Microsoft Research recently publicised their use of Minecraft, and project Malmo, to explore agent based AI approaches to complex problems. Below is a quick GIF of a project Malmo Minecraft agent using Q-learning to learn to avoid lava.
The aim of this game is to reach the blue block as quickly as possible, without dying a horrible death of course. As an added difficulty the agent has no visibility of the space, and must learn to avoid the lava using only trial, error, and reinforcement learning. Using this approach (Q-learning to be exact) the agent can quickly map out the safe path, then learn to navigate it faster.
This is still a pretty simple demonstrative algorithm, but the environment and the ability to use Minecraft characters as experimental agents in an easy to understand and visual way is really cool and I hope to build and share more complex and multi agent scenarios soon!
