Computers Give Other Computers Advice On How To Play Pac Man
[ Watch the Video: Computers Teach Each Other Pac Man ]
Peter Suciu for redOrbit.com – Your Universe Online
It sounded like a classic April Fool’s joke – computers teaching other computers to play Pac Man. But, researchers at Washington State University are in fact experimenting with machine learning by teaching computers how to teach each other.
Instead of teaching the machines advanced programming, the researchers at the School of Electrical Engineering and Computer Science have developed a new method to allow a computer to give advice and teach skills to another computer.
This reportedly mimics how a real teacher and student might interact, even if does sound more like something two gamers would talk about with a video game controller in their hand.
Matthew E. Taylor, WSU’s Allred Distinguished Professor in Artificial Intelligence, led the study, which was published online in the journal Connection Science.
According to the paper’s abstract the research, “introduces a teacher–student framework for reinforcement learning, synthesising and extending material that appeared in conference proceedings.”
“In this framework, a teacher agent instructs a student agent by suggesting actions the student should take as it learns. However, the teacher may only give such advice a limited number of times,” the abstract noted. “We present several novel algorithms that teachers can use to budget their advice effectively, and we evaluate them in two complex video games: StarCraft and Pac-Man. Our results show that the same amount of advice, given at different moments, can have different effects on student learning, and that teachers can significantly affect student learning even when students use different learning methods and state representations.”
For the study, the researchers had the agents – the term for virtual robots – act like a true student and teacher, where one agent taught the other to play either Pac Man or a version of StarCraft. This is a lot more than fun and games, suggest the researchers, as helping robots teach each other to play games is not easy. However, as the agents can learn to teach games they could then teach other tasks.
This could lead to machines that could teach its replacement – such as a housecleaning robot, the researchers suggested. While this sounds like just the sort of scenario in science fiction stories where the robots learn from humans and eventually take over, Taylor said that isn’t a real possibility for now.
“They’re very dumb,” Taylor, an expert on robots, agents and AI said in a statement. He added that the most advanced robots can be easily confused and when they get confused they simply stop working.
However, one goal of the researchers is to eventually have robots and agents teach skills to humans – and not just how to play video games such as Pac Man or StarCraft. In the study, the researchers programmed their teaching agent to focus on action advice and when to tell a student to act.
When to act is important say the researchers for when no advice is given the robot isn’t learning; but if it always gives advice the student may get annoyed and doesn’t learn to outperform its teacher.
“We designed algorithms for advice giving, and we are trying to figure out when our advice makes the biggest difference,” Taylor added.
The work was funded in part by the National Science Foundation (NSF).