June 23, 2009 2:00 PM | Eric Caoili
Several Rutgers University researchers devised a new method for "reinforcement learning" (a sub-area of machine learning) using Object-Oriented Markov Decision Processes, which is described as "a representation that looks at a higher level than usual and considers objects and interactions."
If that sounds complicated, their demonstration makes the concept much easier to understand. They showed the OO-MDPs representation by presenting a system that learned to play the original Pitfall in an Atari 2600 emulator (shown in the video above).
At first, you can see the system's learning algorithm discovering how to progress from the first screen, experimenting with different actions. In the system's second run, it uses what it learned, celebrating with a dance of joy!
According to the reasearchres, OO-MDP will also have plenty of "plenty of 'serious' uses in addition to being used in video game testing and in-game AI".
[Via Nick Montfort]