Microsoft lets AI experiments loose in world of Minecraft

It unveiled the project, then known as AIX, back in March, but at the time only a few academics had access to the code. On Thursday the company made good on its promise to open up the source code by publishing it on Github.

Minecraft, the blocky world-building game that Microsoft paid US$2.5 billion for two years ago, is an ideal place to test how artificial intelligences will interact with one another and with humans.

As it's a simulation, Minecraft is a safe place to test how AIs learn to perform certain kinds of physical tasks: In Minecraft, a rogue machine or runaway car can hurt no one. Since the Minecraft server controls the perceived passage of time, researchers can speed up their simulation so there's no waiting while heavy loads are lifted or lowered, for example. And since everything is simulated, instrumentation is a cinch: The server can measure and monitor every detail for later replay and analysis, making it easier for other researchers to reproduce published results.

This isn't Minecraft's first foray into academia: Microsoft took the wraps off an educational edition of the open world-building tool in January, and plans to begin selling it in September.

Project Malmo allows researchers to modify the underlying code of the Minecraft server, allowing them to introduce AI elements to the virtual world.

"AIs" have long been a component of video games, often controlling the baddies in shoot-em-up games to provide players with more of a challenge than randomly moving enemies would provide. But those AIs are dumb in comparison to what's possible in other fields of endeavor.

With Project Malmo, researchers will be able to build AIs that learn, with the goal of helping them hold conversations, make decisions and complete complex tasks.

The environment will be particularly suitable for the development of reinforcement learning techniques, whereby AIs are given a lot of leeway in how they perform tasks, and rewarded when they advance toward their goals, according to Katja Hofmann, lead developer on Project Malmo at Microsoft's research lab in Cambridge, England. For AIs, a "reward" is confirmation that a decision is an appropriate step toward their goals.

AIs are getting pretty good at talking and parsing human language, written and spoken, but for the most part have no idea what it means. Putting them in a simulated environment where they can associate words and actions will give them the opportunity to learn what those words really mean, just as humans do, Hofmann said in a Microsoft blog post about the release of the Project Malmo code.