a cohesive sentence. More modern games such as Dota 2 have their own complexities, but they don’t create an opportunity for an AI to hone its natural-language skills.

“One reason I’m excited about TextWorld is the way it combines reinforcement learning with natural language,” said Geoff Gordon, a principal research manager at Microsoft. “These two technologies are both really important, but they don’t fit together that well yet. TextWorld will push researchers to make them work in combination.”

Some text adventures can also have extra twists that require even more advanced processing capabilities. For example, an AI might have to find and hold onto a certain item in order to clear the last level, which necessitates the ability to plan ahead. Combined with the fact that the turn-based nature of the genre doesn’t leave much room for mistakes, these factors add up to form a highly sophisticated training environment.

TextWorld enables researchers to define how games should play out by providing high-level specifications. According Microsoft, the tool can customize everything from the number of turns in a text adventure to what objects and furniture the virtual setting should include. It also lets researchers use existing, human-written games to create an even more challenging training environment.

Image: Microsoft