Researchers Are Improving Machine Learning By Making Computers Teach Each Other Pac-Man

I wish this weren’t April 1, because then you’d all appreciate how cool this is, but as far as I can tell, this release posted a few days ago is completely true. In short, researchers at Washington State University are experimenting with machine learning by teaching computers how to teach each other. The plan is to have one computer show another computer how to play Pac-Man in a way that does not involve simply moving the programming from one machine to the other.

“We designed algorithms for advice giving, and we are trying to figure out when our advice makes the biggest difference,” said AI professor Matthew E. Taylor. The teaching computer “shows” the other computer how to play “well” i.e. not get eaten by ghosts and how to grab the most points. The student-teacher pairs also taught each other how to play StarCraft and, at points, the student ended up being better than the teacher.

From the release:

In their study, the researchers programmed their teaching agent to focus on action advice, or telling a student when to act.
As anyone with teenagers knows, the trick is in knowing when the robot should give advice. If it gives no advice, the robot is not teaching. But if it always gives advice, the student gets annoyed and doesn’t learn to outperform the teacher.

The research has applications in machine learning – robotic flocking and the like – as well as in human-computer interaction and learning. If a robot can teach another, arguably stupid robot to play Pac-Man, imagine what advice a robot could give students about math, science, and the like?