U.S. computer scientists say they've come up with a way to let a computer offer advice to another computer and teach it new skills in the same way a human student and teacher would interact.
In a research led by artificial intelligence Professor Matthew E. Taylor of Washington State University, virtual robots dubbed agents were connected in teacher-student pairs while the student agent attempted to learn how to play video games such as Starcraft and Pac-Man.
In the experiments the student computers, aided and instructed by the teacher computers, showed the ability to learn the games and, in the end, surpass the teachers in game-playing skill.
While learning to play a game may seem a frivolous activity for a computer, it represents a serious research effort in robotics, the researchers said.
If robots were able to teach another robot a task, it would free people from having to program each new robot. For example, a robot designed for housecleaning could teach the job to a replacement robot.
While the thought of robots learning from each other raises the specter of their taking control of the world, Taylor says we needn't worry.
"They're very dumb," he says, noting that even the most advanced of robots are prone to becoming easily confused.
Confusion generally leads robots to stop functioning.
While it is easy to teach a robot a new skill by simply switching the "brain" -- the CPU and memory -- from one robot to another, software and hardware compatibility between an older robot and a newer one would likely cause problems, Taylor says.
So rather than a hardware technique, the researchers focused on programming their teaching computer to concentrate on advice about actions to take or telling the student computer the appropriate moment to act.
"We designed algorithms for advice giving, and we are trying to figure out when our advice makes the biggest difference," Taylor says.
As any parent of a teenager knows, knowing when to give advice is almost as important as the advice being given.
It's the same with computers, Taylor says that if the teacher is not giving advice, there is no teaching taking place. But if advice is given constantly, the student -- like a teenager -- gets annoyed and resists learning anything.
Taylor says a goal of the research is to create a curriculum in which the teaching agents would begin with simple work and progress to ever more complex tasks.
