Is there no end to our hubris? As a race that forces robots to hoover our floors, we might think twice about teaching them to navigate a complex world of their own volition, but that's exactly what the team from Microsoft UK are doing with AI in Minecraft (opens in new tab).
Project AIX (opens in new tab) is using Minecraft's procedurally generated landscapes to train AI to learn in a changeable world. Currently, the team is asking it to climb to the top of the highest hill it can find.
It knows nothing about the world it spawns in. For example, until it gets going it doesn't know whether day or night is important to hill climbing. It doesn't know about the dangers of lava pits, and only repeated encounters are going to teach it that no, lava is not conducive to mountaineering. It needs to learn that descending into a cave system is not the best way to go up. This is the difference between machine learning and what we usually refer to as 'AI' in games—AIX is not pre-programmed with the rules.
At the moment, Project AIX is in the hands of select academics, but this summer it will be going open-source. Think of what modders might do with it! Imagine an AI whose goal is to steal diamonds making progressively smarter attempts to access your stronghold as it realises you are the biggest threat to its diamond-gathering mission. It's a long way off that level of smarts, I'm sure, but what can't crowdsourcing fix?
The end goal is not to produce cleverer Minecraft bots, strange though that might seem. Rather, researchers are hoping to use the infinite, dynamic environment to improve AI development at large.