Microsoft uses Minecraft for artificial intelligence research
Microsoft has developed an artificial intelligence platform called AIX, which allows researchers to search Minecraft for ways to improve artificial intelligence. The intelligence learns from its actions within the game.
“It’s a digital baby box for artificial intelligence,” says project officer Fernando Diaz. “It’s an environment where we can develop an algorithm that allows a young artificial intelligence to learn about different concepts in the world.” That is, a learning system starts in the Minecraft world without knowing anything about it. Through a lot of trial and error, a system has to figure out what are important and less important things, and through increasing rewards it has to understand when a certain goal has been achieved.
Known as ‘Project AIX’, the project is powered by the AIX platform, which was developed by researcher Katja Hofmann at the Microsoft lab in Cambridge. This software will be available open source for Windows, Linux and Mac OS from July, so interested parties only need to purchase the PC version of Minecraft to get started. Project AIX was not made for consumers, but the lead developer hopes it will attract many researchers and amateurs. The goal was to develop a platform that supports Microsoft’s research and is also useful to the rest of the research community.
Hofmann chose Minecraft a year ago out of frustration with other platforms that use simple games. Minecraft offers many possibilities, from simple actions such as walking to building complex machines. This open world makes the game extremely suitable for research into artificial intelligence. Microsoft is concerned with developing a ‘general’ artificial intelligence that is not only good at performing a limited number of tasks, but that learns just like a child from all kinds of different stimuli.
It would be possible to develop such a system by incorporating an artificial intelligence into a robot and having it perform various tasks, thereby triggering a learning process. However, this method is expensive and therefore a software platform is a good solution for researchers.
In a recent series of go competitions, Google’s artificial intelligence AlphaGo managed to beat Korean world champion Lee Sedol three times in a row. The system often took surprising decisions, which later turned out to be crucial. Lee did win the fourth game against AlphaGo.
Members of the Project AIX team